# Bridge SEO — full text
Every published page of bridge-seo.com, in build order. Generated from the same sources as the site.
--- https://bridge-seo.com/ ---
# AI SEO: filling the meta tags a page is missing
The tags you wrote stay exactly as they are. Everything else is filled at
the edge, with no plugin and no code change.
## Four places your tags show up before anyone visits
Before a person opens the page, four kinds of screen have already shown
it, and each one is built from the tags at the top.
**Search result.** Google, Bing and other engines show a title and a
snippet under each result. Without a description, the snippet is whatever
line the engine picks from the page.
**AI answer.** ChatGPT, Claude and Perplexity fetch pages while they answer
a question, and read what the page says about itself in its tags.
**Link in chat.** Slack, LinkedIn, WhatsApp and iMessage draw a preview
from Open Graph. Pasted without a card, the link shows as a grey box.
**Social post.** X reads its own card tags and falls back to Open Graph for
the rest. With neither, a post shows the bare link.
## You wrote the first fifty. The site kept growing.
Tags get written while a site is small enough for one person to know every
page. After that, new pages ship with a title and whatever the template
adds.
A description for every page is not a writing job. It is enumeration, and
enumeration is the work nobody gets around to.
## It fills what's empty. It never rewrites what's there.
- A crawler or AI agent requests a page.
- The edge reads the head and finds what is missing.
- Missing tags are written from the page and your brand profile. Yours ship
untouched.
Enrichment fills the tags a page is missing (meta title and description,
Open Graph, Twitter and X cards, JSON-LD) in the rendered copy machines
receive. It never touches a tag somebody already wrote. An authored tag is
kept exactly as it stands, and a generated one appears only where that tag
is absent, so writing your own description later does not invite the
generated version back over it.
## What this never does
**Nothing on the path is blocked or slowed.** Crawlers and AI agents are
served, not screened.
**Google does not participate in IndexNow.** Submission reaches the
engines that do, and the scope there is narrow: discovered and changed
URLs only, no history imported, full-site submission manual.
**It does not move a page up or down.** A filled tag describes a page. An
engine still decides whether to show that description at all, and where the page ranks.
## Four tag families, one composed card
Each family opens its page: what the tags declare and what enrichment
fills. IndexNow is the step after them.
- [Meta title and description](/ai-seo/meta-title-description/)
- [Open Graph](/ai-seo/open-graph/)
- [Twitter and X cards](/ai-seo/twitter-cards/)
- [JSON-LD](/ai-seo/structured-data/)
- IndexNow, the step after them
**Meta title and description.** The pair a result is built from. Where a
page has no description, one is written from what the page says; where it
has no title, a title is written too, ending with the site name when the
brand profile says so.
**Open Graph.** What a pasted link unfolds into. Missing properties are
written from the page and the brand profile: the site's name, and a
default image where one is set.
**X cards.** X reads its own card tags first. Where they are absent, they
are written from the page and the brand profile, including the handle
where you have set one, and the card shape is chosen to match the image
the page has.
**JSON-LD.** A structured statement of what the page is. The type comes
from what the page is, the organisation and site from the brand profile,
and the breadcrumb from the path. A block somebody wrote is never edited.
**IndexNow.** The step after the tags: when a page's tags change, that page can be announced to the engines that accept such
announcements.
Where one is missing, a card can be composed at 1200 by 630 from three
things you already have: the brand colour, the typeface, and a picture of
the page itself. Nothing in it is invented, and a page that already has an image keeps it.
The frame around the screenshot can be drawn in the Windows style or the
macOS one, which sounds like a detail until the card is sitting in
somebody's feed next to four others.
Cards stay switched off even after enrichment is on, until you ask
for them.
## If your site renders in the browser, fix that first
A site that assembles its content in the browser sends a shell first, and
not every reader runs what comes after. Enrichment cannot fill a page that
arrives empty, so prerendering comes first; how it works is explained on
[BotBridge](https://botbridge.io/).
Switch between the three states below to see what changes.
## Is this worth switching on for your site?
Enrichment is worth switching on for a large site built from templates, where the missing tags are the ones nobody got to; a
small site with every description written by hand gains nothing.
**Worth switching on.** Hundreds or thousands of pages, most of them built
from templates, each with a title somebody did write. The missing tags are
the ones nobody got to, and filling those is the whole job.
**Look at rendering first.** A site assembled in the browser may be sending
machines a shell. Prerendering answers that part; enrichment only helps
once the content actually arrives.
**Not worth it.** A small site where somebody writes every description by
hand and keeps them current. Enrichment would find nothing missing and
change nothing.
## A plugin lives in your CMS. This lives in the path the page takes.
There are two places to put a missing tag: inside the system that builds the
page, or on the way the page travels to whoever asked for it. The comparison below sets those two forms side by side and names no product.
One thing a plugin does better: it shows its tags in the editor,
next to the page. Filled tags never appear in your CMS, only in the copy
machines receive, so the way to see them is to fetch the page the way a
machine does.
## Two settings decide what it writes
Two settings decide what it writes: a brand profile that is copied word
for word, and a description of the site that steers the wording. Generated
text is only as trustworthy as the facts it is given, and both settings are
yours to write.
**A brand profile, used word for word.** What the site is called, the
Open Graph site name where it differs, the social handle, the postfix on
a title, a default image and its alt text. These do not change from page
to page, so they are copied as written rather than guessed.
**A description of the site and how it should sound.** What you do, who
reads it, and the rules the copy follows: plain rather than eager,
British spelling, no superlatives. This steers the wording and is never
published on any page. It is there so the tags it writes sound like yours.
## What was filled, and when
Two pages in twelve got nothing, because nothing on them was missing. Every
title in the list is still the one somebody wrote.
## Included in Bridge CDN Pro
AI SEO has no price of its own. It is part of Bridge CDN on Pro and above,
and it stays off for each domain until you switch it on. Plans and prices
live on Bridge CDN.
Writing the missing descriptions by hand is arithmetic: pages times
minutes per page. Put in your own numbers.
## Everything above has a page behind it
Below AEO and GEO sit ordinary technical signals, and a findability problem is often one of them declared wrongly.
- [AI SEO](/ai-seo/)
- [Redirects](/redirects/)
- [Hreflang](/hreflang/)
- [Meta tags](/meta-tags/)
- [Tools](/tools/)
## Switch it on for one domain
Pick one domain, turn enrichment on, and read what machines receive from
its pages. Accounts are created on Bridge CDN.
## FAQ
**Does my JavaScript site get indexed?**
Not reliably on its own. A page assembled in the browser sends a shell
first, and any reader that does not execute that step receives the shell.
Prerendering addresses this by serving rendered HTML to crawlers and AI agents, so what they receive already contains the content.
**Are meta descriptions a ranking factor?**
No. Google has said since 2009 that it does not use the description meta
tag for ranking, even though it sometimes uses it for the snippet shown
under a result. That snippet is what somebody reads before deciding to
click. Google builds it mostly from the page itself and takes the
description when it describes the page better, so a missing one leaves
that choice with nothing to choose from.
**Is it safe to generate descriptions automatically?**
Yes, where writing them by hand is not realistic. Google's own
[documentation on snippets](https://developers.google.com/search/docs/appearance/snippet)
says that on large database-driven sites, programmatic generation of
descriptions can be appropriate and is encouraged. Enrichment works in that space: it writes only where no description exists.
**Do I need to install a plugin?**
No, and there is nothing to add to your codebase either. The work
happens at the edge, on the copy assembled for machines, and that is
also why it works the same on a site rendered in the browser as on one
rendered by a server: the plugin
ecosystem your CMS does or does not have never comes into it.
**What happens if I write my own description later?**
Yours ships. A generated tag is added only where that tag is absent, so
the moment your description exists, machines receive it exactly as you
wrote it, and the generated version does not come back over it.
Availability is Pro and above, and it stays off for a domain until you
turn it on.
**What does the model know about my brand, and does any of it end up on my pages?**
It knows what you put in the brand profile, and those facts are used
exactly as written. The short account of how your
copy should sound steers the wording and is never published: it is read
while a tag is written and goes nowhere near the page. Preview cards are
composed from your own colour, typeface and screenshot.
**What are AEO and GEO?**
Two names for one concern: whether answer engines and generative search
can read a page well enough to use it. AEO stands for answer engine
optimization and GEO for generative engine optimization, and the industry
has not settled on either term. Both start from a page that states what
it is in tags a machine can read without running scripts; the
[AEO and GEO page](/ai-seo/aeo-geo/) goes further.
**Are crawlers or AI agents blocked or slowed down?**
No. Nothing on this path blocks automated readers, throttles them, or
asks them to prove anything. A crawler or AI agent that requests a page
is served that page, and where prerendering is on it receives rendered
HTML instead of a shell.
**Does IndexNow tell Google about my pages?**
No. Google does not participate in IndexNow; it finds changed pages
through its own crawling and the sitemaps you give it. Submission reaches the
engines that do participate, and the scope is specific: discovered and
changed URLs only, no history imported, full-site submission manual.
**What else runs alongside it?**
Two more mechanisms of Bridge CDN are live. Prerendering serves rendered
HTML to crawlers and AI agents instead of a shell. IndexNow submission
tells the engines that participate about discovered and changed URLs;
enabling it imports no history, and full-site submission is manual.
--- https://bridge-seo.com/ai-seo/ ---
# AI SEO: what it fills in, and what it never touches
AI SEO fills the meta title and description, Open Graph tags, X card tags
and JSON-LD that a page is missing, and never touches a tag you wrote. A
page nobody wrote tags for still has to be described somewhere: a
search result needs a title and a line under it, and a pasted link needs
something to unfold into.
Many pages supply a title and stop there. It happens when a site outgrows the person who writes its metadata, and
the gap never shows up in a browser.
## What gets filled
[Diagram: A grid of pages against tag families: the title is present nearly everywhere, the rest is mostly absent]
Four families of tags are filled where a page lacks them, and one step follows them.
**Meta title and description.** The pair a search result is assembled
from. Where a page declares neither, both are written from what the page
actually says.
**Open Graph.** What a link preview reads. Without it, a link pasted into
a chat unfolds as a rectangle with a URL in it.
**Twitter and X cards.** The same intent, a different vocabulary, and a
separate set of tags. A page can preview correctly in one place and badly
in another for no other reason.
**JSON-LD.** Structured data in the form engines read for rich results.
Structured data already on the page stays where it is.
**Then IndexNow.** The step after the tags: when a page's tags change, that page can be announced to the engines that accept such
announcements. Google is not one of them.
Each family has a page of its own:
- **[Meta title and description](/ai-seo/meta-title-description/)** covers the pair a result is built from, and what filling the gap leaves alone.
- **[Open Graph](/ai-seo/open-graph/)** covers the tags a link preview reads, and what a missing set costs where you cannot see it.
- **[Twitter and X cards](/ai-seo/twitter-cards/)** covers the second vocabulary, and why falling back to Open Graph is not quite enough.
- **[JSON-LD](/ai-seo/structured-data/)** covers what is written where a page has no structured data, and what happens to markup it already carries.
- **[Preview images](/ai-seo/preview-images/)** covers the card composed from your colour, your typeface and a picture of the page.
## The boundary
Filling a gap and improving an existing line are different jobs, and only
the first one happens here.
A description written three years ago and long forgotten is treated as
authored, because it is, and is left exactly
as you left it. Every injected node carries `data-bb-seo="1"`, so what
came from you and what was filled in stays distinguishable later, in the
page source, without taking anyone's word for it.
Two settings decide what the model is allowed to know. A brand profile
holds the facts that do not change from page to page: the site's name,
its social handle, the postfix on a title. Those are copied rather than
guessed. A short account of what the site does and how its copy
should sound steers the wording, and is never published on any page.
[Diagram: Three values from the brand profile appear in the written tags unchanged: the same string stands at both ends of every line]
## Where it happens
Enrichment does not edit your site. It edits the copy of the page that is
assembled for machines at the edge, and your own HTML is left where it is,
byte for byte.
A person loading the page
sees what your server sent, generated tags included or not. A crawler
asking for the same address gets the assembled copy, and it is there that
a missing description or a missing card is filled. Nothing is written back
to your origin, so there is no migration to plan, no plugin to install and
nothing to undo if you switch the whole thing off.
It also explains why the result is the same on a site rendered in a
browser and on one rendered by a server. The assembled copy is built the
same way in both cases, and the tags are added to it at the same point.
## How the record is made
The model runs on our own hardware rather than behind a call out to a
third-party API.
Output is checked before it is stored, and a record that fails the check
is discarded rather than injected. The page keeps its gap instead of
receiving something malformed, which is the safer of the two outcomes:
a missing tag and a broken one do not fail the same way. A parser that
chokes on invalid JSON-LD can drop the whole block, including the markup
you wrote yourself, which was fine until something invalid landed next to
it.
## How to see what it did
Every generated node carries `data-bb-seo="1"`, so what was filled can be
checked in a browser. Fetch the page the way a
crawler does, read the head, and the attribute separates what came from
you from what was filled in. Nothing that lacks the attribute was
generated; nothing that carries it was ever yours.
Two more things follow from that. The first is that removing enrichment
removes exactly the marked nodes and nothing else, so the switch is reversible. The second is that an audit of your own tags stays
possible: filter the marked nodes out, and what remains is the state of
your metadata as you wrote it, gaps and all.
## Where the words come from
The niche goes by three names, AI SEO, AEO and GEO, and they are used
interchangeably more often than not. [AEO and GEO](/ai-seo/aeo-geo/) covers what the terms mean,
where the distinction is real, and where it is a distinction without a
difference.
## The signals underneath
Underneath the tags sit ordinary technical declarations, and a
findability problem is often one of those declared wrongly.
[Meta tags](/meta-tags/) covers what a title and description are read
for, and [duplication across a site](/meta-tags/duplicates/) covers
what it costs when the same pair appears on many pages at once.
## FAQ
**Does AI SEO rewrite tags I already wrote?**
No. The check is for absence, so injection fills what is missing and
nothing else. A description you wrote years ago and forgot about is kept
exactly as it stands, whether or not it is any good, because deciding that
it is not good enough would be a different job from filling a gap. A
generated tag appears only where that tag is absent.
**Do I need a plugin or a code change?**
Neither. Your site sits behind Bridge CDN, and enrichment happens to the
copy assembled for machines at the edge. That is why it works the same on
a site rendered in the browser as on one rendered by a server.
**Is it on by default?**
No. It stays off on each domain until you switch it on, and it is
available on Pro and above. The default is deliberate: enrichment changes
what machines read about your pages, and that is not a change anyone
should discover after the fact. Turning it on is a decision you make per
domain, not once for an account.
**What happens to a generated tag when I write my own?**
Yours takes over. The check is for absence, so a tag that exists is a tag
that is left alone, whether it was there yesterday or added this morning.
The generated node stops appearing once the gap it filled is gone.
**Can I tell which tags were generated?**
Yes, and without asking anyone. Generated nodes carry `data-bb-seo="1"`,
so fetching the page the way a crawler does and reading its head separates
them from your own. That marker is also what makes switching enrichment
off a clean operation: what disappears is
exactly the marked nodes, and your own tags are not part of the
transaction.
--- https://bridge-seo.com/hreflang/ ---
# Hreflang tags: what they declare, what they do not
Hreflang tells a search engine that several pages are versions of the same
thing intended for different audiences, and which audience each version is
for. It is a declaration about relationships, not an instruction to send
anybody anywhere, and much of the trouble people have with it comes from
expecting the second.
## What it actually declares
Hreflang declares a set of alternates. Each page in the set names every page in the set,
including itself, and states which language and optionally which region
each one is meant for.
That is the whole model. It does not redirect anybody, does not detect
where a visitor is, and does not choose a version on their behalf. What it
does is remove a search engine's ambiguity: without it, several pages with
similar content look like duplication, and one of them tends to be shown
everywhere. With it, they are understood as a set, and the engine has a
basis for showing the appropriate one to each searcher.
The consequence worth internalising early is that hreflang affects what
appears in results, and nothing about what happens after somebody clicks.
## The three places it can live
Hreflang can live in the page head, in a sitemap or in an `HTTP` header,
and the three are equivalent in meaning.
**In the page itself.** As elements in the document head. A common
choice, and the most expensive at scale, because every page must list every
alternate, and the total number of declarations across the site grows
with the square of the set.
**In an XML sitemap.** As annotations on each entry. This keeps the
declarations out of the pages entirely, which suits large sets and sites
where templates are awkward to change.
**In an HTTP header.** The only option for files that are not documents,
such as those distributed as downloads.
Choosing between them is an engineering decision rather than an SEO one.
What matters is that exactly one method is used consistently, because
maintaining two is how they come to disagree. The syntax of each, with
Google's own examples, is covered under
[hreflang implementation](/hreflang/implementation/).
## Reciprocity: the rule that makes it work
[Diagram: Three language versions of a page, each naming the other two and itself; the set holds only while every page agrees]
Every page in the set must name every other page, and each must name
itself. Break it on one page and the whole set stops being reliable.
The reason it exists is that a declaration nobody can verify is worthless.
If a page claims another is its alternate but that page makes no such
claim, the relationship is asserted by one party alone, and an engine has
no reason to believe it: anybody could point at anybody.
The practical consequence is that the set is only as good as its least
maintained member. Adding a new language means editing every existing page
in the set, not only the new one, and forgetting one of them breaks the
relationship for the whole group rather than for that page alone.
## Language and region codes
Language codes come from [`ISO 639-1`](https://www.iso.org/iso-639-language-code) and optional region codes from
[`ISO 3166-1`](https://www.iso.org/iso-3166-country-codes.html), two standards used together and often confused.
The language is expressed with a code from `ISO 639-1`. The region,
optionally, with one from `ISO 3166-1`, and it qualifies the language
rather than replacing it. A region on its own is not a valid declaration,
because a region does not tell an engine what language the page is in.
A frequent error here is inventing a region where a language would do. A
page in a language spoken across many countries usually wants the language
alone; adding a region narrows the declaration to that country and leaves
every other speaker outside the set, which is rarely what anybody intended.
Another is treating the region as a marketing decision. It
describes who the page is for in a way an engine can act on, and inventing
combinations that do not correspond to real usage produces a set that
matches nobody.
## The default entry
One entry in the set can be marked as the fallback for anybody the other
entries do not describe.
This exists because no set covers the world. A site with versions for three
languages has nothing to offer a searcher in a fourth, and without a
fallback the engine chooses for itself. Marking one version as the default
makes that choice yours.
It is optional, and it is worth having whenever the set does not aim to be
exhaustive, which is the usual case. The [default entry](/hreflang/x-default/)
has a page of its own: what it declares, and when it earns its place.
## Where hreflang stops
Hreflang does not redirect anybody, and it is regularly expected to do
things it does not do.
**It does not redirect.** Sending visitors to a regional version is a
separate decision, made by your application, and combining automatic
redirection with hreflang is a common way to make both worse.
[Redirects](/redirects/) have their own failure modes, and a locale
redirect layered on a declaration set hides both.
**It does not fix thin or duplicated content.** Pages that differ only in a
currency symbol are still nearly identical, and declaring them as
alternates does not change what they are.
**It does not apply across unrelated pages.** The set is versions of one
page. A category listing in one language and an article in another are not
alternates, however related they feel editorially.
How hreflang and canonical declarations work together is covered under
[hreflang and canonical](/hreflang/canonical/).
## What this section covers
The pages below take five parts separately: how to choose a
[URL structure for each country](/hreflang/url-structure/), how each of the three methods
is written, the errors that break a set
and how to find them, the default entry and whom
it catches, and the one rule that keeps hreflang compatible with
canonical declarations.
## FAQ
**What does hreflang do?**
It declares that several pages are versions of the same thing for different
audiences, so a search engine can show the appropriate one rather than
treating them as duplicates. It affects what appears in results and does
nothing at all after somebody clicks.
**Does hreflang redirect visitors?**
No. It is a declaration about relationships between pages, not an
instruction to send anybody anywhere. Redirecting visitors to a regional
version is a separate decision made by your application, and combining the
two carelessly tends to make both harder to reason about.
**Why does my hreflang not work?**
Often because the declaration is not reciprocal. Every page in the set
must name every other page and also itself, and a set where one member has
been forgotten breaks for the whole group rather than for that page alone.
**Do I need a region code?**
Only when the page is genuinely intended for one country. A language code
alone covers everybody speaking that language, while adding a region narrows
the declaration and leaves speakers elsewhere outside the set entirely,
which is rarely what anybody intended when they added it.
--- https://bridge-seo.com/meta-tags/ ---
# Title tags and meta descriptions: what really works
Two tags decide how your page is offered to somebody scanning a page of
results. One is a strong signal about what the page is; the other is a sales
line with no direct effect on position. Both are suggestions rather than
instructions. A third, the canonical, names which address the result
points to.
## What each of the two does
[Diagram: The title becomes the headline of a result and the description becomes the line beneath it]
The title names the page and the description offers a line under it;
they are routinely discussed together and they do different jobs.
**The title.** Written as `
` in the document head, it names the page. It
is used as the heading of your result, carries weight in how the page is
understood, and it is the single piece of
text most likely to decide whether somebody chooses your result over the one
above it. If you improve only one thing on a page, this is usually it.
**The description.** Written as ``, it does not name
the page; it sells it. It has no direct influence on where the page appears. What it can do is change how many people who see your result decide
to click it, which is worth real attention for that reason.
Everything else that lives in the head of a document is outside this
section. Instructions to crawlers about indexing, such as `meta robots`,
belong to the site in this network that covers indexing, and the tags
controlling how a link looks when shared on a social platform are covered
in this site's own [AI SEO section](/ai-seo/open-graph/). This section is
about the pair that appears in search results, and only that pair.
## They are suggestions, not instructions
A search engine may show text of its own choosing instead of the title or
description you wrote.
For the title, it may substitute something else: a heading from the page,
text from a link pointing at the page, or your site's name appended or
removed. For the description, substitution is not the exception but the
common case, because engines routinely assemble a snippet from whatever part
of the page matches the query being asked.
That behaviour is not a fault, and it is not a sign that your tag was
rejected. It follows from the engine answering a specific query: text you
wrote in advance cannot be tailored to a question you did not know would be
asked, while a passage lifted from the page can be. The practical
consequence is that a description is best treated as the version shown when
nothing better fits, which is a useful thing to control and a
smaller thing than it is usually assumed to be.
Rewriting is far more likely when the tag is generic, when it is stuffed
with terms, when it duplicates other pages, or when it simply does not
answer the query. Those are the levers you have, and they are the
same levers that make a tag worth reading in the first place.
All of that assumes a tag exists. Where none was written, something is
still assembled and shown, and none of these levers is in play at all:
[the pair written where it is missing](/ai-seo/meta-title-description/)
is a different problem from the pair written badly.
## Where the limit really is
The cut-off is measured in rendered width, not in characters: what gets
truncated depends on how much space the text occupies on the screen that
shows the result. Google sets no limit on description length and cuts the
snippet to fit; where that cut falls, a working target, and how to write the
part that survives are covered under
[meta description length](/meta-tags/meta-description/), and the same
account for titles under [title tags](/meta-tags/title-tag/).
What is stable is the technique. Put what matters at the front, so that
losing the tail loses nothing important. Read the tag as somebody scanning
twenty results would read it rather than as its author. And check what is
actually being displayed for your pages rather than what you wrote, since
those are frequently not the same text.
## Writing ones that get chosen
Titles and descriptions get chosen when they are specific, distinct and
front-loaded, and a short set of habits gets you there.
**Be specific before being clever.** A result competes with nine others that
are broadly about the same subject. Anything true and concrete about your
page beats a well-turned phrase that could sit on any of them.
**Do not repeat the site name in every title.** It occupies space in every
result, adds nothing after the first one somebody sees, and is often
appended by the engine anyway.
**Write the description as an offer, not a summary.** It should say what
somebody gets by clicking rather than restating the title in longer form.
Duplicating the title in the description wastes the second field entirely.
**Never write the same pair twice.** Identical tags across many pages are
among the strongest signals that neither is worth using, and
[what duplication costs](/meta-tags/duplicates/) is a page of its own.
On large sites it is usually a template's doing.
**Match what the page actually delivers.** A tag that oversells earns the
click and loses the visit, and the pattern is visible in behaviour long
before anybody notices it in a report.
## What this section covers
Six pages go further. One covers the title tag: how long it can be,
what truncation removes, and when Google shows a different title. One
covers meta description length: where the cut falls, a working target,
and how to write the part that survives. One covers how to find
duplicated titles and descriptions across a site, what the duplication
costs, and what to do about it, starting with the sources that produce it
wholesale. One covers the canonical tag: which of several
addresses for one page search should keep, and how to say so. One explains
why Google ignores [meta keywords](/meta-tags/meta-keywords/), and the last
covers [featured snippets](/meta-tags/featured-snippets/), the passage Google
lifts into an answer box.
## FAQ
**How long should a title tag be?**
Long enough to say something specific, and front-loaded so that truncation
costs you nothing important. The real limit is measured by rendered width
rather than by character count, it varies by context, and it changes over
time, so use a current figure with a date rather than a number remembered
from an article.
**Does the meta description affect rankings?**
Not directly. It influences whether somebody chooses your result once they
see it, which is worth optimising for its own sake. Treating it as a ranking
field leads to keyword-stuffed lines that read badly, get rewritten by the
engine anyway, and persuade nobody to click.
**Why does Google show different text than what I wrote?**
Because it is answering a specific query and can often find a passage on
your page that fits that query better than text written in advance. This is normal. Generic, stuffed, or duplicated tags make
substitution more likely; specific and relevant ones make it less so.
**Do I need a meta description on every page?**
Not strictly, and a generic description that could sit on any page is
worse than none, since the engine will assemble something from the page
regardless. Write them where
the click matters and where you can say something specific, and leave the
rest rather than filling a field for the sake of a green checkmark.
**What about meta keywords?**
`` is not used by major search engines and has not
been for many years.
Filling it costs nothing except the time spent and the confusion of anybody
who later reads the page source and assumes it means something. There is no
version of this field that helps.
--- https://bridge-seo.com/on-page/ ---
# On-page SEO: the signals a page sends about itself
On-page SEO is work on what a page says about itself through its own
markup and text: headings, address, the text behind its images, its links
and its place in the site's structure. A reader and a search engine meet
these signals first, and Google's advice for them is the one it gives for
all of Search: write for people.
## What counts as on-page
On-page signals are the ones you set inside the page and its address,
without asking anyone else. This section covers six of them.
**Headings.** The outline of the page, from the H1 down, that lets a reader
scan and lets a machine see the structure.
**The URL slug.** The readable part of the address that says what the page
is before anybody opens it.
**Alt text.** The text that stands in for an image when it cannot be seen,
read by screen readers and by search engines alike.
**Anchor text and internal links.** The words that link one page to another
and the paths they make through a site.
**Breadcrumbs.** The trail that shows where a page sits in the site's
hierarchy.
**The XML sitemap.** The list of addresses a site gives search engines so
nothing important is missed.
## What lives elsewhere on this site
Some page signals have sections of their own here. Titles, descriptions
and canonical tags are covered under [meta tags](/meta-tags/).
Machine-readable markup about the page, such as JSON-LD, is covered under
[structured data](/ai-seo/structured-data/). Language versions are covered
under [hreflang](/hreflang/), and moved addresses under
[redirects](/redirects/).
What this site does not cover is whether a page can be crawled or indexed
at all: robots rules, noindex and indexing status belong to
[BotBridge](https://botbridge.io/), which documents how pages reach
crawlers.
## What Google says about these signals
Google's guidance on these elements is consistently practical. Its SEO
Starter Guide, last updated on 2025-12-10 and checked on 2026-10-12, asks
you to "provide headings to help users navigate your pages," calls alt
text "a short, but descriptive piece of text that explains the relationship
between the image and your content," and describes link text as "the text
part of a link that you can see."
None of these is presented as a lever that moves a page by itself. Each
helps a reader understand the page, and helps Google understand it in the
same way. That is also Google's position on newer AI search features: its
guide to generative AI features, last updated on 2026-07-10, calls
optimizing for them "still SEO."
## How to approach the section
Start with what is broken rather than with what could be polished. A page
without an H1, images without alt text, slugs made of ID numbers and pages
no link points to are faults a reader notices. Fix those first, then refine
the wording.
Keep each signal true to the page. A heading that promises what the section
does not deliver, alt text stuffed with keywords and anchor text that says
"click here" all describe something other than what is there, and that is
the opposite of what these signals are for.
## A checklist for one page
A single page can be checked in a few minutes with nothing but the browser
and its view of the source. Work down the list and note what fails.
**One H1 that names the page.** Find the main heading in the source and
check that it says what the page is, and that lower headings follow in
order without skipping levels.
**A readable address.** Read the URL aloud. Words separated by hyphens pass;
a string of ID numbers or parameters does not.
**Alt text where images carry meaning.** Every image that informs has a
short description of what it shows in context; purely decorative images
have an empty one.
**Links in and out.** At least one other page on the site links here with
words that describe it, and the page links on to related pages the same
way.
**A place in the structure.** The page sits under a parent a reader can
reach, shown by a breadcrumb trail where the site uses one, and it appears
in the XML sitemap.
**Title and description present.** Both exist and describe this page, not
the site; the meta tags section covers how to write them.
Each item on this list has its own page in this section or the next one.
Run the list again after a redesign or a template change, since templates
are where these signals most often break across many pages at once.
## Fixing at the template, not the page
Most on-page faults are not written by hand; they are produced by a
template and repeated on every page that uses it. A theme that renders the
site name as the H1, a product template that leaves images without alt
text, a CMS that builds addresses from database IDs: each makes the same
mistake hundreds of times. Fix it once in the template and every page
inherits the fix, then check a few pages of each type to confirm the
change came through.
## What this section covers
The pages below take the on-page signals one at a time:
[headings and the H1](/on-page/headings/), [URL slugs](/on-page/url-slug/),
[alt text](/on-page/alt-text/), [anchor text](/on-page/anchor-text/),
[internal links](/on-page/internal-links/), [breadcrumbs](/on-page/breadcrumbs/)
and the [XML sitemap](/on-page/xml-sitemap/). Each one covers what the signal is, what Google and the web
standards say about it, and how to get it right on a real site.
## FAQ
**What is on-page SEO?**
On-page SEO is work on what a page says about itself: its headings, URL,
image alt text, links and place in the site structure. These signals are
set inside the page and its address, unlike links from other sites. Google's
advice for them is to make the page clear for people, which also makes it
clear for search.
**What is the difference between on-page and technical SEO?**
The line is blurry, and usage varies. On this site, on-page covers what a
page says about itself; whether it can be crawled and indexed at all is
covered by BotBridge, and metadata such as titles and structured data have
sections of their own. Most of the work overlaps in practice.
**What are the most important on-page elements?**
There is no published ranking of them. Start with what readers notice when
it is missing: a clear H1 and heading structure, readable URLs, alt text on
meaningful images, and links that connect related pages. Titles and
descriptions matter as much and are covered under meta tags.
**Does on-page SEO still matter with AI search?**
Google says so directly for its own search. Its guide to generative AI
features describes optimizing for them as "still SEO," with no special
markup or files required. A page that is clear to a reader, with honest
headings, text and links, is the page these features can use.
--- https://bridge-seo.com/redirects/ ---
# URL redirects for SEO: which code to use, and when
A redirect is a response telling the client that what it asked for lives
somewhere else. Which status code you send is not cosmetic: it states
whether the move is permanent, whether the client may reuse the answer, and
whether it should repeat its request the same way. Those three things
decide almost everything that follows.
## What a redirect actually is
[Diagram: A request follows one redirect and arrives; each further hop is another round trip before anything is served]
A redirect is a response, not a rule: the server answers with a status code in the
redirection class and a header naming the new location, and the client
decides what to do with that.
The semantics of each code are defined in [`RFC 9110`](https://www.rfc-editor.org/rfc/rfc9110), which is the useful
place to start because it is stable. Search engine behaviour built on top
of those semantics is not stable, changes without notice, and is documented
separately by each engine. Keeping the two apart in your head is worth the
effort: the protocol tells you what a code means, and an engine's
documentation tells you what that engine currently does about it.
## The two questions behind every code
Use `301` or `308` when the move is permanent and `302` or `307` when it is
not; of each pair, the newer code, `307` or `308`, is the one specified to
preserve the request method. Two questions settle which of the four applies.
| Code | Permanent | Request method preserved |
|---|---|---|
| `301` | Yes | Not required |
| `302` | No | Not required |
| `307` | No | Yes |
| `308` | Yes | Yes |
**Is this permanent?** A permanent redirect says the old address is
finished and the new one replaces it. A temporary one says the original
address is still the right one and something about right now is different.
The distinction matters because clients are permitted to remember a
permanent answer, and because search engines treat the two differently when
deciding which address to show.
**Must the request method survive?** Older codes were specified loosely
enough that clients commonly changed a form submission into a plain
retrieval when following them. Newer codes exist specifically to remove
that ambiguity, preserving whatever the original request was. If the
redirect can be reached by anything other than an ordinary page load, this
question is not academic.
The newer pair is covered in full under
[307 and 308 redirects](/redirects/307-308/). The choice is usually obvious
once the questions are asked in that order.
## What redirects cost
Every redirect is a round trip that produces no content, and that cost is
paid by the visitor.
A single one is negligible. A [chain](/redirects/chains/), where the
first address points to a second which points to a third, multiplies it,
and chains form without anybody deciding to make one: a rule added years ago, a canonicalisation
rule added since, and a marketing address added last month can compose into
three hops that nobody designed.
Chains also degrade differently from other faults. Clients follow a limited
number of hops before giving up, and different clients give up at different
points, so a chain that grows past somebody's limit fails for them and
works for everybody else. [Loops](/redirects/loops/) are the same fault taken
further and fail for everyone.
The rule that avoids both is to point at the final destination rather than
at the next step. Where a chain already exists, collapsing it is usually a
matter of editing the first rule rather than removing the later ones.
## Where redirects are configured
The same redirect can come from several places, and knowing which one
issued it is most of the diagnosis.
A web server can issue it from its own configuration. An application can
issue it in code. A content system can issue it from a database somebody
edits through an interface. A layer in front of your server can issue it
before the request ever arrives. In many sites, more than one of these is
active and none of the people responsible knows about the others.
When a redirect behaves unexpectedly, the useful first question is not what
it does but where it comes from. A rule you cannot find is usually being
issued by a layer you were not thinking about, and the way to establish
that is to ask each layer directly rather than to read configuration files
in the order you happen to remember them.
## Before you change canonicalisation
Changing which form of your domain is canonical and applying redirects in
bulk both deserve a warning ahead of the step rather than a note
afterwards.
**Changing which form of your domain is canonical.** The choice between a
bare domain and a prefixed one, or between protocols, moves every address
on the site at once. Done correctly it is invisible; done with a
rule that matches too broadly it produces a loop that takes the whole site
down rather than part of it. Test the rule against a handful of addresses
before it applies to all of them.
**Applying redirects in bulk.** A bulk change during a restructure carries the same shape
of risk. A mapping with one wrong entry sends one page somewhere odd; a
pattern with one wrong assumption sends thousands. Patterns are worth
writing, and they are worth testing against a sample that includes the
awkward cases: addresses with parameters, with unusual characters, and the
ones that were already redirects before you started.
Both changes are also the two where keeping a record matters most, because
the thing you will want in six months is the list of what used to be where.
## What this section covers
The pages below take the working decisions apart: the permanent 301 and
how to set one up, the temporary 302 and
when it is the honest answer, .htaccess rules on Apache and why their
default is that same 302, return and rewrite on nginx, which share it, moving a whole site to HTTPS
in one hop, the three ways WordPress issues redirects, the
meta refresh for pages whose server cannot be changed, JavaScript
redirects and why Google may not see them, the modern pair of codes that keeps a
request's method intact, finding and collapsing chains, and diagnosing
loops, where a rule fights itself until nothing arrives at all.
## FAQ
**Which redirect code should I use?**
Answer two questions: is the move permanent, and must the request method
survive. Permanent moves use a permanent code so clients and search engines
treat the new address as the replacement. Where the original address is
still correct and something temporary is happening, a temporary code is the
honest answer.
**Do redirects slow down a site?**
One is negligible; chains are not, because each hop is a round trip
producing no content. Chains also form by accumulation rather than by
design, as rules added at different times compose, which is why a periodic
check pays even when nothing has been reported.
**What happens if redirects form a loop?**
Clients follow a limited number of hops and then stop, so a loop fails for
everybody. A long chain fails for some clients and not others, depending on
where each gives up, which makes it harder to diagnose than a loop and
easier to leave in place.
**Where do I set up a redirect?**
In your web server's configuration, in your application's code, in a
content system's interface, or in a layer in front of your server. More
than one is often active at once, which is why identifying the source is
usually more of the work than writing the rule.
--- https://bridge-seo.com/tools/ ---
# SEO checks: what each reads and what it leaves alone
A short index of the checks this section explains: what each one reads, in
what order, and what the result can and cannot tell you. Each is paired
with the part of the site that covers the subject behind it in full.
## What is here
**Redirect trace.** How an address is followed from the first request to
the final destination, hop by hop, and why some hops never appear in a
server response. Written up in full under
[redirect tracing](/tools/redirect-trace/).
A check is deliberately narrow, and the narrowness is the method. One
that audited a page as a whole would produce a list nobody acts on,
whereas a check that answers a single question gives you something you can
do next. That is why every page in this section stays on one question, and
why each states what it does not report as plainly as what it does.
Narrow also means repeatable. A check worth trusting reads the same things
in the same order every time, so that two runs of it disagree only when
the site changed between them. The write-ups here describe checks in
exactly those terms: what is requested, what is read from the answer, and
at which step a difference means a fault rather than a preference. Reading
a check this way is slower than pasting an address into a form, and it
pays back the first time a result needs to be defended to somebody else:
you can say what was measured, not only what a page of results asserted.
## What these checks deliberately do not cover
Each check answers its own question and stops there.
None of them reads canonical declarations or crawler directives, even
though a page being checked carries both. Those subjects belong to other
sites in this network, and a check that reported on everything it happened
to see would be answering questions this site has not earned the right to
answer. The tags used when a link is shared are this site's subject, but
they are explained in the AI SEO section rather than checked here.
The explanations live in the sections on
[redirects](/redirects/) and [hreflang](/hreflang/), and each check links
back to the part of them it belongs with.
## FAQ
**Are these interactive tools?**
Not on this site. Each page here explains a check: what it reads, in what
order, and what the result settles. The subjects themselves are covered in
the sections on redirects and hreflang, and nothing here asks you to sign
up for anything.
--- https://bridge-seo.com/ai-seo/aeo-geo/ ---
# AEO vs GEO: two names for machine answers, compared
AEO aims to make a page the passage a system lifts and shows; GEO aims to make it one of the sources a system composes its own
answer from. The two are often used as synonyms, but they describe
different outcomes. Both are terms of the [AI SEO](/ai-seo/) niche, and
this page sets them side by side.
## The two in one table
Each term has its own page with the definition and sources:
[what is AEO](/ai-seo/aeo/) and [what GEO means](/ai-seo/geo/). Side by
side, they differ on every practical point except the page they start from.
| | AEO | GEO |
|---|---|---|
| Outcome sought | Your passage is shown as the answer | Your page is among the sources of a composed answer |
| What the system does | Extracts a passage | Synthesises from several sources |
| Where it shows | A quoted passage, often with a link | A paragraph that may not name its sources |
| How it fails | Visibly: another page is quoted | Quietly: your page is left out or paraphrased wrongly |
| What helps most | A direct answer that stands alone | Specific, attributed, consistent facts |
| Origin of the term | Industry usage, no single definition | A 2023 research paper (arXiv 2311.09735) |
## Extraction against synthesis
The practical split is the unit. Extraction works on a passage: one
paragraph has to answer the question cleanly without the paragraphs
around it. Synthesis works on facts: the system reads several sources,
reconciles them and writes its own text, so what matters is whether your
facts are specific enough to be used and consistent enough to be trusted.
## Which one applies to your page
Usually both, in different places. A page that answers a common question
in its opening sentence is a candidate for extraction; the same page,
with precise and sourced facts further down, is a candidate for synthesis.
The choice between the terms matters less than the two habits they point
to: answer early, and state facts in a form that can be checked.
## Where the distinction stops being useful
[Diagram: One system lifts a passage out of a single page; another composes an answer from several pages at once]
The two terms overlap heavily and are used interchangeably by most people
who use either. The distinction worth keeping is extraction against
synthesis: whether you are trying to be quoted, or trying to be one of the
sources behind a paragraph nobody attributes.
Everything else that gets attached to the acronyms is marketing around a
fairly small idea: that one is newer, that one is about chatbots and the
other about search.
## AEO vs SEO, and GEO vs SEO
Neither replaces SEO; each changes what a page competes for. SEO competes
for a position in a list of links, and the reader then chooses a result.
AEO competes to be the passage an answer engine lifts and shows as the
answer. GEO competes to be among the sources a generated answer is built
from, and to be represented accurately when it is.
All three start from the same page. A page that cannot be fetched or does
not say what it is about fails at all three, which is why the groundwork is
shared: content a machine receives, a direct answer early, and facts stated
where they can be read.
Google takes the same view for its own engine. Its guide to generative AI
features in Search, last updated on 2026-07-10 and checked on 2026-10-09,
says: "From Google Search's perspective, optimizing for generative AI
search is optimizing for the search experience, and thus still SEO." The
statement is about Google Search, including AI Overviews and AI Mode. For
other engines the terms still describe what a page competes for, which is
how this page uses them.
## They fail differently
Extraction fails visibly. A passage is lifted that answers a slightly
different question, or nothing is lifted at all and somebody else's page
is quoted instead. The remedy is structural: say the thing plainly, early,
and in a passage that survives being read on its own.
Synthesis fails quietly. Your page is one of five behind a paragraph that
names none of them, and where two of the five disagree, the composed
answer picks a side without telling anyone. You do not see this happen and
there is no listing to check. The remedy is not structural but factual:
be specific enough to be worth resolving toward, and consistent enough
that nothing on your own site argues with you.
Notice that the second remedy is mostly ordinary editorial discipline. It
is also the harder of the two, because it cannot be verified by looking.
## What neither of them promises
No arrangement of a page obliges an engine to use it.
Neither term is a lever: both describe how a page is treated once a
machine has it, and the machine's decision remains the machine's. A page
written for extraction is easier to extract from, which is a real
advantage and a modest one. No arrangement guarantees a citation.
What holds up is smaller. Pages that
state what they are, answer what they were asked, and declare their basic
facts in a form a program can read are the pages that can be used. The
rest is out of your hands.
## What either one asks of a page
Both assume the machine received the page. That is a larger assumption
than it sounds: a site that assembles its content in the browser sends a
shell first, and a reader that does not execute that step receives the
shell. Nothing written for extraction or synthesis survives being absent.
Both also assume the page says what it is about somewhere a machine
looks. Google notes that structured data "isn't required for generative AI
search," and its systems read the page itself. Readers that rely on
declared metadata are another matter: to them, a page with no description
and no [structured data](/ai-seo/structured-data/) is legible only through
its text, which is the gap AI SEO closes.
## FAQ
**Is GEO just a new name for AEO?**
Nearly. The useful split is extraction against synthesis: AEO is about a
passage being liftable as the answer, GEO about being among the sources an
answer is composed from. Most advice applies to both, and where a piece of
advice applies to only one of them, it is usually about structure rather
than about substance.
**Do I need to write differently for each?**
No. Writing that answers a question in its first sentence and stands alone
out of context serves both. The rest is the same editing you would do
anyway: cut the throat-clearing, put the claim before the reasoning, and
make each section survive being read without the one above it.
**Does either one replace search optimisation?**
No. A page still has to be found the ordinary way before any of this
applies. These terms describe what happens to a page after a machine
already has it, which makes them a layer on top of the older work rather
than a replacement for it. Sites that treat them as a replacement tend to
discover the omission slowly.
**Do I need to do anything technical for AEO or GEO?**
Less than the acronyms imply, and the little there is overlaps entirely
with ordinary hygiene. The page has to arrive whole at something that does
not run scripts, and it has to declare what it is. Neither requirement was
invented by answer engines; both were true for search engines first.
**Is any of this measurable?**
Partly, and less than you would like. Extraction leaves traces: you can
see a passage quoted. Synthesis usually leaves none, because the paragraph
that used your page names nobody. Anyone reporting a precise share of
generated answers is measuring a proxy, and the proxy deserves saying out
loud.
--- https://bridge-seo.com/ai-seo/aeo/ ---
# What is AEO? Answer engine optimization, plainly put
AEO, answer engine optimization, is work to make a page the answer a
system shows directly: a passage lifted into a search result, an
assistant's reply, a spoken answer. It has no single founding definition.
It is one of the terms of the [AI SEO](/ai-seo/) niche, and Google treats
it, for its own search, as SEO.
## What the term covers
AEO covers the case where a system answers instead of listing. The reader
asks a question and receives a reply, and the reply is built from a passage
on somebody's page. The work is to make your page the one that passage
comes from.
The unit that matters is the passage, not the page. A page can rank and
still not be quoted if no single part of it answers the question cleanly;
a modest page can be quoted if one paragraph does. That shifts attention
from the page as a whole to the sentences that open each section.
Google's guide to generative AI features in Search, last updated on
2026-07-10 and checked on 2026-10-12, describes AEO and GEO together as
"terms you may see used to describe work specifically focused on improving
visibility in AI search experiences." Beyond that, usage varies: some
writers use AEO for featured snippets and voice answers, others for any AI
assistant. This page uses it for being the passage a system shows.
## AEO and GEO
AEO and GEO are often used as synonyms, and the difference is worth
keeping. AEO aims at being extracted: the system takes your passage and
shows it. [GEO](/ai-seo/geo/) aims at being composed from: the system
writes its own answer and your page is one of its sources. A side-by-side
comparison of the two is covered under [AEO vs GEO](/ai-seo/aeo-geo/).
## What Google says about AEO
For Google Search, Google does not treat AEO as separate work. The same
guide says: "From Google Search's perspective, optimizing for generative AI
search is optimizing for the search experience, and thus still SEO." It
also advises reviewing third-party "AEO" or "GEO" advice against its own
guidance on evaluating SEO advice.
The guide is equally direct about one popular technique. It lists
"chunking" content among things you can ignore: "There's no requirement to
break your content into tiny pieces for AI to better understand it." So
AEO, read correctly, is not about cutting pages into fragments. It is about
making the important sentences say the important thing.
## What AEO asks of a page
AEO asks for answers that can stand on their own. A system lifting a
passage reads it without the paragraphs around it, so the passage has to
make sense alone.
**Answer in the first sentence.** Each section opens with the answer to
its own heading, and the explanation follows. A reader skimming, and a
system extracting, both get the point first.
**Make passages self-contained.** A paragraph that depends on "as
mentioned above" fails when it is lifted. Name the subject again rather
than pointing back to it.
**Use the question's own words.** A heading that matches how people ask
helps both the reader find the section and the system match the passage to
the question.
**Keep facts where they can be read.** A number in an image, or one that
only appears after a script runs, cannot be lifted as text.
These are habits of clear writing as much as of optimization, and they
help the person reading the page in full just as much.
## Where the title and description fit
Before any passage is read, a system sees what the page says about itself.
The [title tag](/meta-tags/title-tag/) and the meta description are that
first statement, and structured data adds the type of page and its place
in the site. A page missing them is harder to match to a question before
its text is even considered. Filling the missing ones, without touching
those somebody wrote, is the part of this work that AI SEO enrichment
covers.
## Checking whether a page is quoted
There is no single report of where a page has been quoted, so checking is
manual and approximate. Ask the questions your page is meant to answer, in
the systems your readers use, and note whether the answer comes from your
page, from another, or from none. Repeat after you change the page; one
check, on one day, says very little on its own.
For Google's AI features, its guide names a first-party source: the
Generative AI performance report in Search Console shows how content
performs in those features. For other systems, the honest record is the
one you keep yourself, with the date of each check, because answers change
as the systems and their sources do.
## What AEO does not promise
No arrangement of a page obliges a system to quote it. A clean passage
makes extraction possible and more likely; the choice stays with the
system, and it is made per question. Advice that promises a page will
become "the answer" promises something nobody controls, and a quote won
today can move to another page tomorrow when the question or the sources
change.
## FAQ
**What does AEO stand for?**
AEO stands for answer engine optimization: work to make a page the answer a
system shows directly, such as a passage in a search result or an
assistant's reply. The term has no single founding definition, and usage
varies between writers, but the common core is being the passage that is
lifted and shown.
**Is AEO the same as SEO?**
For Google Search, Google says that optimizing for its generative AI
features "is optimizing for the search experience, and thus still SEO."
The outcome differs: SEO competes for a place in a list of links, AEO for
being the passage a system shows. The groundwork, a readable page that
answers early, is the same.
**What is the difference between AEO and GEO?**
AEO aims at being extracted: a system lifts your passage and shows it. GEO
aims at being one of the sources a system composes its own answer from.
Both start from the same page qualities, and the two terms are often used
interchangeably, which is why a side-by-side comparison helps.
**How do I optimize for answer engines?**
Open each section with a direct answer to its heading, make paragraphs
readable on their own, use the words people actually ask with, and keep
facts in text rather than images. Google adds that you do not need to break
content into tiny pieces. None of this guarantees a quote; it makes one
possible.
--- https://bridge-seo.com/ai-seo/ai-overviews/ ---
# AI Overviews: how Google chooses the pages it links
AI Overviews are AI-generated summaries Google shows above its results for
some searches, linking the pages they draw on. A page becomes one of those
links the way it reaches ordinary results: Google says "There are no
additional requirements to appear in AI Overviews." It is a feature the
[AI SEO](/ai-seo/) niche is built around.
## What AI Overviews are
AI Overviews summarise a topic and point to sources. Google's page on AI
features and your website, last updated on 2025-12-10 and checked on
2026-10-12, describes them as helping "people get to the gist of a
complicated topic or question more quickly," providing "a jumping off point
to explore links to learn more."
They do not appear on every search. The same page says that "AI Overviews
are only shown when our systems determine that it is additive to classic
Search, and as such, often don't trigger." AI Mode, a separate experience
for longer, exploratory questions, works on the same principles but "may
use different models and techniques, so the set of responses and links they
show will vary."
## How the linked pages are chosen
Google builds the answer from its index, and it may search more than once.
Both features "may use a "query fan-out" technique," which Google
describes as issuing multiple related searches across subtopics and data
sources to develop a response. That lets them "display a wider and more diverse set of
helpful links" than a single search would.
That has one practical consequence. A page can be linked for a question it
does not target directly, because one of the related searches matched it.
Covering a subject well, rather than one phrasing of it, is what fan-out
rewards.
## What makes a page eligible
The requirements are the ones for Google Search itself. Google states: "To
be eligible to be shown as a supporting link in AI Overviews or AI Mode, a
page must be indexed and eligible to be shown in Google Search with a
snippet." It adds, as it does for all of Search, that meeting the
requirements "doesn't mean that Google will crawl, index, or serve its
content."
Its list of practices that continue to matter is the ordinary one: crawling
allowed, content findable through internal links, a good page experience,
important content in text, good images and video where they help, and
structured data that "matches the visible text on the page."
## What Google says you do not need
Google is explicit about what AI Overviews do not require. On its AI
features page: "You don't need to create new machine readable files, AI text
files, or markup to appear in these features. There's also no special
schema.org structured data that you need to add."
Its guide to generative AI features, last updated on 2026-07-10, goes
further and names specific ideas you can ignore for Google Search: an
[llms.txt](/ai-seo/llms-txt/) file, "chunking" content into small pieces,
rewriting text just for AI systems, and chasing inauthentic mentions. It
describes optimizing for these features as "still SEO", which is the same
conclusion reached from the other side under [GEO](/ai-seo/geo/).
## Where structured data fits
Structured data is not a requirement for AI Overviews, but it is not
irrelevant either. Google's guide says it "isn't required for generative AI
search," and in the same sentence advises continuing to use it "as part of
your overall SEO strategy, as it helps with being eligible for rich
results." The rule that applies everywhere applies here: markup has to
match what the reader sees.
What markup can and cannot do for a result's appearance is covered under
[rich results](/ai-seo/structured-data/rich-results/); what a page's
[structured data](/ai-seo/structured-data/) contains is covered on its own
page.
## How to see whether a page appears
Google's AI features page says sites appearing in AI Overviews and AI Mode
are counted in Search Console's overall search traffic, in the Performance
report under the "Web" search type. Its newer guide also points to a
Generative AI performance report for content in generative AI features.
Neither shows the text of the overview or which sentence came from which
page. For that, the only method is to search the questions your page
answers and read what appears, noting the date, since overviews change as
the index does.
## What this means for a page
Read together, Google's statements describe a short list. Make sure the
page can be crawled and indexed and is allowed a snippet. Answer the
questions it covers clearly and early, because a related search from the
fan-out may land on any section. Keep facts in text and keep structured
data in step with what the reader sees. Beyond that, the page competes as it
does in ordinary results, and the decision to link it stays with Google,
query by query.
## Limiting what appears
Google treats AI Overviews as part of Search, so the controls are the
search controls: crawling rules for Googlebot, and the directives that limit
what is shown from a page in results. How those directives work, and how
crawlers and bots receive pages, is covered by
[BotBridge](https://botbridge.io/); this site does not cover them.
## FAQ
**How do I get my page into AI Overviews?**
Through the same route as Google's results. Google says a page must be
indexed and eligible to show with a snippet, and that there are "no
additional requirements" or special optimizations. Covering a subject well
helps, because AI Overviews may run several related searches and link pages
matched by any of them. Nothing guarantees inclusion.
**Can I keep my pages out of AI Overviews?**
Google treats AI Overviews as part of Search, so the controls are the
search ones: crawling rules for Googlebot and the directives that limit
what is shown from a page. Those directives also affect ordinary results.
They are covered by BotBridge, which documents how pages reach crawlers.
**Do AI Overviews use structured data?**
Google says structured data "isn't required for generative AI search" and
that there is no special schema.org markup to add. It still recommends
structured data for rich results, and asks that markup match the visible
text. So markup is neither a requirement for AI Overviews nor a reason to
remove it.
**Does llms.txt help with AI Overviews?**
No. Google's guide to generative AI features says Google Search does not
use llms.txt files, and that keeping one "will neither harm nor help" a
site's visibility or rankings in Google Search. It is fine to keep one for
other systems that read it, but not for AI Overviews.
--- https://bridge-seo.com/ai-seo/geo/ ---
# Generative engine optimization (GEO): what it means
Generative engine optimization, GEO, is work to make a page one of the
sources a generative system draws on when it composes an answer, and to
have it represented accurately there. The term comes from a 2023 paper.
It is one of the terms of the [AI SEO](/ai-seo/) niche,
and Google treats it, for its own search, as SEO.
## Where the term comes from
The name comes from a paper, "GEO: Generative Engine Optimization", by
Pranjal Aggarwal and five co-authors, first posted to arXiv on 2023-11-16
and accepted at the KDD 2024 conference. Its authors describe generative
engines as systems that answer queries "by synthesizing information from
multiple sources and summarizing them using LLMs", and they introduce GEO
as a way "to aid content creators in improving their content visibility in
generative engine responses."
The paper's starting point is the publisher's problem. In a list of links,
a page is either shown or not. In a composed answer, a page can be used
without being visible, paraphrased without being named, or left out
entirely, and its authors note that creators "have little to no control
over when and how their content is displayed."
## What the research measured
The paper tested changes to source pages against a benchmark of its own,
GEO-bench, a set of queries across several domains with the web sources
that answer them. It reports that the methods it studied "can boost
visibility by up to 40% in generative engine responses."
Two qualifications come with that figure. It is a result measured on the
authors' benchmark and visibility metrics, not a forecast for any site.
And the paper itself adds that "the efficacy of these strategies varies
across domains." The useful conclusion is narrower than the headline
number: how a source is written affects whether a generator uses it, and
the effect depends on the subject.
## What GEO asks of a page
GEO asks a page to be worth composing from: specific, attributable and
consistent. A generator assembling an answer from several sources has to
choose between them where they differ, and a vague or self-contradicting
page gives it nothing to choose.
**Say the specific thing.** A concrete figure, name or condition is more
useful to a composed answer than a general claim that every source
repeats.
**Attribute what you state.** A fact with its source and date can be
checked and reused; an unsourced one competes with every other unsourced
version of it.
**Keep the site consistent.** Two pages that disagree about the same fact
make the site an unreliable source on that fact, whichever page is read.
**Make the text available.** A page whose content a system cannot read, or
reads only after scripts run, cannot be composed from at all.
None of these is a trick. They are the same qualities that make a page
useful to a person who reads it.
## What Google says about GEO
Google addresses the term directly. Its guide to generative AI features
in Search, last updated on 2026-07-10 and checked on 2026-10-12, defines
AEO and GEO as "terms you may see used to describe work specifically
focused on improving visibility in AI search experiences," and adds: "From
Google Search's perspective, optimizing for generative AI search is
optimizing for the search experience, and thus still SEO."
Its page on AI features and your website, last updated on 2025-12-10, says
of AI Overviews and AI Mode: "There are no additional requirements to
appear in AI Overviews or AI Mode, nor other special optimizations
necessary." Among the things it says you do not need are special files and
markup; that includes an [llms.txt](/ai-seo/llms-txt/) file. How
Google chooses the pages it links there is covered under
[AI Overviews](/ai-seo/ai-overviews/).
That position covers Google Search, including its AI features. Other
generative systems choose their sources in their own ways, which is why the
term remains a useful description of what a page competes for outside
Google.
## Measuring it
Measuring GEO is harder than measuring a ranking, because a composed answer
can use a page without linking it. For Google's own AI features there is a
first-party source: its guide points to the Generative AI performance
report in Search Console as the way to see how content performs in
generative AI features on Search and Discover.
For other systems there is no shared equivalent, and the same guide warns
against tools that "promise ranking success or claim to use "internal"
Google metrics." A sensible approach is modest: check the questions your
pages answer in the systems your readers use, note whether and how the
page is cited, and repeat after changes rather than reading a single
check as a trend.
## GEO next to AEO
GEO and AEO are often used as synonyms, but they describe different
outcomes. [AEO](/ai-seo/aeo/) aims at being the passage an answer engine lifts and shows;
GEO aims at being among the sources behind a composed answer. A side-by-side
comparison of the two, and of each against ordinary SEO, is covered under
[AEO vs GEO](/ai-seo/aeo-geo/).
## Where tags fit in
A composed answer is built from text, and a page's own description of
itself helps a system decide what the page is before it reads further.
Titles, descriptions and [structured data](/ai-seo/structured-data/) are
that description, and a page without them leaves the system to infer it.
Filling the ones that are missing, without touching the ones somebody
wrote, is the part of this work that AI SEO enrichment does.
## FAQ
**What is GEO in SEO?**
GEO, generative engine optimization, is work to make a page one of the
sources a generative system uses when it composes an answer, and to have it
represented accurately. The term comes from a 2023 research paper. For
Google Search, Google describes this work as part of SEO rather than
something separate from it.
**Is GEO different from SEO?**
For Google Search, Google says no: optimizing for its generative AI
features "is optimizing for the search experience, and thus still SEO,"
with no special optimizations required. The difference is in the outcome
being competed for: a place among the sources of a composed answer rather
than a position in a list of links.
**Who coined the term GEO?**
It was introduced in "GEO: Generative Engine Optimization" by Pranjal
Aggarwal and co-authors, posted to arXiv in November 2023 and accepted at
KDD 2024. The paper defines generative engines, proposes ways to make
sources more visible in their answers, and measures them on its own
benchmark.
**Does GEO really increase visibility by 40%?**
The paper reports that its methods "can boost visibility by up to 40%" on
its own benchmark, and notes that results vary across domains. That is an
experimental finding under the authors' conditions, not a figure any site
can expect. Treat it as evidence that writing affects use, not as a target.
--- https://bridge-seo.com/ai-seo/llms-txt/ ---
# llms.txt: what the file declares, and who reads it
`llms.txt` is a plain markdown file at the root of a site that offers
language models the site's own account of what it is and where its
important pages are. It sits beside the tags covered under
[AI SEO](/ai-seo/). Advice to add one rarely
says what the file claims or who has agreed to read it.
## What the file is
`llms.txt` is a plain markdown file at the root of a site, proposed in
September 2024 by Jeremy Howard of Answer.AI and specified at
[llmstxt.org](https://llmstxt.org/). It offers a language model a short, curated account of
what the site is and where its important pages are, written for a reader
that arrives with a question rather than a browser.
[Diagram: An llms.txt file is offered to a reader, which may take it or ignore it]
The shape is deliberately small: an `H1` with the site name, an optional
blockquote summary, then lists of links with one line of description
each. A companion file, `llms-full.txt`, carries the full text of those
pages in one place, so a model can read the substance without following
every link.
That is the whole proposal. It is a summary a site writes about itself,
in a format a machine does not have to guess at.
## What it is not
`llms.txt` is not a standard in the sense that a status code or a `rel` value is
a standard. No standards body has ratified it, and no major engine has
published a commitment to consume it. Treat any claim that a particular
assistant reads your `llms.txt` as unverified unless that assistant says
so in its own documentation.
It is also not an access-control file. It grants nothing and forbids
nothing. A crawler that ignores it is not violating anything, and a
crawler that reads it gains no permission it did not already have.
## How it differs from robots.txt
`robots.txt` governs **access**: which paths a crawler may fetch.
`llms.txt` offers **content**: what the site is about and where its
important pages are. A site can have both, and they never contradict each
other; the full comparison is under
[llms.txt vs robots.txt](/ai-seo/llms-txt/vs-robots-txt/).
## What it can plausibly do
The file can do three modest things, each narrow and each real.
**It removes a guess.** A model that would otherwise infer what a site is
from whatever page it happened to land on gets a description the site
wrote on purpose.
**It shortens a path.** Where the important pages are named in one place,
a reader that follows links does not have to find them by crawling.
**It survives your navigation.** A menu is designed for people and often
hides structure behind interaction. A flat list does not.
None of that is a ranking claim, and none of it is measurable in the way
a redirect or a tag is. The file is a statement of intent, and its value
depends entirely on whether the reader chose to read it.
## The ceiling
`llms.txt` cannot make a page eligible for anything. A page that is not fetchable
does not become fetchable because a summary mentions it.
Nor does it substitute for the tags a page is missing: a model reading
a page directly reads that page's own head, and an `llms.txt` two
directories up does not fill a missing description. Those gaps are the
subject of the tag families, not of this file. And there is
no console, no status and no confirmation that any reader used it, which
is why claims about its effect deserve more scepticism than most.
Google has said so for its search engine directly. Its guide to generative
AI features in Search, last updated on 2026-07-10 and checked on
2026-10-09, lists llms.txt among things you can ignore: Google Search
"doesn't use them," and maintaining one "will neither harm nor help your
site's visibility or rankings in Google Search, as Google Search ignores
them." The same guide adds that it is "completely fine" to keep one for
other services or systems that use it.
## Whether your site needs one
A file this small is rarely wrong to have, and the question is better
asked as what it would say.
A site whose structure is genuinely hard to infer from its navigation
gains the most: documentation with many entry points, a knowledge base
whose hubs are not obvious, a product whose name does not describe it.
A site of eight pages with a plain menu gains little, and a stale file
that names pages you deleted is worse than none.
Writing it is a short editorial job, not an engineering one. The cost of
keeping it accurate is the real cost, and it is the reason a file added
once and forgotten stops being useful within a release or two.
## What this site does with it
This site publishes its own `llms.txt` and `llms-full.txt`, and they are
generated from the same sources as the pages, so they do not quietly
drift from what the pages say.
Nothing on this site generates an `llms.txt` for your site. Enrichment
fills missing tags in the copy served to machines and does not write
files at your root. Where a page here describes what is shipped, that is
the list, and this file is not on it.
## FAQ
**Does an llms.txt affect rankings?**
Not in Google. Google's guide to generative AI search, updated on
2026-07-10, says Google Search ignores llms.txt files and that keeping one
"will neither harm nor help" a site's visibility or rankings. The file
addresses readers that answer questions rather than rank pages, so keep it
for those readers, not as a ranking lever.
**Which assistants actually read it?**
No major provider has published a commitment to read it. Some smaller tools and readers do, and say so in their own
documentation. Any list claiming a large assistant consumes the file is
worth checking against that assistant's documentation before you repeat
it, because such lists circulate faster than the confirmations behind
them.
**Is it a replacement for robots.txt?**
No, and the two do not overlap at any point. `robots.txt` governs access,
has an established format, and is honoured by the major crawlers.
`llms.txt` offers content, is a proposal rather than a format anyone has
ratified, and is honoured by nobody in particular. Keep both if you want
both, and expect neither to do the other's job.
**Should llms-full.txt include every page?**
It should include the pages you would hand somebody who asked what the
site covers, which on a small site may well be all of them. A dump of
everything is easy to generate and hard to read, and on a large site it
defeats the point of writing a summary in the first place. The test is
whether a reader could answer a real question from it.
**How often should it be updated?**
Whenever the pages it names change enough to make it wrong. A file that
lists deleted pages misdescribes the site more confidently than no file
at all, which is why generating it from the same source as the pages is
worth the small effort.
--- https://bridge-seo.com/ai-seo/llms-txt/examples/ ---
# llms.txt examples: two good files and one done badly
Three [`llms.txt`](/ai-seo/llms-txt/) files, read line by line: a small
product site, a documentation set, and the same file done badly. All
three are for invented sites, so nothing here pretends to describe
somebody's real one. The [format](/ai-seo/llms-txt/format/) is short, and
a blank file is still hard to start.
## A small product site
A small product site still needs the file, because its name rarely says
what it does. Eight pages, a plain menu, one thing to sell: the
temptation is to skip the file entirely.
[Diagram: Three shapes of an llms.txt file: a short one, a sectioned one, and a dump of every URL that defeats the point]
```text
# Example Analytics
> Server-side analytics for teams that cannot ship
> a browser tag.
## Product
- [What it does](/product/): the measurement model,
and what it cannot see
- [Setup](/setup/): the two integration paths, with
the tradeoff of each
- [Pricing](/pricing/): what each plan includes,
and where the limits are
## Reference
- [API](/api/): endpoints, auth, and rate limits
- [Data retention](/retention/): how long each
record is kept, and why
```
Five links out of eight pages. The three left out are the ones a reader
with a question would not need: an about page, a contact form, a terms
page. Leaving them out is the selection that gives the file its value.
The notes carry a second job that is easy to miss. Each one names a
limit as well as a subject: what it cannot see, the tradeoff of each,
where the limits are, and why. A reader deciding between two links is
usually deciding which one will disappoint it less, and a note that
admits a boundary answers that faster than a note that praises the
page.
Notice what the blockquote does. "Server-side analytics" places the
category; "for teams that cannot ship a browser tag" places the reader.
A summary that stopped at the category would have described half the
sites in the category.
## A documentation set
For a documentation set, the mistake to avoid is mirroring the sidebar.
Sixty pages, a sidebar, and a structure that makes sense only if you can
already see the whole of it: this is the case the format was designed for.
```text
# Example DB
> An embedded database with a SQL interface and no
> server process.
## Start here
- [Install](/docs/install/): platforms, and what
each package contains
- [First query](/docs/first-query/): the shortest
path to a working read
## Guides
- [Transactions](/docs/transactions/): isolation
levels, and which is default
- [Migrations](/docs/migrations/): the file format,
and rollback behaviour
- [Backups](/docs/backups/): what a backup contains,
and what it omits
## Optional
- [Internals](/docs/internals/): storage layout, WAL
- [Changelog](/changelog/): releases, dated
```
Sixty pages became seven links, grouped by what a reader would be trying
to do. `Start here` and `Guides` are tasks; the sidebar's own categories
were probably nouns.
The grouping is the editorial work, and it is the part that cannot be
generated. A build step can list every page and keep the list accurate,
but only somebody who knows the product can say that installing and
running a first query are one journey while transactions and backups are
another. Where the list is generated, the section names and the notes are
still worth writing by hand once and reviewing when the shape of the
product changes.
The `Optional` section earns its place here. Internals and a changelog
are skippable by a reader answering a practical question, and
labelling them honestly means the reader drops the right two rather than
a guess.
## The same file, done badly
The file below is valid and close to useless. It parses, it sits at the
right path, and nothing in it is malformed.
```text
# Example DB - The Fastest Embedded Database
> The best embedded database for modern teams.
> Blazing fast, developer first, trusted by thousands.
## Links
- [Home](/)
- [Docs](/docs/)
- [Blog](/blog/)
- [About](/about/)
- [Careers](/careers/)
- [Privacy](/privacy/)
## Optional
- [Install](/docs/install/)
- [First query](/docs/first-query/)
```
Four things went wrong, and each is common.
**The H1 carries a slogan.** A reader quoting the site now quotes a
superlative, which is worse for the site than a plain name.
**The summary makes claims it cannot support.** "Trusted by thousands"
gives a reader nothing to verify and reads as noise.
**The notes are gone.** Every link is a category label, so the reader
learns only what the URL already said.
**`Optional` holds the two pages that matter most.** A reader honouring
the label drops exactly the pages the site most wanted read.
## Reading your own file back
Read the file as if you had never seen the site, then answer three
questions from it alone. The test takes a minute and needs no tooling.
What does this site do? If the blockquote does not answer it, rewrite the
blockquote before touching anything else.
Which link would I follow first, given a real question? If two links look
interchangeable, the notes are doing no work.
What would I still not know? That gap is either a page you forgot to list
or a page you have not written, and both are worth knowing.
## What ages and what does not
The blockquote rarely goes stale, because what a site is changes slowly.
The link list goes stale constantly, because pages move and get renamed
and get retired, and nothing warns you.
That asymmetry is the argument for generating the list from the same
source as the pages. A summary written by hand is a second description of
the site that has to be kept in step with the first, and the failure mode
is silent: the file still parses, still fetches, and now names pages that
are gone.
Where hand-writing is unavoidable, put the file where a deploy would
notice it, and read it back on the same schedule you review the
navigation.
There is one more thing worth checking on that schedule, and it takes
seconds: fetch the file the way a reader would, from outside your
network, with a plain client. A file that answers from your machine and
returns a challenge or a redirect to everybody else has stopped existing
for the readers it was written for, and nothing in the file itself would
ever tell you.
## FAQ
**Can I copy one of these files?**
Copy the shape, not the content. The sections, the note lengths and the
use of `Optional` transfer to any site; the words do not, because the
whole value of the file is that it describes this site rather than a
category. A copied summary is a generic one, and a generic summary is the
guess the file exists to replace.
**How many links is too many?**
There is no limit in the format, and the practical ceiling is the point
at which a reader stops being able to choose. The two working files above
list five and seven links. If a section needs thirty, the honest
move is usually to link a hub page and let it do the enumerating.
**Should the notes repeat keywords?**
No. The file is read by something that already has the page text
available if it wants it, so a keyword in a note buys nothing and costs
clarity. Write the note so that a reader choosing between two links can
tell which one answers the question in front of it.
**Do I need llms-full.txt if my site is small?**
Rarely. Its purpose is to save a reader many fetches, and a site of eight
pages does not present that problem. Where the pages are long or the
reader is likely to want several at once, it starts to earn its place,
and it should be generated rather than maintained.
**What if my sections do not fit these names?**
Use your own. Only `Optional` has a reserved meaning, and every other
heading is yours to name. Naming them after what a reader is trying to do
works better than naming them after how the site is organised internally,
because the internal organisation is the thing the reader cannot see and
was never meant to learn.
--- https://bridge-seo.com/ai-seo/llms-txt/format/ ---
# llms.txt format: the four parts, in order, and why
An [`llms.txt`](/ai-seo/llms-txt/) file has four parts in a fixed order:
an H1 with the site's name, a blockquote summary, sections of links with
one-line notes, and an `Optional` section a reader may skip. Only the H1
is required. So few rules leave room for invention, which is what makes
one file useful and another noise.
## The shape, in one look
A conforming file is markdown, and it has four parts in a fixed order,
with room for a line or two of free prose between the summary and the
first section.
Only the first is required.
[Diagram: The four parts of an llms.txt file in reading order; the Optional section is the one a reader may drop]
```text
# Site name
> One sentence on what this is
Any prose that helps a reader place the site.
## Docs
- [Page name](/url): one line on what the page answers
- [Another page](/url): one line
## Optional
- [Deep reference](/url): may be skipped when context is short
```
| Part | Required | What it is |
|---|---|---|
| `# Site name` | yes | an H1, and nothing above it |
| `> One sentence` | no | a blockquote, directly after the H1 |
| Free prose | no | plain markdown, no headings |
| `## Section` | no | H2 sections holding link lists |
| `## Optional` | no | the one section name with a reserved meaning |
The order matters because the file is read top to bottom by something
that may stop early. The name, then the sentence, then the links.
## The H1 and the blockquote
The H1 is the site's name, not a slogan and not a keyword line. A reader
that quotes you will quote this.
The blockquote directly under it is the one sentence you would say if
somebody asked what the site is. It is optional in the specification and
close to mandatory in practice: without it, a reader has to infer the
subject from the link list, which is the guessing the file exists to
prevent.
Keep it declarative. A summary written as a promise reads as marketing to
a person and as an unsupported claim to a machine, and neither is what
you wanted from it.
## The link sections
Sections are ordinary H2 headings, and their names are yours: `Docs`,
`Guides`, `Reference`, whatever the site actually has. Each holds a flat
list of links, and each link carries a short note after a colon.
The note is the part people skip, and it is the part that does the work.
Compare a bare link with the same link carrying one:
```text
- [Pricing](/pricing/)
- [Pricing](/pricing/): what each plan includes, and where the limits are
```
The first tells a reader nothing it could not guess from the URL. The
second tells it whether to follow the link at all, which is the only decision the reader is making.
Write the note as an answer to "what would I learn here", not as a
category label. A short phrase is usually enough. A reader
choosing between six links reads six notes, so a note that repeats the
link text spends the reader's attention and returns nothing.
Order the sections the way you would explain the site out loud, not the
way the navigation happens to be built. The file is read from the top,
and a reader that stops halfway should have stopped in the right place.
## The Optional section
One section name is reserved: `## Optional`. Links under it may be
skipped by a reader working with a short context window.
This is the only prioritisation the format offers, and it is worth using
honestly. Put the pages that a reader can do without there: deep
reference, changelogs, archives. Do not use it as a second list of things
you would like read anyway, because a reader that trusts the label will
drop them first, and a reader that does not trust it gains nothing.
If everything on your site is essential, the file does not need the
section. An empty `Optional` heading is worse than none.
## llms-full.txt
The companion file carries the full text of the pages, in one document,
in the order the summary lists them. It exists because following twenty
links costs a reader twenty fetches, and a single document costs one.
Two things make it useful rather than large. It should contain the same
pages the summary names, so the two cannot disagree. And it should be
generated from the same source as the pages themselves, because a
hand-maintained copy drifts within a release and then misdescribes the
site with confidence.
## Serving it
The file lives at the root: `/llms.txt`, alongside `robots.txt`. Serve it
as `text/plain` or `text/markdown`, and do not put it behind a redirect
chain or a bot challenge, since the readers that would use it are
precisely the ones least likely to survive either.
There is no registration step, no submission, and nothing to verify. The
file is either fetchable or it is not.
That also means nothing warns you when it breaks. A deploy that changes
how static files are served, a rule that starts challenging unfamiliar
user agents, a rename that leaves the old path returning a redirect: any
of these takes the file out of use quietly, because nothing was watching
it. Fetching it the way a reader would, occasionally, is the whole of the
available monitoring.
## Where files go wrong
**Links that do not resolve.** A summary naming deleted pages describes
the site worse than no summary at all.
**Notes copied from meta descriptions.** Those are written to be clicked;
these are written to be chosen between. The register is different.
**A wall of every URL.** The value is in the selection. A file that lists
everything has made no selection and saved the reader nothing.
**Marketing voice.** Superlatives read as noise to a reader that cannot
verify them, and the file has no place to put evidence.
**A file nobody owns.** The summary describes the site, so it goes stale
exactly when the site changes, which is when everyone is busy. Generating
it from the same source as the pages removes the problem rather than
scheduling it.
One file done badly is read line by line among the
[llms.txt examples](/ai-seo/llms-txt/examples/).
## FAQ
**Is any of this enforced, and does it have to be markdown?**
Nothing enforces it: there is no validator with authority, no registry
and no error to receive, so a malformed file is read as whatever the
reader makes of it. The specification says markdown for a practical
reason, because headings and lists are the structure readers already
parse, and the shape is the only thing a reader can rely on.
**How long should it be?**
Short enough that a reader arriving with a question can take in the whole
of it, which for most sites means well under a hundred lines. If a
section needs its own explanation, that explanation belongs on a page
that the file links to rather than inside the summary itself.
**Can I list pages that are not indexed?**
You can, and it is occasionally right: a page blocked from search may
still be the best answer to a question. Be aware that you are asking a
reader to fetch something the rest of your setup keeps out of results,
and make sure that is what you meant rather than an oversight.
**What goes in the notes for a large documentation set?**
Group by what a reader would be trying to do, not by how your navigation
is organised. Half a dozen sections with honest one-line notes beat a
faithful mirror of a sidebar, because the sidebar was designed for
somebody who can already see the whole of it.
--- https://bridge-seo.com/ai-seo/llms-txt/vs-robots-txt/ ---
# llms.txt vs robots.txt: two files that never overlap
The two files never overlap: `robots.txt` decides which paths a crawler
may fetch, and [`llms.txt`](/ai-seo/llms-txt/) describes what the site is
and which of its pages matter. Both are plain text at the root and both
address machines, and that is the entire resemblance; having one says
nothing about needing the other.
## What each one is for
`robots.txt` is about **access**. It tells a crawler which paths it may
fetch. It is subtractive by nature, it has an established format that
predates most of the web anyone remembers, and the major crawlers honour
it as a matter of long practice. The depth of it belongs to
[BotBridge](https://botbridge.io/), which documents the request path.
[Diagram: robots.txt governs access and llms.txt offers content; the two files do not overlap]
`llms.txt` is about **content**. It offers a reader the site's own
account of what it is and where its important pages are. Nothing about
it is enforced: the [format](/ai-seo/llms-txt/format/) is a proposal, and honouring it is a choice
each reader makes for itself.
One says what may be fetched. The other says what is worth reading. A
file cannot do both, and neither tries.
## Where the confusion comes from
Three surface facts do the damage: same location, same file type, same
audience in the loosest sense. Add an era in which everybody is being told
to "do something about AI crawlers", and two unrelated files end up in
the same sentence.
The other source is the name. `llms.txt` echoes `robots.txt` closely
enough that people assume a family resemblance and then assume a shared
mechanism. There is none. The naming is a convention borrowed for
recognisability, not a statement that the files work alike.
## What robots.txt does and llms.txt cannot
`robots.txt` can keep a crawler out of a path. Nothing in `llms.txt` withholds
anything; listing a page there does not open it, and omitting a page does
not close it.
It is honoured by the major crawlers. Decades of practice have made
compliance the default among the major ones, and `llms.txt` has nothing
comparable behind it.
It is unambiguous about intent. A disallowed path is a statement with an
agreed meaning, and both sides know what was claimed.
## What llms.txt does and robots.txt cannot
`llms.txt` describes rather than restricts. A crawler reading `robots.txt` learns
where it may go and nothing about what it would find; the summary is the
only place a site says what it is in its own words.
It orders things. A list written by the site puts the important pages
first, which no access rule can express.
Nothing in `llms.txt` is enforced, and nothing about it can be checked
the way an access rule can. The file is an offer, and its value depends on
the reader choosing to take it.
## Using both
`robots.txt` and `llms.txt` never contradict each other, because they
answer different questions. A page can be disallowed and still be described, though that
combination is usually a mistake worth noticing: it tells a reader the
page matters and then refuses it the page.
The one real interaction is the reverse. A file that is itself blocked
cannot be read, and an `llms.txt` sitting behind a rule that challenges
unfamiliar clients has stopped existing for exactly the readers it was
written for. Fetching it the way an outside reader would is the whole of
the available check.
Beyond that, treat them as independent. Changing one is never a reason to
change the other.
## The question that matters
The useful questions are what nobody should fetch and what you would say
if a machine asked what the site is; "which file do I need" assumes a
choice the two files do not offer.
**What should nobody fetch?** That is an access decision, it is old, it
has an established answer, and it belongs in `robots.txt`.
**What would I say if a machine asked what this site is?** That is an
editorial decision, it is new, it has no enforcement behind it, and it
belongs in `llms.txt`.
A site that answers only the first is invisible in its own words. A site
that answers only the second has described pages it may not have meant to
open. Most sites should answer both, and the answers are written by
different people on different days.
## What neither of them fixes
Neither file fills a missing tag. A page with no description still has no
description, whichever files sit at the root, because a reader arriving
at that page reads that page's own head. Those gaps are the subject of
[the tag families](/ai-seo/) rather than of either file here.
Neither one reports on itself. `robots.txt` at least has established
tooling around it; `llms.txt` has no console, no status, and no
confirmation that any reader used it.
And neither is a ranking lever. Claims that adding one improved
visibility are the kind that cannot be checked, which is reason enough to
treat them carefully.
## FAQ
**If I have robots.txt, do I still need llms.txt?**
The two files answer different questions, so having one says nothing about needing
the other. The test is whether a reader could tell what your site is from
its navigation alone. Where the structure is genuinely hard to infer, the
summary earns its place; where eight pages sit behind a plain menu, it
adds little beyond a file to keep accurate.
**Can llms.txt block AI crawlers?**
No. It withholds nothing and grants nothing, and a reader that ignores it
breaks no rule. Access decisions belong in `robots.txt`, and the
behaviour of individual crawlers on the request path is documented on
BotBridge, because that is the layer that owns it.
**Should the two files list the same pages?**
There is no reason they would. One names paths to keep out of, the other
names pages worth reading, and a well-run site usually has many pages in
neither list. Trying to keep them in step is work that buys nothing and
creates a second thing to go stale.
**Does robots.txt affect whether llms.txt is read?**
Only in one direction, and it is worth knowing: a rule that blocks or
challenges the client fetching `/llms.txt` makes the file unreadable to
the readers it was written for, and nothing anywhere warns you that it
happened. The reverse never occurs, since `llms.txt` carries no authority
over what any crawler may fetch and never claimed any.
**Which one should I write first?**
The access file, because getting it wrong has consequences and getting
the summary wrong does not. Once you know what should not be fetched, the
summary becomes easier to write anyway: you are describing what remains,
which is the part you wanted read.
--- https://bridge-seo.com/ai-seo/meta-title-description/ ---
# Meta titles and descriptions, written where missing
Where a page has no meta description, [AI SEO](/ai-seo/) writes one from
what the page says; where it has no title, it writes a title too, and tags
you wrote stay as they are. A page often has a title and no description.
The title arrived from the CMS almost by accident; the description had to
be written by somebody who decided to write it, page by page, and past a
certain size of site nobody decided. The pair gets shown either way. The title becomes the line somebody
clicks; the description becomes the sentence under it, when the engine
chooses to use one. Where neither was written, both are still assembled
by something that has never read your brief.
## What each one is read for
[Diagram: The title becomes the headline of a result and the description becomes the line beneath it]
The title is a name, not a summary. `` is the string a result list
shows, the label a browser tab carries, and the text a bookmark keeps. It is read
by people scanning twenty of them at once, which is why the specific words
at the front matter more than the ones at the end.
`` is not a ranking field, and has not been one
for a long time. It is an
argument for the click, and the engine is free to ignore it and assemble
its own line from the page instead. That freedom is often mistaken for the
tag being pointless. It is not: a page with nothing to offer gives the
engine nothing to prefer.
## What happens where neither is declared
The engine writes both anyway. It has to show something, so it takes the
first heading it trusts and a passage that looks relevant to the query.
The result is usually adequate and occasionally wrong in a way nobody
notices until somebody searches for the page by name.
Enrichment fills the gap from what the page actually says. Where a title
exists, it stays; where a description was never written, one is written
from the page's own content, not invented around it.
## What the model is given
The model is given two inputs, the page itself and your brand profile, and the difference between them is the whole design.
The first is the page itself. A description is written from what the page
actually says, which is why the result reads like a summary of your
content rather than a summary of your industry. Nothing is pulled from a
competitor, a directory or a general idea of what a business like yours
probably does.
The second is your brand profile: the site's name, the social handle, the
postfix that belongs on a title. Those are facts you set once, and they
are copied, never worked out. It matters which side of the line each
input sits on. The parts of a tag that have to be exact are the parts
nobody is generating, and the parts being generated are the ones where
wording is a matter of judgement rather than fact.
A short account of what the site does and how its copy should sound can
steer the wording. It is read while the tag is written and never published
on the page.
## When the page itself says little
Generation describes what is there. A page with three sentences on it
gives a generated description three sentences to work from, and the result
will be as thin as its source.
Nothing here invents a benefit the page does not claim, a
statistic it does not carry or an audience it never named. Where a page is
thin, the honest outcome is a thin description, and the fix is on the page
rather than in its head.
The same applies to a page that says several unrelated things. A
description has room for one idea. Where the page holds four, the
generated line picks the one the page leads with, which is usually the
right choice and occasionally a sign that the page needs splitting.
## What it does not touch
A title you wrote is kept exactly as it stands, including the ones you
regret: the absent gets written, and the present stays yours.
The same brand profile feeds the
[Open Graph](/ai-seo/open-graph/) set, which is why a page filled in
both places does not contradict itself. The brand postfix is the one place
the two meet. A generated title can end
with the site's name because the brand profile says so. That is a fact you
set, applied verbatim, not a rewrite of anything.
## The two are not the same job
Generating a title and generating a description are different problems,
though the two sit next to each other in the head and are usually
discussed as a pair.
A title is close to a fact. It names the page, it has to match what the
page is called elsewhere on the site, and it carries the postfix you set.
There is little room for interpretation, which is why so much of
it comes from the brand profile rather than from the model.
A description is closer to an argument. It has to say why this page and
not another, in one sentence, to somebody who has not arrived yet. That is
a matter of wording rather than of fact, and it is where a generated line
earns its place: not because it beats a written one, but because the
alternative on a page that has none is whatever the engine improvises.
## FAQ
**Will a generated description outrank a written one?**
Neither one ranks. A description is not a ranking field and has not been
one for a long time. What it influences is whether somebody chooses your
result once they can already see it, which is worth having for its own
sake. Treating the tag as a lever on position is the mistake; treating it
as the sentence that argues for the click is not.
**What if the engine rewrites it anyway?**
It often will, and that is normal behaviour, not a fault. Engines substitute
their own line when they judge the declared one a poor match for the
query. Generic or [duplicated text](/meta-tags/duplicates/) makes
substitution more likely, and specific text makes it less so. A page with no description gives the
engine nothing to prefer, so substitution there is not merely likely but
the default.
**Does the postfix count as changing my title?**
No. It is appended to titles that are generated here, never to one you
wrote yourself. The postfix is a fact from your brand profile and not
an edit, and it is applied at the point a title is written from nothing.
A title you wrote keeps whatever ending you gave it, including none.
**Can a generated description be wrong about the page?**
It can be thin, and it can pick the wrong idea from a page that holds
several. What it does not do is state something the page never claimed.
The source is your own content, so an inaccuracy in the description is
almost always an inaccuracy already sitting in the copy above it.
**What happens when I edit the page?**
The description was written from the page as it stood. Substantial edits
make it less accurate over time, in the same way a hand-written one drifts
after a rewrite nobody re-checked. The tag stays under your control: fill
in your own, and the generated line steps aside for good.
--- https://bridge-seo.com/ai-seo/og-image-size/ ---
# OG image size: the working numbers and what matters
The working size for an `og:image` is 1200 by 630 pixels, a ratio of
roughly 1.91:1, as checked on 2026-09-01, and a composed card from
[AI SEO](/ai-seo/) is built at exactly that size. The number is a moving
convention, and two smaller decisions, the ratio and the declaration, do
more work than the number itself.
## The working size, with its date
As checked on 2026-09-01, the working size for an `og:image` is 1200 by
630 pixels, an aspect ratio of roughly 1.91:1.
The date is part of the answer. Platform requirements change without
notice and without ceremony, which is why this page carries one number
with a date instead of a table of many without one; a figure nobody
re-checks is a rumour with formatting. If you are reading this long
after that date, treat the digits as a starting point and the method
below as the answer.
Of the two halves, the ratio matters more than the pixel count. A
sharper image at the same proportions ages well as screens improve; an
image at the wrong proportions loses its edges today, at every
resolution, for the reason
[the composition page](/ai-seo/preview-images/) explains: platforms
scale a matching shape and cut a mismatched one.
## What other sizes actually get
A wrong size is scaled, cropped or dropped, and nothing reports an error, which is the trouble.
An oversized image is scaled down and survives. An image at a different
ratio is cropped toward the slot it lands in, edges first, and whatever
lived near those edges, a logo in the corner, the end of a headline,
simply leaves the frame. A small image is the worst case: platforms
that consider it too small fall back to a plain link, and the page
previews as if it had declared nothing at all.
None of this is visible from your own browser, because your page does
not render its own card. It shows up where links are pasted, which is
also where it is cheapest to check.
## Declaring the size beside the address
Declare the size with `og:image:width` and `og:image:height` beside the address; Open Graph has these companion properties for the purpose:
```html
```
The pair does not change what the image is. It tells a platform what
the image is before the file arrives, and the practical effect is on
the first share: a consumer that knows the dimensions can lay the card
out immediately instead of fetching, measuring and reflowing. The
[Open Graph page](/ai-seo/open-graph/) places these two among the tags
worth carrying; here it is enough that they should agree with the file.
A declaration that contradicts the actual pixels is worse than none,
because there is no rule for which side a reader believes.
## Minimums, weights, and how to check them
Every platform sets a floor below which an image is ignored and a
weight above which it is refused, and this page will not print those
figures, because they are the ones that move most and are checked
least.
The method is worth more. The platforms that read `og:image` publish
their own card documentation and their own preview tools; the figure
that matters is the one in those documents on the day you ship, for the
platforms your audience actually uses. A practical floor holds without
any table: an image already at 1200 by 630 clears the documented
minimums of the major platforms by a wide margin, which is one more
argument for the working size over a clever small one.
## The file behind the number
Size is settled at the moment of export, and two habits keep it
settled. Export the card as its own file at exactly the working size,
instead of pointing `og:image` at whatever hero image the page already
uses: page images are sized for the page, and the crop that follows is
the one described above. And keep the address stable, because platforms
cache fetched cards; a renamed file forces every earlier share to
re-resolve, while a replaced file at the same address updates on their
schedule, not yours. Neither habit needs a figure, and both outlive
whichever figure is current.
## One file for every platform
One 1200 by 630 file serves almost every platform, and the exception has its own vocabulary.
A single 1200 by 630 file serves the platforms that read Open Graph,
which is most of them. The separate case is the one described under
[Twitter and X cards](/ai-seo/twitter-cards/): a second vocabulary with
its own frames, where the shape you request changes how the same image
is presented. A page that declares both vocabularies can still point
them at one file; what it cannot do is make one file look right in a
frame built for a different shape, so check the two previews
separately.
Where a page declares no image at all, a card can be
[composed](/ai-seo/preview-images/) at exactly the working size, which
is one reason the composed cards need no size decisions from anybody.
## FAQ
**What size should an og:image be?**
1200 by 630 pixels, at roughly 1.91:1, as checked on 2026-09-01. The
ratio is the half worth protecting: it decides whether platforms scale
your image or cut it. If you are far from that date, verify the figure
against the platform documentation before shipping anything expensive.
**Can one image serve every platform?**
Usually. The Open Graph readers share the same expectations closely
enough for one 1200 by 630 file to serve them all. The Twitter and X
card vocabulary frames images its own way, so a page declaring both
should be previewed in both places, even when a single file stands
behind the two declarations.
**What happens if my image is smaller than a platform's minimum?**
The platform ignores it, and the link previews as though no image was
declared: a bare card, or whatever the consumer scrapes on its own.
Nothing warns you. Shipping at the working size clears documented
minimums with room to spare, which is cheaper than tracking every
floor separately.
**Do og:image:width and og:image:height have to match the file?**
Yes, or they should be absent. The declaration exists so a consumer
can lay out the card before fetching the file; when the two disagree,
different readers resolve the conflict differently, and the result is
the kind of inconsistency nobody can debug from the page itself.
**Does image size affect where the page ranks?**
There is no evidence that it does, and nobody credible has claimed a
ranking effect on the record. The size decides how the link looks when
a person shares it, which influences whether people click a share.
That is an effect on readers, not on rankings, and it is the honest
reason to get the size right.
--- https://bridge-seo.com/ai-seo/open-graph/ ---
# Open Graph tags: the link that unfolds into nothing
Open Graph tags, often called OG tags, decide what a pasted link unfolds
into: four or five
properties in the head of the page, led by `og:title`, `og:description`
and `og:image`. Where they are missing, [AI SEO](/ai-seo/) writes them
from the page and your brand profile, and tags you declared stay as they
are.
Without them, somebody pastes your URL into a chat and sees a grey
rectangle with the address in it.
That is the whole of the problem, and it is invisible from your own site.
## Which tags are read
`og:title` and `og:description` do what their names suggest, though the
platform decides how much of each to show. `og:url` is the canonical
address the preview points at. `og:site_name` is the small line naming the
publisher. `og:image` is the one that changes whether anybody looks at the
preview at all.
There is also `og:type`, rarely interesting and occasionally decisive: an
article and a website are previewed differently in some places.
## What a missing set costs
Nothing measurable, and that is why it survives. A link with no
preview is still a working link. It is simply less likely to be opened
than the one below it that shows a title, a sentence and a picture.
The cost lands where you cannot see it: in conversations you were not part
of, about pages you did not know were being shared. No report records the
link that nobody opened.
## A partial set is its own problem
[Diagram: A page declares only the title; the platform fills the rest of the preview by guessing, so the same link looks different in different places]
A half-present set is harder to notice than an absent one.
A page carrying `og:title` and nothing else previews as a line of text
with an empty space where the picture belongs. Some platforms fill that
space by guessing: they take the first sizeable image in the document,
which might be a logo, a banner or somebody's avatar. Others leave it
blank. The same link therefore looks different depending on where it was
pasted, and the version you checked is not necessarily the version anyone
else saw.
This is why filling the set matters more than filling any single tag in
it. A preview is assembled from a group, and a group with a hole in it
falls back to whatever the platform improvises.
## Where the preview is decided, and when
The tags are read once, by whatever fetched the page first, and the result
is usually kept.
Most platforms cache what they found. A link shared before the tags
existed keeps its empty preview in that platform's store, and adding the
tags afterwards does not travel back to the messages already sent. Some
platforms expose a way to ask for a refresh; others simply expire the
entry on their own schedule.
The practical consequence concerns order, not urgency. Tags added
before a page starts circulating are read on the first fetch and cost
nothing to correct. Tags added after it has been shared widely fix the
next share and not the last hundred. It is the ordinary argument for
filling gaps quietly and early, not in response to somebody
noticing.
## The address the preview points at
`og:url` is the one tag in the set that is not about appearance.
A page is often reachable at more than one address: with tracking
parameters, with and without a trailing slash, through a campaign
redirect. Each of those is a separate link as far as a
sharing platform is concerned, and the counts, where a platform keeps them, are separate too.
Declaring the canonical address collapses them back into one.
Where the tag is missing it is derived: the canonical address is a fact
the page already carries, with nothing to compose.
## What gets filled
Where a page declares no Open Graph tags, they are written from what the
page says and from the brand profile: the site's name, the social handle,
a default image where one is set. Tags you declared stay as they are.
The half-present case works the same way, tag by tag rather than set by
set. A page carrying `og:title` and nothing else keeps that title exactly
as written and receives only the members of the group it was missing. This
is the more useful behaviour of the two, because the partial set is the
common one: somebody added a title years ago, the rest was never filled
in, and the preview has been assembling itself out of guesses ever since.
Where no image exists at all, a card can be composed at 1200 by 630 from
your brand colour, your typeface and a picture of the page, which
[preview images](/ai-seo/preview-images/) covers on its own. That part
stays off until you ask for it, and a page that already has an image keeps
it.
## FAQ
**Do I need Open Graph if I already have Twitter cards?**
Yes. They are separate vocabularies, not two spellings of one
thing, as [Twitter and X cards](/ai-seo/twitter-cards/) covers. Some
readers fall back from the Twitter set to Open Graph where the former is
absent, but the fallback is a courtesy, not a rule, and it does not
run in the other direction.
**What size should the image be?**
1200 by 630 pixels, and a composed card is built at exactly that; the
number, its date and the ratio behind it are under
[OG image size](/ai-seo/og-image-size/). Smaller images are often shown as
a thumbnail beside the text rather than above it, which changes the
preview from something people look at into something they scroll past.
**Will it replace the og:image I set myself?**
No. Composition applies only where the tag is absent, and a declared
image keeps its place whatever anyone thinks of it. The gap being
answered is a missing tag, never a mediocre one; the longer answer,
dimensions and age included, lives under
[preview images](/ai-seo/preview-images/).
**Why does my link still preview as blank after I fixed it?**
Because the platform is showing you what it cached on its first fetch.
Adding tags does not reach back into messages already sent. Where the
platform offers a refresh, asking for one updates its copy; otherwise the
old preview stands until the entry expires on that platform's own
schedule.
**Which of these tags are generated and which are derived?**
`og:title` and `og:description` are written from the page. `og:site_name`
comes from your brand profile verbatim. `og:url` is derived from the
canonical address the page already declares, because that is a fact rather
than a judgement. `og:type`
follows what the page already is. `og:image` is the profile's default image where one is set;
a composed card stays off until you enable it.
--- https://bridge-seo.com/ai-seo/preview-images/ ---
# Preview images: a card composed rather than imagined
A preview image is the 1200 by 630 card a link unfolds into.
[AI SEO](/ai-seo/) can compose one for a page that has none, from your
brand colour, your typeface and a screenshot of the page itself, and the
feature stays off until you switch it on.
A link with no image previews as a grey box, and most pages have none,
because somebody would have had to make one for each of them. The gap is
arithmetic: one card per page, multiplied by every page a site has ever
published.
## What goes into it
[Diagram: A brand colour, a typeface and a picture of the page are composed into a card of 1200 by 630 pixels]
A card is composed from three inputs, all yours. The brand colour, given as a hex value, becomes
the card's background. The typeface sets the title. A picture of the page
itself sits in a window frame beneath the title, so the card shows the
thing it links to, not a stock abstraction.
The window frame can be drawn in the Windows or the macOS style. It sounds
like a detail until the card is sitting in a feed next to four others.
The screenshot is the part that does the work. A card built around a
picture of the actual page tells somebody scrolling past what they would
be opening, which is more than a logo on a coloured field can do. It also
ages honestly: when the page changes enough to make its own card wrong,
the wrongness is visible instead of hiding behind an abstraction that was
never specific in the first place.
## Composed rather than imagined
Nothing in the card is invented. The colour is the one you set, the words
come from the page's own title, and the image is a screenshot rather than
an illustration of what the page might be about.
A generated illustration could show something the page does not contain;
a composed card shows only what the page and your settings already hold.
## Why 1200 by 630
1200 by 630 is the shape most platforms crop toward, and cropping is the operation
that decides what survives. The number and the date it was checked are
under [OG image size](/ai-seo/og-image-size/).
A card is almost never shown at the size it was made. It is scaled into
whatever slot the reader has, and where the aspect ratio does not match,
the difference is taken off the edges rather than added to them. Composing
at the ratio platforms expect means the crop is a scale rather than a
cut, and the parts you cared about are still in the frame.
There is a second reason to pick one size and keep it. A set of cards at
different ratios previews inconsistently across the same feed, and the
inconsistency reads as carelessness even to somebody who could not say
what was wrong. One shape, applied everywhere, is the cheaper kind of
polish.
## Where composition stops
A composed card cannot know what the page is for. It shows the page, and the page has
to carry its own meaning.
A screenshot of a page that is mostly body text reads, at card size, as a
grey rectangle with a lighter grey rectangle inside it. Pages built from
the same template produce cards that resemble each other, because the
thing being photographed resembles itself. Neither is a fault in the
composition; both are the page showing through.
Where a page deserves a picture made for it, a composed card is not the
better option and was never meant to be. What it is better than is the
grey box, and that is the comparison that applies to the several hundred
pages nobody was ever going to art-direct.
## Which tags the card lands in
A composed card is a file, and a file does nothing until something points
at it.
The address goes into `og:image`, which is what most readers look for and
what [Open Graph](/ai-seo/open-graph/) covers, and into
`twitter:image`, which is what the
[second vocabulary](/ai-seo/twitter-cards/) looks for. The
description of the picture goes into `og:image:alt` and
`twitter:image:alt`. All four are filled only where the page declares
none, on the same rule as everything else here.
Dimensions are declared alongside the address where the reader accepts
them. That spares the reader from fetching the file to discover its shape,
and it is the difference between a preview that appears at once and one
that appears a moment late, after the layout has already settled without
it.
## What stays off
This is the one part that does not run until you switch it on, per domain.
A page that already declares an image keeps it, and no card is composed
for it.
The result is written into the image tags of the copy served to machines,
so the preview appears where previews are read, and the page a person
opens is unchanged.
## FAQ
**Can I use my own template instead?**
Not today. The card is composed from the brand colour, the typeface and
the screenshot, and the frame style is the choice on offer. The narrowness
is deliberate rather than temporary shorthand: a template system is a
design tool, and this is a gap-filler for pages that were never going to
get design attention in the first place.
**Will it replace an image I already set?**
No. Composition applies only where the page declares none. A page carrying
an image keeps it whatever its dimensions, whatever its age and whatever
anyone thinks of it. The absence of a tag is the condition being answered
here, and a poor image is not an absence.
**Does the screenshot update when the page changes?**
It reflects the page as it was captured. A page that changes shape will
want a fresh one, in the same way a hand-made card goes stale after a
redesign. The difference is that nobody has to remember which pages have
cards, because the card belongs to the page, not to a folder
somebody maintains.
**What happens on a page that is mostly text?**
The card shows a page that is mostly text. It will be plainer than one
built around an illustration, and it will still be a card with your
colour, your typeface and your title on it rather than a grey box with a
URL. Where that is not good enough, the answer is an image of your own,
and the composed one steps aside the moment you declare it.
**Does the card carry alt text?**
Yes, written from the same page the card was built from and declared in
`og:image:alt` and `twitter:image:alt`, only where the page declares none.
A preview image without a description reads to a screen reader as silence
where the picture was.
--- https://bridge-seo.com/ai-seo/structured-data/ ---
# Schema markup: JSON-LD written where there is none
Structured data is a block of JSON-LD in a page's markup that states what
the page is in a shape a program does not have to infer. Where a page has
none, [AI SEO](/ai-seo/) writes it, and markup already on the page stays.
Nothing breaks when structured data is missing, which is most of the
reason it stays missing. No error appears, nothing looks wrong in a
browser, and the page goes on being read by machines that have to work out
what it is instead of being told. A machine reading prose has to decide
what the page is about; a machine reading a declaration only has to
believe it or not.
## What it is, in one paragraph
A block of JSON in the page's markup, conventionally in the head, naming
a type from a shared
vocabulary and filling in its fields. It says the same kind of thing as
the [title and description](/ai-seo/meta-title-description/), in a
form that does not have to be read as prose first. An `Organization` says who publishes
the site. A `WebSite` names the site itself. A `BreadcrumbList` says where
the page sits. A `WebPage` or `BlogPosting` says what this particular page
is.
None of it is a promise about how a result will look. It is a statement of
fact in a form that can be read without parsing prose.
## What is written where none exists
The organisation and the site come from the brand profile, so they are the
same on every page and correct by construction rather than by luck. The
breadcrumb comes from the path. The page type comes from what the page is.
A home page is not given a breadcrumb, because it has nowhere to come
from. That kind of small correctness is the whole reason to generate this
rather than paste it.
The type is the part most often got wrong by hand. A tool page is not an
article and a listing is not a blog post, and declaring the wrong one is
worse than declaring nothing: it invites the reader to expect fields the
page will never have, such as an author or a publication date on something
that has neither. Deriving the type from what the page is avoids the
question rather than answering it badly.
## Why a broken block is worse than none
A broken block is usually discarded whole, so invalidity can cost a page
more than absence does: absence only costs it the chance to say something.
A parser meets one malformed block and has no obliging way to recover
half of it. The usual outcome is that the block is discarded whole, and
where your own markup sat in the same block, it goes with it. A page that
described itself correctly until a stray comma arrived now describes
itself not at all, and nothing on the page looks different.
This is the reason a record is checked before it is stored rather than
after it is injected. A record that fails the check is dropped, and the
page keeps its gap. Given the choice between saying nothing and saying
something a parser will reject, the first is plainly the safer of the two,
and it is the one that leaves your existing markup untouched.
## How the blocks refer to each other
[Diagram: Three blocks on a page point at one organisation defined once, instead of each restating it and drifting out of step]
Several blocks on one page are not several unrelated statements. They form
a graph, and `@id` is what stitches it together.
An `Organization` is defined once, with a name and a logo, and every page
that needs a publisher points at that identifier rather than restating it.
The alternative, which is common enough on sites that pasted their markup
in by hand, is a publisher block repeated on every page and drifting out
of step with itself: a name updated here, a logo path stale there, and a
reader with no way to know which copy to believe.
References are worth checking in the same breath as validity. A pointer to
an identifier that nothing defines parses cleanly and says nothing, which
is the quietest failure in the set.
## What must not be marked up
Structured data is supposed to describe what the page actually contains. A
review the page does not show, a price it does not state, a rating nobody
gave: marking those up is not an optimisation but a misrepresentation, and
engines treat it as one.
Generation stays on the right side of that line by construction rather
than by good intentions. The record is written from the page in front of
it, so it can only describe what is there. Where the page does not claim
something, there is nothing for the block to say about it.
## What happens to markup you already have
Markup you already have stays. Structured data already on the page is left alone, and nothing
is merged into it. Two blocks describing different things are normal and
expected; a generated block never edits a written one.
That includes markup a plugin put there and nobody has looked at since.
Whether it is good is not the question being asked. The question is only
whether the page declares a given thing at all, and where it does, the
answer is to leave it be.
## FAQ
**Will this give me rich results?**
It makes them possible where they were not, and no markup obliges an
engine to show one. Engines also retire features: Google stopped showing
FAQ rich results on 2026-05-07, as covered under
[FAQ schema](/ai-seo/structured-data/faq-schema/). Its guide to generative AI search adds that
structured data "isn't required" there, while it still helps with being
eligible for rich results.
**What if my structured data is wrong?**
Then it stays wrong. Nothing here validates or corrects what you wrote:
a page with a broken block keeps it, and a generated block is never
merged into it. Checking your own markup remains your affair, exactly
as it was before enrichment existed.
**Does the vocabulary version matter?**
Rarely in practice. The types in play here have been stable for years, and
the fields that matter for them have not moved. Version differences bite
at the edges of the vocabulary rather than at `Organization`, `WebSite`
and `BreadcrumbList`, which is most of what a page needs to say about
itself.
**Can I have two blocks describing the same thing?**
You can, and it is usually a sign of two systems writing into one head
without knowing about each other. Nothing generated here creates that
situation, because the check is for absence. Where duplication already
exists, it is worth resolving on the page: a reader confronted with two
answers has no rule that tells it which one you meant.
**How do I check what the page declares?**
Read the source. Every generated node carries `data-bb-seo="1"`, so the
blocks written for you and the blocks you wrote are distinguishable
without guessing. Beyond that, the usual validators take a URL and report
what they parsed, which is the only account that matters.
--- https://bridge-seo.com/ai-seo/structured-data/faq-schema/ ---
# FAQ schema in 2026: markup without the rich result
FAQ schema is `FAQPage` markup from the schema.org vocabulary, usually
written as JSON-LD, that labels a page's questions and answers for
machines. Google stopped showing FAQ rich results on May 7, 2026, so the
markup no longer changes how a result looks there. It remains one of the
[structured data](/ai-seo/structured-data/) types a page can declare.
## What FAQPage markup is
schema.org defines the type in one line: "A FAQPage is a WebPage presenting
one or more "Frequently asked questions"". Each question is a `Question`
with a `name`, and its answer is an `Answer` attached as `acceptedAnswer`.
This page carries such a block, built from the questions at its end.
Shortened to one question, it reads:
```json
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [{
"@type": "Question",
"name": "Is FAQ schema still worth adding?",
"acceptedAnswer": {
"@type": "Answer",
"text": "…"
}
}]
}
```
The full version is in the page's source, beside the other blocks this
site declares.
## What Google changed
Google withdrew the FAQ rich result in three steps, each recorded on its
documentation updates page, checked on 2026-10-09.
| When | What Google recorded |
|---|---|
| September 2023 | The FAQ documentation was updated "to state that the feature is only shown for well-known, authoritative government and health websites." |
| May 8, 2026 | "This feature will no longer appear in Google Search starting May 7, 2026." |
| June 15, 2026 | "Removed documentation for the FAQ rich result feature… The FAQ rich result feature is no longer shown in Google Search results." |
The old documentation address now leads to that last entry. For most sites
the expanded questions under a result had already gone in 2023; since May
2026 they are gone for every site. Which other
[rich results](/ai-seo/structured-data/rich-results/) remain, and who
decides when they show, is covered separately.
## What the markup still does
The markup still describes the page accurately, and nothing in Google's
announcements treats it as an error. What stopped is a display feature,
not support for the vocabulary: `FAQPage` remains a schema.org type, and
structured data in general is still part of Google's advice.
Google's guide to generative AI features in Search, last updated on
2026-07-10, puts the general position in two sentences: "Structured data
isn't required for generative AI search, and there's no special schema.org
markup you need to add. However, it's a good idea to continue using it as
part of your overall SEO strategy, as it helps with being eligible for rich
results on Google Search."
So the honest summary is narrow. In Google Search, `FAQPage` markup now
earns no rich result and is not needed for AI Overviews or AI Mode. Whether
other systems that read your pages use it is a question for their own
documentation, and this page makes no claim about them.
## Should you keep it or remove it
Keep it if it is correct, and remove it if keeping it costs effort. The
markup does no harm when it matches the page, so there is no reason to
strip working blocks in a hurry. There is also no reason to add new ones
expecting a change in Google's results.
Two cases deserve a look. Markup generated by a plugin or template can
drift from the visible text when the page is edited, and a block that
disagrees with the page misleads every program that reads it. And markup
added only to win the old rich result, on pages that are not really
questions and answers, describes something the page is not; that is worth
removing on its own merits.
This site keeps its own `FAQPage` blocks, and a check at build time fails
if any block differs from the questions and answers a reader sees.
## Not a reason to switch to QAPage
A neighbouring type, `QAPage`, still has a Google rich result, and it is
tempting to relabel an FAQ to keep a richer listing. Google's QAPage
documentation, last updated on 2026-09-08 and checked on 2026-10-09,
rules that out. It describes Q&A pages as "web pages that contain data in a
question and answer format, which is one question followed by its answers,"
and lists among the invalid uses "An FAQ page written by the site itself
with no way for users to submit alternative answers."
The two types describe different pages. An FAQ is a set of questions the
site answers itself. A Q&A page is one question that its users answer,
as on a forum or a support community. Relabelling one as the other turns
an accurate description into an inaccurate one, which is the opposite of
what markup is for.
## The rule that did not change
Structured data must describe what is on the page. A question in the markup
should be visible on the page, and its answer should say the same thing in
both places. That rule never depended on a rich result: it is what makes
the markup a description rather than a claim.
The practical test is simple. Read the questions in the markup, then find
each one on the page. If any is missing, or if the answer differs, fix the
page or the block until they agree, and repeat the test whenever the page
is edited.
## What enrichment writes, and what it does not
AI SEO enrichment does not write `FAQPage` markup. Where a page is missing
JSON-LD, it fills the blocks that describe the organisation, the site, the
breadcrumb and the type of page, all derived from the page and the brand
profile. Questions and answers are content a person writes; a block that
invented them would describe a page that does not exist.
What the filled blocks contain, and why a broken block is worse than none,
is covered on the parent page about structured data; the whole set of tags
enrichment fills is described under [AI SEO](/ai-seo/).
## FAQ
**Is FAQ schema still worth adding?**
Not for Google's results: since May 7, 2026, Google no longer shows FAQ
rich results, and its guide says structured data is not required for
generative AI search. The markup does no harm when it matches the page, so
keep correct blocks you already have, but do not add new ones expecting a
change in Google.
**Does Google still show FAQ rich results?**
No. Google recorded on its documentation updates page that the feature
"will no longer appear in Google Search starting May 7, 2026," and in June
2026 it removed the documentation, noting the feature is no longer shown.
Since 2023 it had already been limited to well-known government and health
websites.
**Should I remove FAQPage markup from my site?**
Only if it is wrong or costly to maintain. Correct markup that matches the
visible questions does no harm, and schema.org still defines the type.
Remove blocks that disagree with the page or that describe pages which are
not really questions and answers, since those mislead any program that
reads them.
**Does FAQ schema help with AI search?**
Not in Google. Its guide to generative AI features says structured data
"isn't required for generative AI search" and that no special schema.org
markup is needed. For other AI systems, check their own documentation;
this page makes no claim about how they treat FAQPage markup.
--- https://bridge-seo.com/ai-seo/structured-data/rich-results/ ---
# Rich results: what Google shows, and who decides it
A rich result is a search result with extra elements, such as ratings,
prices or a breadcrumb trail, that Google builds from a page's
[structured data](/ai-seo/structured-data/). Markup makes a rich result
possible, never certain: Google says it "does not guarantee that your
structured data will show up in search results, even if your page is
marked up correctly."
## What a rich result is
Google's own definition, from its gallery of supported markup, last
updated on 2026-06-15 and checked on 2026-10-12: "Google uses structured
data to understand the content on the page and show that content in a
richer appearance in search results, which is called a rich result."
The names vary. "Rich snippet" is the older term and is still the one
people search for; "rich result" is the one Google's documentation uses
now. They describe the same thing: a result that shows more than a title,
an address and a description.
## Which kinds exist
Google keeps a gallery of the features it supports, grouped by the kind of
site they suit: ecommerce, organisations, news, jobs, food, education and
others. Each feature has its own structured data type and its own required
properties, documented on its own page.
A few are general. Breadcrumbs, described in the gallery as "Navigation
that indicates the page's position in the site hierarchy," apply to most
sites. Others, such as recipes or job postings, apply only to pages that
really are those things. The gallery is the place to check what exists
today, because it changes.
## Who decides
Google decides, per query and per result. Its structured data policies,
last updated on 2026-07-10, put it in one sentence: "Using structured data
enables a feature to be present, it does not guarantee that it will be
present." The gallery adds that "The actual appearance in search results
might be different."
So a correct block is an entry ticket, not an outcome. A page can carry
valid markup for months without the feature appearing, and the feature can
appear for some queries and not for others.
## Features Google has withdrawn
Google also retires features, and its documentation updates page records
when. In September 2023 it removed the How-to documentation "as this rich
result is no longer shown in search results," and limited FAQ results to
well-known government and health websites. On May 7, 2026 it stopped
showing FAQ rich results altogether, a change covered under
[FAQ schema](/ai-seo/structured-data/faq-schema/).
The practical lesson is to treat any rich result as revocable. Markup for a
withdrawn feature does no harm when it describes the page accurately, but
it no longer changes how the result looks.
## What can take a rich result away
Two things can remove a rich result that was showing: Google changing the
feature, and a problem with the markup itself. The policies are specific
about the second: "If your page contains a structured data issue, it can
result in a manual action," and such an action "means that a page loses
eligibility for appearance as a rich result; it doesn't affect how the
page ranks in Google web search."
The most common issue is markup that describes something the reader cannot
see. Google's introduction to structured data asks you not to "add
structured data about information that is not visible to the user, even if
the information is accurate." Markup that matches the visible page is the
condition every feature starts from.
## Keeping markup eligible
Most rich results are lost to drift rather than to a single mistake. A
template changes, a field is renamed, a price moves into a script, and the
markup that matched the page last month no longer does. Eligibility then
fails quietly, page by page.
**Generate markup from the same data as the page.** A block filled from
the same source as the visible text cannot disagree with it. A block pasted
by hand will, sooner or later.
**Use the type the page actually is.** A product page marked as an article,
or a listing marked as a single product, asks for properties the page does
not have.
**Check after template changes, not only after launch.** The status reports
catch problems across a whole site, which is where a template change
shows up first.
**Keep only what you can maintain.** A smaller set of correct blocks is
worth more than a large set that drifts.
## How to check markup
Google offers two tools, and they answer different questions. The Rich
Results Test checks a single page or code snippet and shows which features
it is eligible for. The rich result status reports in Search Console show
the same for every page Google has found on a verified site, including
errors that appear after a template changes.
Both report eligibility. Neither predicts whether a feature will be shown,
which is why a passing test and an ordinary-looking result are not a
contradiction.
## Where enrichment fits
AI SEO enrichment fills structured data that a page is missing: the
organisation, the site, the breadcrumb and the type of page, derived from
the page and the brand profile, and never altering markup somebody wrote.
That makes a page eligible for the features those types support, such as
breadcrumbs. Whether Google shows them remains Google's decision, as above.
## FAQ
**What are rich snippets?**
Rich snippets is the older name for rich results: search results that show
extra elements, such as ratings, prices or breadcrumbs, built from a
page's structured data. Google's documentation now uses "rich result."
Both names describe the same feature, and the same rule applies: markup
makes it possible, not certain.
**How do I get rich results?**
Add structured data of a type Google supports, with its required
properties, describing what is visible on the page. Check it with the Rich
Results Test and the status reports in Search Console. Google then decides
per query whether to show the feature; its policies say markup "does not
guarantee that it will be present."
**Do rich results improve rankings?**
Google's policies say that losing eligibility through a manual action
"doesn't affect how the page ranks in Google web search," which treats
rich results as a matter of appearance rather than ranking. A rich result
changes how a result looks, not where it is placed.
**Why did my rich result disappear?**
Either Google changed or withdrew the feature, as with How-to in 2023 and
FAQ in May 2026, or something changed on the page: markup removed by a
template, properties no longer valid, or a manual action for a structured
data issue. Search Console's status reports show which of the two applies.
--- https://bridge-seo.com/ai-seo/twitter-cards/ ---
# Twitter card tags: a second preview vocabulary for X
Twitter card tags are X's own preview vocabulary: `twitter:card`
chooses the shape, and `twitter:title`, `twitter:description` and
`twitter:image` fill it. They carry the same intention as
[Open Graph](/ai-seo/open-graph/) in a different set of tags, and
[AI SEO](/ai-seo/) fills the ones a page is missing.
A page can carry a perfectly good Open Graph set and still preview badly
on X, because nothing there is obliged to read it. The second vocabulary
is not a duplicate of the first, and where it differs is where things go
wrong.
## What the tags declare
`twitter:card` chooses the shape: a small square thumbnail beside the text,
or a large image above it. `twitter:title` and `twitter:description` carry
the words. `twitter:image` carries the picture, and `twitter:site` names
the account the page belongs to.
The shape matters more than it sounds. A summary card and a large-image
card occupy very different amounts of a reader's screen, and the choice is yours to declare.
## The two shapes, and what they cost
[Diagram: Two card shapes: a summary card puts a small square image beside the text, a large-image card puts a wide one above it]
`summary` puts a small square image beside the text. `summary_large_image`
puts a wide one above it. Everything else about the card is the same, and
the difference in how much attention it draws is not small.
The shape is declared rather than inferred. A wide picture does not
produce a wide card on its own; the tag has to say so. This catches people
out regularly: the image is right, the card is still a thumbnail, and
nothing in the page looks wrong.
The mismatch runs the other way too. A card declared large with a square
image behind it gets cropped to fit, and what survives the crop is the
middle. Where the interesting part of the picture is not in the middle,
the preview shows the part that is.
## Why it falls back, and why that is not enough
Readers usually fall back to Open Graph when a Twitter tag is missing,
which is why the gap so often goes unnoticed. The fallback is a courtesy,
not a rule, and it does not cover everything: the card shape has no Open
Graph equivalent to fall back to.
So a page with only Open Graph usually previews, and usually as a small
card, whether or not that was the intention.
## When the two sets disagree
Nothing keeps them aligned. A page can declare one title in Open Graph and
another in the Twitter set, and both are valid; the reader simply picks
the one addressed to it.
Disagreement is rarely deliberate. It arrives when somebody updates the
Open Graph title and does not think to look ten lines further down, or
when two systems write into the same head at different times. The result
is a page that describes itself differently depending on where it is
shared, and it is nobody's job to notice.
Checking is awkward for the same reason it goes wrong: the two sets are
read by different readers, so seeing one correct tells you nothing about
the other. Reading the head directly is the only reliable check, which is
also why a marker on generated nodes is worth having.
## The handle, and what it is for
`twitter:site` names the account the page belongs to, and `twitter:creator`
names the person who wrote it. Neither affects the picture or the words.
What they do is attach an attribution line to the card, and on a platform
where accounts matter more than domains, that line is what makes a shared
link recognisably yours.
Both come from your brand profile verbatim. A handle is a fact rather than
a phrasing, and generating one would be an odd thing to attempt. Where you
have set no handle, the line is simply left off rather than filled with a
guess.
## The line nobody sets
`twitter:image:alt` describes the picture for somebody who cannot see it.
It is the easiest tag in the set to argue for: a
card without it is a card that reads, to a screen reader, as a title and
then silence where the image was.
It is also the one place in the family where a generated line is doing
something a person would rarely get around to. Alt text for a preview
image is written once, seen by almost nobody, and often left out of
templates. Where the image itself is a
[composed card](/ai-seo/preview-images/), the text describing it is
written from the same page the card was built from.
## What is filled
Where the tags are absent, they are written from the page and the brand
profile, including the handle, where you have set one. Platforms will not
infer the card shape, but enrichment can: where one is not declared,
one is chosen to match the image the page actually has. A tag you wrote is
kept.
## FAQ
**Is `twitter:` still the right prefix?**
Yes. The attribute names did not change when the platform's name did, and
readers still look for exactly those strings. Renaming them in your head
is easy; renaming them in your markup breaks the card. This is one of the
places where the older name survives in the plumbing long after it stopped
being the name anyone says out loud.
**Do I need both Twitter tags and Open Graph?**
For predictable previews, yes. Relying on the fallback works until it does
not, and the failure is silent: the link still previews, just not as the
card you had in mind. The one thing the fallback cannot supply is the
shape, because Open Graph has no equivalent tag to fall back to.
**Can the two sets disagree?**
They can, and when they do the Twitter tags win on X. That is occasionally
useful and more often an accident, because the disagreement usually starts
as an edit to one set that nobody mirrored into the other. Reading the
head is the only way to see it; the preview itself shows you one answer
and hides the other.
**What shape is chosen where I declared none?**
The one that matches the picture the page actually has. A wide image gets
a large card and a square one gets a summary, because declaring a shape
the image cannot fill produces a crop rather than an improvement. Where
you declare a shape yourself, that declaration stands whatever the image.
**Does any of this affect what X shows in search?**
No. These tags decide what a shared link unfolds into, and nothing else.
They are not ranking inputs on the platform, they carry no weight with
search engines, and a page that previews beautifully is not thereby easier
to find. The gain is recognition at the moment somebody is deciding
whether to open it.
--- https://bridge-seo.com/hreflang/canonical/ ---
# Hreflang and canonical: using the two tags together
[Hreflang](/hreflang/) and canonical sit in the same head and are read together, yet
they answer different questions: one says which version suits this
reader, the other says which address counts as the page. Nothing about
them conflicts by design. Every conflict people meet has been built, and
built quietly.
## What each tag declares
The two declarations do different jobs, and neither can do the other's.
**Hreflang is an offer of choice.** It tells an engine that the page in
front of it belongs to a set, and that a reader in another language or
region has a better member waiting. It says nothing about which address
is authoritative. It assumes every member stands on its own.
**Canonical is a claim of identity.** It names the address that should
stand for this content when more than one address carries it. It says
nothing about languages: a canonical never suggests that a reader in
French would be better served somewhere else.
Because the questions differ, using both is not a compromise between
rivals. An international site of any size usually needs both, and the
pages that carry one often turn out to carry the other.
## The two lines, side by side
In the head of a French member of a set, the pair looks like this:
```html
```
Three details carry the weight. The `rel="canonical"` names this page's
own address, not a parent or a variant. The `rel="alternate"` entry for
`fr` repeats that same address, so the set and the canonical agree about
where the French version lives. And the language codes are [`ISO 639-1`](https://www.iso.org/iso-639-language-code)
values, the same vocabulary the whole set uses; the codes themselves are
covered in the hreflang overview, not here.
The fallback line is ordinary in this arrangement: `x-default` points at
whichever address you chose for
[everybody the set does not describe](/hreflang/x-default/), and that
address obeys the same rule as any member.
## The one rule that keeps them compatible
Point hreflang at canonical addresses, and at nothing else.
An hreflang entry is an invitation to fetch a page. If the fetched page
then declares a different address as canonical, the invitation was
written to the wrong door: an engine has been offered an address as an
alternate and told on arrival that the real page lives elsewhere. Each
declaration is legible on its own. Together they describe a set that
does not quite exist.
The working form of the rule is short. Every address in the set is its
own canonical, every member lists the set, and no member's canonical
looks sideways. A page whose canonical points elsewhere is not a member
at all; the address its canonical names is the candidate to list in its
place.
## How the combination breaks
The typical sequence starts with ordinary duplication. A parameterised
address acquires a canonical to its clean twin, which is the right move
on its own. The hreflang set, written months earlier, still lists the
parameterised form. Nothing errors, nothing warns, and every page still
looks correct when read alone.
What the set now offers is an address whose own head disowns it. An
engine holding both signals has to discount one of them, and which one
it discounts is its decision, made per engine and revisable without
notice. A set consistent under the rule never puts the question.
The shape should be familiar from [hreflang errors](/hreflang/errors/):
a declaration that one side makes and the other side fails to honour.
There the failing side is a missing return entry; here it is a canonical
looking away. The set degrades the same way in both cases, quietly and
as a whole.
## Checking the pair
The pass is the one described under
[hreflang errors](/hreflang/errors/), with one extra reading per page.
Fetch each address the set declares and confirm it answers directly.
Confirm the fetched page names the set in return. Then read its
canonical and confirm it names itself. A member that fails the last
reading is the place to start, because that failure is invisible to any
report that reads tags one page at a time: the alternate looks valid on
the page that declares it, and the contradiction sits on the page it
points to.
The same reading catches the reverse arrangement, a canonical that
points at a page carrying no hreflang at all. Whether that is a fault
depends on what you meant, which is why it is worth seeing on a
list instead of assuming.
## Where this page stops
Choosing a canonical, collapsing duplicate addresses, and how canonical
interacts with redirects are a subject of their own, covered under
[canonical tags](/meta-tags/canonical/).
This page needs a single fact from it: whatever address you canonise is the address your
hreflang set must name. Hold that, and the two tags have nothing left to
disagree about.
## FAQ
**Can I use hreflang and rel canonical together?**
Yes, and on a site with language versions you usually should:
they answer different questions and neither replaces the other. The one
requirement is agreement, meaning the set lists canonical addresses and
every member's canonical names itself. Kept to that, the combination is
ordinary and unremarkable.
**Does canonical override hreflang when they disagree?**
Not reliably either way. An engine holding two contradicting signals
discounts one of them by its own rules, and those rules change without
notice, so the outcome cannot be predicted. What you control is whether the signals
contradict at all, and a set that follows the rule on this page leaves
nothing to override.
**Should every page in an hreflang set be self-canonical?**
As a working rule, yes. A member whose canonical points elsewhere is
telling engines that the listed address is not the real page, which
unpicks the invitation the set just made. If some address should not be
self-canonical, that is usually the sign it should not be in the set;
list the address it canonises instead.
**Does the x-default address need its own canonical too?**
It is a member like any other: it lists the set, the set lists it, and
its canonical names itself. Nothing about the fallback role changes the
rule. A default whose canonical looks elsewhere disowns the offer
exactly the way any ordinary member would, and it breaks for the widest
audience of all.
**Can I point hreflang at a page whose canonical looks elsewhere?**
You can, in the sense that nothing will stop you, and it is the exact
mistake this page exists to name. The alternate you offered is disowned
by its own head on arrival. Fix either side, but fix one: name the
canonical address in the set, or make the listed address self-canonical.
--- https://bridge-seo.com/hreflang/errors/ ---
# Hreflang errors: the ones that break the whole set
Many [hreflang](/hreflang/) problems are not mistakes in the markup. They are declarations
that one side of a relationship never returns, addresses that point at
something other than the page they claim, and two methods quietly disagreeing.
All three look fine from the page you are reading and are visible only from
the other end.
## Why one error affects the whole set
[Diagram: One page names another as its translation, the other does not name it back, and the pair is ignored]
A set of pages is a set of claims about each other, and a claim only one
party makes is not usable by anybody.
This is the property that makes hreflang unusual to debug. A broken image
affects one page. A page missing from the set breaks the relationship for
every other member, because each of them now names something that does not
name it back. The fault is in a pair, but the damage is in the group.
The practical consequence: the page that breaks the set is almost never
the page anybody is looking at when the problem is reported.
## The declaration that is never returned
A declaration that its target never returns is a frequent fault, and the
one worth checking first.
Page A names page B as its alternate. Page B does not name A. The
relationship is asserted by one side only, and there is no way for anybody
receiving A's claim to confirm it, so the claim does not do the work it was
written to do. Every page in a set must name every other, and must name
itself.
Four maintenance patterns produce this, and none of them looks like a
mistake at the time:
**Adding a language.** The new pages list every existing version, and the
existing versions are not edited to list the new one. The new pages are
correct in isolation; the set is not.
**Removing a page.** The other members keep naming an address that no longer
exists or no longer carries the declarations. Everyone remembers to delete
the page and nobody remembers the references to it.
**Generating one side from a template.** A template that emits the set on
some page types and not others produces a set whose membership depends on
which template rendered the page, which is invisible until you compare pages
of different types.
**Editing one region's pages independently.** Where different teams own
different markets, each team maintains its own side of the relationship, and
the sides drift.
The check that finds all four is the same one described below: fetch each
declared address and read what it declares back.
## Entries that point at the wrong thing
The address in a declaration is a promise, and three kinds of address break
it.
**Addresses that redirect.** A declaration naming an address that then sends
the client elsewhere is claiming a relationship with a page that is not
there. The relationship you meant is with the destination, so name the
destination. Where the destination itself redirects, the entry is pointing
into a [chain](/redirects/chains/), and every hop in it is one more
place the set can come apart.
**Addresses that do not return the page.** Entries left pointing at removed
pages, at error responses, or at addresses excluded from being fetched by
your own configuration. Each of these is a member of the set that cannot
participate in it.
**Addresses from another environment.** Staging hostnames are the classic
case, usually arriving through a template that was correct in the
environment where it was written. These are worth checking explicitly,
because they look completely normal to anybody reading the page in the
environment where they work.
## Two methods disagreeing
The declarations can live in the page, in a sitemap, or in an `HTTP` header,
and the trouble starts when two of those are in use at once.
Nothing forbids using more than one method. What causes damage is that they
are maintained by different people at different times: the pages are edited
during a redesign, the sitemap is generated by a system nobody has looked at
in a year, and the two now describe different sets. Which one is believed is
not a question with a satisfying answer, and the honest fix is not to be in
that position.
Pick one method, remove the others, and make that choice explicitly rather
than by accident. Where a site has drifted into two, the older one is usually
the sitemap, and it is usually the one still describing the structure the
site had two redesigns ago.
## Conflicts with the canonical declaration
Strictly from the hreflang side: a set can be undermined by what its members
say about their own canonical address.
The rule that avoids the conflict is that each version's canonical
declaration names itself. When a version instead names another version as
canonical, the page is saying two incompatible things at once: that it is a
distinct version for a distinct audience, and that it is a duplicate of a
different page. The second statement disagrees with the whole point of the
first.
This is a frequent outcome of consolidating duplicates, where somebody
correctly removes duplication across a site and applies the same treatment to
translations, which are not duplicates. Depth on canonical belongs elsewhere;
what matters here is only the interaction: the set requires each member to
stand as its own address.
## How to check
Fetch every address each page declares and read what comes back; the
check is mechanical, and it is the only one that finds the faults above.
For each page in the set, take every address it declares, fetch it, and read
the declarations that come back. Three questions settle it: does the
address return the page directly rather than redirecting, does the page name
your original page in return, and does it name itself. A set that passes all
three for every pair is correct regardless of which method carries the
declarations.
Doing this by hand is feasible for a handful of pages and unreasonable for a
large set, which is what a validator that follows and re-reads each address
is for: a mechanical pass over the same three questions, one address at a
time.
Two checks are worth adding to the same pass: that every language code is a
valid [`ISO 639-1`](https://www.iso.org/iso-639-language-code) value and every region a valid [`ISO 3166-1`](https://www.iso.org/iso-3166-country-codes.html) one, and that
the fallback entry, if the set has one, appears in exactly one place.
## What to fix first
Fix non-returned declarations first and unfetchable addresses second,
because those two break sets and the rest are tidying.
Non-returned declarations come first: they are frequent and they
damage the whole group. Addresses pointing at redirects or missing pages come
second, since a member that cannot be fetched cannot participate whatever
else is correct. Method disagreements come third, and are usually resolved by
deleting rather than editing. Code and fallback problems come last: they
narrow or widen who a page is offered to, which matters, but they do not stop
the set from working.
Recheck after each change rather than at the end. Sets are edited in more
than one place, and a fix applied to one side is a change to the pair rather
than to the page.
## FAQ
**Why does one missing page break the whole set?**
Because every other member names it, and each of those claims is now
unreturned. Hreflang describes relationships between pairs, so removing a
member does not subtract one page from the group; it leaves every remaining
page holding a claim that nothing confirms.
**Do I have to list a page in its own set?**
Yes. Each page names every member including itself, which is what makes the
set consistent from whichever member an engine reads it. A set where members
name each other but not themselves is incomplete in the same way as one that
misses a member entirely.
**Can I use both page markup and a sitemap?**
Nothing forbids it, and it is still worth avoiding. Two methods are
maintained by different people on different schedules and eventually describe
different sets, at which point the question of which is believed has no
answer you can act on. Pick one and remove the other.
**Should each version point its canonical at the main language?**
No. Each version's canonical names itself, or the page is claiming to be a
duplicate of another version while also claiming to be a distinct version for
a distinct audience. That combination is a common way a correctly
built set stops working.
--- https://bridge-seo.com/hreflang/implementation/ ---
# Hreflang implementation: tags, headers or a sitemap
Hreflang is declared in one of three places: `` elements in each
page's ``, an HTTP `Link` header, or an XML sitemap. Google's
documentation says "The three methods are equivalent from Google's
perspective," so the choice is practical, not tactical. The
[hreflang overview](/hreflang/) covers what the declaration means; this
page covers how each method is written.
## Choosing a method
Pick the method that your site can keep correct, and use only one. Google's
documentation, last updated on 2026-09-21 and checked on 2026-10-09, adds a
warning to its statement of equivalence: "While you can use all three
methods at the same time, there's no benefit in Search (in fact, it maybe
be much harder to manage three implementations instead of just picking
one)."
| Method | Suits | Weak point |
|---|---|---|
| HTML `` elements | Most sites, where templates can be edited | Every page lists every version; large sets mean long heads |
| HTTP `Link` header | Files with no ``, such as PDFs | Set by server configuration, invisible in page source |
| XML sitemap | Large sites, or templates that are hard to change | Kept apart from the pages, so it can drift from them |
Whatever the method, the rules are the same: each version lists itself and
every other version, and every address is written in full.
## In the HTML head
The HTML method adds one `` element per version to the `` of
every version. Google's documentation gives the form of each element:
```html
```
and an example for a company with English pages for the UK, the US and
everyone else, a German page, and a default page with a language selector:
```html
Widgets, Inc
```
The same set appears on all five pages. Google describes it plainly: "The
set of links is identical for every version of the page." The elements must
sit inside a well-formed ``, and hreflang should not be combined
with other attributes, such as `media`, in a single ``.
## In an HTTP header
The header method sends the same information with the response, in a
`Link` header. Google calls it "useful for non-HTML files (like PDFs)",
which have no `` to hold the elements. Its example for a PDF in
English, Swiss German and German:
```http
Link: ; rel="alternate"; hreflang="en",
; rel="alternate"; hreflang="de-CH",
; rel="alternate"; hreflang="de"
```
As with the HTML method, "The Link: header returned for every version of a
page is identical." The header is set in the server's configuration or by
the application that serves the file.
## In an XML sitemap
The sitemap method moves every declaration out of the pages into the
sitemap. Each version gets its own `` entry with a ``, and each
entry lists every version, itself included, as an `` child
with `rel="alternate"` and an `hreflang` value. The sitemap declares the
namespace at the top:
```xml
xmlns:xhtml="http://www.w3.org/1999/xhtml"
```
Google spells out the consequence for a set of three: "your sitemap will
have entries for the URLs of each version, and each entry will have 3
identical child entries." It also notes that these child elements "don't
count towards the URL limit for sitemaps," and that the order of the
children does not matter, though a fixed order makes mistakes easier to
spot.
The sitemap has to sit where it is allowed to describe those addresses: "a
sitemap can only contain descendant URLs of the directory where the sitemap
is hosted from."
## Switching from one method to another
A move between methods is safest as a swap, not an overlap. Build the new
declarations completely, for every version, and check them against the
served pages or the published sitemap. Only then remove the old ones, in
the same release, so there is no period in which the two disagree.
Moving from HTML elements to a sitemap is the common direction on growing
sites, because the heads stop getting longer with every new language. The
risk in that move is losing the link between the declarations and the
pages: when a page is renamed or removed, the template used to update
itself, and the sitemap now has to be regenerated from the same source.
## Keeping the set current
Most broken hreflang is not written wrong; it goes wrong later. A new
language added to some pages and not others, a version that moved without
its siblings being told, a page deleted while the others still point to
it: each leaves the set asserting something that is no longer true.
The defence is to generate the declarations from one list of versions,
wherever that list lives, rather than editing them by hand on each page.
When the list changes, every page and every sitemap entry changes with it.
## WordPress and other CMSs
On a CMS, hreflang usually comes from a multilingual plugin or from the
theme, which writes one of the three methods for you. That saves the
typing and moves the risk: the output depends on how each translation is
linked inside the CMS, and a missing link there becomes a missing return
tag on the page.
Check what the site actually serves, not the settings screen. Open the
source of two versions of the same page and compare their sets; they
should be identical, with each page present in both.
## Checking the result
A correct implementation passes three checks, whichever method you chose.
- **Every version lists itself and all the others.** Google ignores a pair that does not point both ways; the [hreflang errors](/hreflang/errors/) page covers this and the other common faults.
- **Every address is complete.** Google requires fully qualified URLs, "including the transport method (http/https)."
- **The canonical agrees.** Each version should name itself as canonical, as covered under [hreflang and canonical](/hreflang/canonical/).
If the set has a language selector or a page for unmatched visitors, it
belongs in the set as the default entry, covered under
[x-default](/hreflang/x-default/).
## FAQ
**Which hreflang method is best?**
None, from Google's side: its documentation calls the three methods
equivalent. Use HTML link elements if you can edit your templates, the HTTP
header for files without a head such as PDFs, and a sitemap for large sites
or awkward templates. What matters is choosing one and keeping it correct.
**Can I use HTML tags and a sitemap together?**
You can, but Google says there is no benefit in Search and that managing
several implementations is harder than picking one. Two methods also give
the declarations two chances to disagree. Choose the method your site can
keep accurate and remove the other once the switch is checked.
**How do I add hreflang in WordPress?**
Usually through a multilingual plugin or the theme, which writes the tags,
the header or the sitemap entries for you. The output depends on how
translations are linked inside the CMS, so check the served pages: two
versions of one page should list the same set, each including itself.
**Does hreflang go in the sitemap or on the page?**
Either, and not necessarily both. Google treats the HTML elements and the
sitemap entries as equivalent. On the page the set sits in each ``;
in the sitemap each version gets its own entry listing every version.
Pick the one that is easier to keep complete on your site.
--- https://bridge-seo.com/hreflang/url-structure/ ---
# International SEO: choosing a URL for each country
A site aimed at several countries can give each one its own country domain,
a subdomain or a subfolder of one main domain. Google lists all three as
workable, each with trade-offs, and recommends against using URL parameters
to separate countries. Whatever the choice, the versions still need
[hreflang](/hreflang/) to say which serves whom.
## Multilingual or multi-regional
Two different situations are often called international SEO, and Google
separates them. Its guide to multi-regional and multilingual sites, last
updated on 2025-12-10 and checked on 2026-10-12, defines a multilingual
website as "any website that offers content in more than one language," and
a multi-regional website as "one that explicitly targets users in different
countries."
The difference matters for addresses. A site in English and French for one
market needs language versions; a shop selling in the United States and
Canada with different prices needs country versions, even in one language.
Many sites need both.
## The four options Google lists
Google's guide sets out four ways to structure addresses by country, with
pros and cons for each. Summarised:
| Option | Example shape | For | Against |
|---|---|---|---|
| Country domain | a domain ending in the country code, such as .de for Germany | Clear geotargeting; server location irrelevant; easy separation of sites | Expensive, with limited availability; more infrastructure; strict rules in some countries; one country only |
| Subdomain on a general domain | `de.example.com` | Easy to set up; different server locations possible; easy separation | Users might not recognise the targeting from the address alone ("de": language or country?) |
| Subfolder on a general domain | `example.com/de/` | Easy to set up; low maintenance on one host | Users might not recognise the targeting; one server location; harder to separate |
| URL parameters | `example.com/?loc=de` | none listed | "Not recommended": hard to segment, targeting not visible |
The table is Google's assessment of practical trade-offs. It does not rank
the first three; each suits a different kind of site.
## Domains that look national but are not
Some country codes are used as general domains, and Google treats them so.
Its guide says: "We also treat some vanity ccTLDs (such as .tv and .me) as
gTLDs, as we've found that users and website owners frequently see these
more generic than country-targeted." A site on such a domain is not
targeted at that country by its address alone.
## How Google works out the target country
Google combines several signals, and its guide lists them. Country domains,
"tied to a specific country," give "a strong signal to both users and
search engines." Hreflang statements, in tags, headers or sitemaps, say
which version serves which audience. Server location "can be a signal," but
"is not a definitive signal," since many sites use content delivery
networks. Other signals include "local addresses and phone numbers on the
pages, the use of local language and currency, links from other local
sites."
Some things it does not use. Google "ignores locational meta tags (like
geo.position or distribution)," and it uses the visible text, not `lang`
attributes or the URL, to work out a page's language.
## Do not redirect by location
Sending visitors to a version automatically, by their IP address or browser
language, is a common idea that Google advises against. Its guide says:
"Avoid automatically redirecting users from one language version of a site
to a different language version," because such redirects "could prevent
users (and search engines) from viewing all the versions of your site." On
IP detection it is blunter: "Don't use IP analysis to adapt your content,"
since "most, but not all, Google crawls originate from the US."
The alternative it recommends is a choice: links on every page that let a
visitor pick their region or language. A visitor who lands on the wrong
version can then move in one click, and every version stays reachable.
## Choosing between them
The choice usually follows the organisation more than search. A separate
country domain fits a business that runs each country as its own operation,
with its own team and content. A subfolder fits a single site managed in
one place, where most pages exist in every country version. A subdomain
sits between the two, with separate hosting possible but one brand domain.
Three questions settle most cases:
- **Who maintains each version?** One team on one system points to subfolders; separate teams point to domains or subdomains.
- **How different are the versions?** Mostly translated pages suit subfolders; different products and prices per country make separation more useful.
- **Can you get the domains?** Country domains can be costly or restricted, and some countries limit who may register them.
## Every option still needs hreflang
The address structure tells a reader where they are; hreflang tells search
engines which version is for which language or country. Google's guidance
on localized versions applies to all four options alike: each version lists
itself and every other version, in HTML, a header or a sitemap. How to
write those declarations is covered under
[hreflang implementation](/hreflang/implementation/).
A country domain does not replace hreflang either. Without it, a search
engine sees several similar sites and has to guess which belongs to which
audience.
## Same language, several countries
Versions for different countries in the same language, such as English for
the United States and for the United Kingdom, can look like duplicates.
Google's guide addresses this case directly: where similar content in one
language sits on several addresses, "pick a preferred version and use the
rel="canonical" element and hreflang tags to make sure that the correct
language or regional URL is served to searchers." Differences that matter
to the reader, such as currency, prices and delivery terms, also make the
versions genuinely distinct.
## Changing the structure later
Moving from one structure to another, for example from country domains to
subfolders, changes every address on the site. It is a site move: every old
address needs a permanent redirect to its new one, as covered under
[301 redirects](/redirects/301/), and every hreflang declaration has to be
updated to the new addresses at the same time. Choosing carefully at the
start avoids the largest single migration a multi-country site can face.
## FAQ
**Should I use a country domain, a subdomain or a subfolder?**
Google lists all three as workable with different trade-offs. Country
domains give clear targeting but cost more and cover one country each.
Subfolders are easy to set up and maintain on one host. Subdomains sit
between. Choose by how you run each country, then add hreflang.
**Does server location matter for international SEO?**
A little. Google says the server's location "can be a signal about your
site's intended audience," but "not a definitive signal," because many sites
use content delivery networks or host abroad. Country domains and hreflang
are the stronger, explicit signals, and Google's comparison calls server
location irrelevant for country domains.
**Can I use URL parameters for different countries?**
Google lists URL parameters, such as a loc parameter in the address, as
"Not recommended," because segmenting the site becomes difficult and users
cannot see the targeting from the address. Use a country domain, a
subdomain or a subfolder instead, with hreflang.
**Do I need hreflang if I use country domains?**
Yes. A country code in the domain says where a site is aimed, but hreflang
connects the versions to each other and tells search engines which one to
show for which language or region. Without it, the versions look like
separate, similar sites rather than one site in several versions.
--- https://bridge-seo.com/hreflang/x-default/ ---
# The x-default entry: whom your hreflang set misses
A [set of hreflang language versions](/hreflang/) describes the audiences you built pages for. The
fallback entry, written as `x-default`, names the version to offer everybody
else. It is the only part of the declaration that is about people you did not
plan for, which is why it is the part easiest to leave out.
## What it is for
[Diagram: A request in a language the set does not name falls through to the default entry]
The fallback entry names the page to show a visitor your set has no
version for, instead of leaving that choice to whatever measure is applied.
A site with versions for three languages has nothing specific to offer a
visitor searching in a fourth. Something will still be shown to them, because
one of your pages is the closest match by whatever measure is applied. The
fallback entry replaces that measure with your decision.
That is the whole of its purpose, and it explains why it belongs on nearly
every set. Exhaustive sets are rare, and a set that aims to be exhaustive
today acquires a gap the moment somebody in an unplanned market finds you.
## What it is not
The fallback is not the main language version, not a substitute for a
missing entry and not a ranking signal. Each of those readings looks
reasonable, and each produces the wrong page.
**Not the main language version.** The fallback is not a way of saying which
version came first or which market matters most. Those are business facts and
the declaration has no field for them. Marking the original version as the
fallback is sometimes correct, but only if it is genuinely the right thing to
show somebody the set does not describe.
**Not a substitute for a missing entry.** Where a version exists for an
audience, that audience gets its own entry with its own [`ISO 639-1`](https://www.iso.org/iso-639-language-code) language
code. The fallback is for the audiences that have no version, not a shortcut
around declaring the ones that do.
**Not a ranking or priority signal.** It describes coverage, not importance.
A set can name its smallest, least commercially interesting page as the
fallback and be entirely correct in doing so.
## Choosing which page it points at
Point it at a content version when your versions differ mostly by region,
and at a selector page when they differ by language with no common one.
**A selector page.** It offers the visitor the list and lets them pick.
Honest, and it never shows anybody the wrong language. It also costs a step
and delivers no content to somebody who arrived from a search wanting an
answer, so it converts curiosity into a decision at the worst possible moment.
**A content version.** Usually the one whose language reaches the widest
audience outside your set. It delivers something immediately, at the cost of
being wrong for some of the people who see it.
The deciding question is what your versions differ by. Where they
differ mostly by region while sharing a language, a content version is nearly
always right: an unplanned visitor gets readable content and only the regional
details are off. Where they differ by language with no common one among them,
a selector page is the safer answer, because the failure it avoids is showing
somebody a page they cannot read at all.
## Where it sits in the set
The fallback appears once in a set, at one `URL`, and follows the same
reciprocity as every other entry.
It is not declared once per page type, and not on the main version only:
each member of the set carries the whole set, including this one. A
fallback declared on the primary page and nowhere else is a quiet fault,
because the page that declares it looks correct in isolation.
That last point is worth connecting to the general case. The whole set
inherits the weakness of whichever page is looked after least, which is
the subject of [hreflang errors](/hreflang/errors/); the fallback entry is
simply one more
thing every member must carry.
## Three ways it goes wrong
The fallback goes wrong when it is declared more than once, when it
points at a page that cannot be fetched, or when it points at an address
that redirects by location. All three survive review, because none of
them looks wrong on the page where it is written.
**Declared more than once.** Two entries claiming to be the fallback is a
contradiction, and it is usually the result of two people each adding one to
a different template. Nothing announces the duplication; the set simply
stops making a single statement.
**Pointing at something that cannot be fetched.** A fallback naming a page
excluded by your own `robots.txt`, or one that answers with an `HTTP` error,
nominates a page nobody can receive. This is a particular risk with selector pages, because selector pages are the ones people are
tempted to exclude from being fetched on the grounds that they hold no
content.
**Pointing at an address that redirects by location.** Before you set up
automatic redirection based on where a visitor appears to be, note what it
does to the rest of the set: a page that sends everybody somewhere else based on
their apparent location cannot serve as the fallback, cannot be verified by
you from your own country, and makes the whole set untestable, since every
check you run returns the version your own location selects. If you need
location-based routing, offer it rather than impose it, and keep every
version reachable at its own address. Redirection itself, and how to keep it
from compounding, is covered under [redirects](/redirects/).
## When a set does not need one
A set can do without a fallback in two cases: a single-language site
with no regional versions, and a set that genuinely describes every
audience the site can serve.
A site with a single language and no regional versions has no set to speak
of and nothing to fall back to. And a set that genuinely describes every
audience the site can serve, which in practice means an audience defined by
something other than language, such as a service available only in named
countries, may reasonably decide that anybody outside it is not a visitor to
plan for.
Everywhere else, the absence of a fallback is a decision made by not making
one, and the result is that some part of your audience is shown a version
selected by rules you never saw.
## How to check it
Check the fallback in the same pass as the rest of the set: fetch each
member and read its declarations back. The fallback should appear
exactly once per member, name the same address every time, and that address
should return a page directly rather than redirecting. If different members
name different fallbacks, you have two sets that overlap rather than one set
with a fallback. A validator that follows and re-reads each declared address
reports this without the arithmetic.
## FAQ
**Does every hreflang set need an x-default entry?**
Nearly every one benefits from it, because sets are almost never exhaustive
and the alternative is that something is chosen for you. The exceptions are a
site with one language and no regional versions, and a set whose audience is
genuinely closed, such as a service offered only in named countries.
**Should the fallback point at my English version?**
Only if English is genuinely the right thing to show somebody your set does
not describe. That is often true and is not automatic. The fallback describes
coverage rather than importance, so the question is what an unplanned visitor
can use, not which version came first.
**Can I use a language selector page as the fallback?**
Yes, and it is the safer choice when your versions differ by language with
nothing in common. It costs the visitor a step and delivers no content
immediately, which is the trade. Where versions differ mainly by region while
sharing a language, a content version usually serves better.
**Why does my fallback seem to be ignored?**
Check that it is declared on every member of the set rather than on the main
version alone, that only one entry claims it, and that the address it names
returns a page directly. Those three cover the usual cases, and all three
look correct when you read only the page that declares it.
--- https://bridge-seo.com/meta-tags/canonical/ ---
# Canonical tags for SEO: which address search keeps
A canonical tag is a `` element in a page's
`` that names the main address among several showing the same
content. Google calls it "a strong signal that the specified URL should
become canonical": a signal, not a command. Among the
[meta tags a page carries](/meta-tags/), it is the one that speaks for
other addresses.
## Where copies of one page come from
Copies of one page come from addresses, not from content: the same page is
reachable at more than one URL, and each URL looks like a separate page
until something says otherwise. On example.com, all of these can return the
same product:
- `https://example.com/atlas-40/`
- `https://example.com/atlas-40` (no trailing slash)
- `http://example.com/atlas-40/` and `https://www.example.com/atlas-40/`
- `https://example.com/atlas-40/?utm_source=newsletter`
- `https://example.com/atlas-40/?colour=green`
- `https://example.com/print/atlas-40/`
Nobody built six pages. A server that answers on both protocols, a tracking
parameter, a filter and a print template did it between them. Search sees
six addresses and has to decide which one to show.
## What the tag says, and how strongly
The tag states a preference and leaves the choice to the engine. Google's
documentation lists the methods in order of how strongly they influence
the choice:
- **Redirects:** "A strong signal that the target of the redirect should become canonical."
- **`rel="canonical"` link annotations:** "A strong signal that the specified URL should become canonical."
- **Sitemap inclusion:** "A weak signal that helps the URLs that are included in a sitemap become canonical."
The same page adds that none of these is required: if you state no
preference, Google "will identify which version of the URL is objectively
the best version to show to users in Search." The tag matters when you know
better than the engine which address should represent the page, and you
usually do. That documentation was last updated on 2026-07-10 and checked
on 2026-10-09.
## How to write the tag
The tag goes in the `` and carries one absolute address:
```html
```
**Use the full address.** Google supports relative paths but does not
recommend them, because they "can cause problems in the long run", for
example when a test copy of the site gets crawled and every page on it
points at itself.
**Put it on the main page too.** Google recommends a canonical on the
canonical page itself, a self-referential canonical. It costs one line and
removes any doubt when parameters are added to that address from outside.
**Say the same thing everywhere.** Pick one method per page, or make all of
them agree. A sitemap listing one address while the tag names another is
two contradicting statements, and Google's list of best practices names
exactly that case.
**Use the header for files.** A PDF has no ``. For non-HTML files,
the same statement goes in an HTTP `Link` header with `rel="canonical"`.
## Mistakes Google names
Google's documentation names a short list of things not to do, and each is
a way of saying something other than what was meant.
- **A fragment as canonical.** Google generally does not support URL fragments, so an address ending in `#section` names nothing it can use.
- **Different canonicals by different methods.** The tag says one address, the sitemap another; the signals cancel rather than add up.
- **Blocking or removing instead of declaring.** Google advises against using `robots.txt`, `noindex` or the URL removal tool to pick a canonical within a site, because they hide pages rather than point between them. Those directives are a subject of their own and are not covered here.
A further, quieter mistake is pointing the tag at an address that does not
return the page: a redirect, an error or a different page entirely. The tag
then names something that cannot be the main version, and the engine has
to fall back on its own choice.
## Canonical tag or redirect
A redirect is for an address that should stop existing; a canonical tag is
for an address that should keep working. When an old URL has moved, a
visitor and a crawler should both land on the new one, and a
[redirect](/redirects/) does that while also being the strongest signal on
Google's list.
When the copy has to stay reachable, as with a filtered view, a tracking
link or a print version, a redirect would break the thing people use. The
canonical tag lets the address keep serving visitors while asking search to
credit the main one.
## Canonical and hreflang together
Language versions are not duplicates of each other, and the canonical tag
should not merge them. Google's guidance is to "specify a canonical page in
the same language, or the best possible substitute language if a canonical
page doesn't exist for the same language." Each language version names
itself, and the hreflang set names all of them. How the two tags work
side by side, and how the combination breaks, is covered under
[hreflang and canonical](/hreflang/canonical/).
## Checking what is declared
The check reads the delivered page rather than the template. Open the page
source of each address variant and confirm that every one names the same
absolute URL, that the URL it names returns the page with a success status,
and that the sitemap lists that same URL. Copies of a page often carry
copies of its title as well; finding those across a site is covered under
[duplicated titles and descriptions](/meta-tags/duplicates/).
## FAQ
**What is a canonical tag?**
It is a `` element in the `` that names the
main address among several addresses showing the same content. Google
treats it as a strong signal for which URL to show in search results. It
does not remove the other addresses; they keep working for visitors.
**Is a canonical tag a directive?**
No. Google lists it as a strong signal, alongside redirects, among the
methods that influence which URL becomes canonical; the choice stays with
the engine. Consistent signals carry more weight than one tag on its own,
so the tag, internal links, the sitemap and any redirects should all name
the same address.
**Should every page have a self-referencing canonical?**
Google recommends one on the canonical page itself, which it calls a
self-referential canonical. It costs a single line and keeps the
preference clear when someone links to the page with tracking or filter
parameters added. Use the full absolute address, not a relative path,
since relative paths can point a crawled test copy at itself.
**Canonical tag or 301 redirect?**
Use a redirect when the old address should stop serving the page, such as
after a move; it is the strongest signal on Google's list. Use a canonical
tag when the copy must keep working for people, such as a filtered or
tracked version, while search credits the main address.
--- https://bridge-seo.com/meta-tags/duplicates/ ---
# Duplicated titles and descriptions: what they cost
Two pages carrying the same [title tag](/meta-tags/) are not being punished for it. What they
are doing is failing to tell anybody which is which, including the
engine deciding which of them to show. That decision then gets made without
your input, and the cost of duplicates is entirely contained in that
sentence.
## Where they come from
[Diagram: Two addresses carry the same description, and the engine is left to choose which one to show]
Duplicates come from templates and structure rather than from
carelessness, and none of the sources below is a mistake on its own.
**A template that names the section rather than the page.** Every `HTML`
page in a section inherits the same pattern, and the variable part is either
missing or sits so late that the first words are identical across hundreds of
pages.
**Pagination.** The second, third, and fortieth page of a listing are
generated by the same code and described by the same words. It is often the largest source by count and the least noticed, because nobody
looks at page fourteen.
**Filters and facets.** Every combination of options produces a `URL`, and
every one of them inherits the description of the thing being filtered. A
modest catalogue can generate thousands of pages this way, all correctly
described and all described identically.
**Near-identical items.** Product variants, locations, or entries that differ
in one attribute and share everything else. Here the duplication in the
metadata honestly reflects duplication in the content, which is a different
problem wearing the same clothes.
## What they actually cost
Duplicates cost you the choice of which page is shown, set your pages
competing with each other, and make the description less likely to be
used; the first cost is usually the one that matters.
**You stop choosing which page is shown.** When several of your pages answer
a query equally, something has to pick, and if your metadata gives no basis
for preferring one, the pick is made on other grounds. Sometimes it selects
the page you would have chosen. Often it selects a filtered view, a
paginated page, or an old variant, and the visitor arrives somewhere
adequate rather than somewhere good.
**Your pages compete with each other.** Two pages aiming at the same phrase
divide the evidence that either is the right answer. Neither is stronger for
the other's existence, and the version of this that hurts is the one where a
strong page and a weak page both target the same thing.
**The description is less likely to be used at all.** Descriptions are
suggestions rather than instructions, as
the section on meta tags sets out, and a description that
describes a hundred pages equally is the least useful suggestion available.
Generic text is the text most likely to be replaced by something
drawn from the page.
Absent from the list: there is no penalty here, and treating
duplicates as an infraction leads people to fix them cosmetically rather than
structurally.
## Pagination deserves its own answer
A paginated page should say plainly what it is, including its position in
the sequence; a cleverer title does not help page two.
A paginated page usually should not be competing for anything. It exists so
that people can reach further items, and its metadata should say plainly what
it is rather than repeat the ambitions of the first page. Making the position
in the sequence part of the title is enough to distinguish it, and it also
stops the page from claiming to be the main answer to a query it should not
win.
If pages deep in a sequence do attract visitors, that is worth understanding
before optimising: it usually means the content people wanted is buried, and
the fix is in the structure.
## Filters are a structural decision
With facets, the metadata is downstream of a decision nobody wrote down.
The question is not how to describe ten thousand filtered views. It is which
of them should exist as separate addresses at all. A filter combination that
matches real demand deserves a page and its own description. One that exists
only because the interface can produce it does not, and no amount of writing
will make it distinct, because it is not distinct.
Where such pages should exist but should not be found separately, the
instrument is a canonical declaration or an exclusion rather than a rewritten
title. That is a decision of a different class from writing metadata.
## Finding them
Export the pairs and count, rather than looking page by page.
Take every `URL` you publish with its title and its description, then count
exact repeats. Sort the result by how much traffic the affected pages
receive, because a duplicate on two pages nobody visits is not worth
an hour and a duplicate across your busiest section is worth a day.
Then look for near-duplicates, which the exact count misses: identical
openings with a different word at the end, or strings differing only by a
number. Those behave like duplicates to a reader scanning results, since the
part that distinguishes them arrives after the point where the eye stops.
## Fixing them, in order
1. **Make the pages different, or merge them.** If two pages genuinely say the
same thing, one page saying it well is better than two saying it identically.
2. **Lead with what distinguishes.** Where pages differ in one attribute, that
attribute belongs at the front rather than after the words they share. This single change resolves much of the duplication a template produces.
3. **Decide which pages should be found at all.** Pagination and facets are
answered by structure, not by wording, and once that decision is made the
metadata question usually disappears with it.
4. **Do not manufacture variety.** Padding titles with filler to make strings
differ produces different text and identical meaning. It satisfies a report
and helps nobody, and it makes the next person believe the problem was
solved.
## FAQ
**Are duplicate titles a penalty?**
No. They are a failure to distinguish your own pages from one another, which
costs you the ability to influence which of them is shown for a query. The
harm is real and it is not a sanction, which matters because the fix is structural.
**Do I need a unique description for every page?**
For pages you want found, yes, because a generic description is the one most
likely to be replaced by text drawn from the page. For pages that should not
be competing at all, the better question is whether they need to be
separately findable in the first place.
**What about paginated pages?**
Say what they are, including their position in the sequence, and stop there.
They should not be competing with the first page, and a title that
distinguishes them is enough. If deep pages attract visitors, look at your structure first.
**How do I fix thousands of filtered pages?**
Not by writing thousands of descriptions. Decide which filter combinations
correspond to real demand and deserve their own address, and handle the rest
structurally. The metadata problem is a symptom of the addressing decision
underneath it, and rewriting the symptom leaves the decision unmade.
--- https://bridge-seo.com/meta-tags/featured-snippets/ ---
# Featured snippets: the answer box above the results
A featured snippet shows the passage from a page first and the link to it
second. Google describes featured snippets as
"special boxes where the format of a regular search result is reversed,
showing the descriptive snippet first." Google chooses the passage; no
markup decides it. Of the [search result elements](/meta-tags/), it is
built from page text.
## What a featured snippet is
A featured snippet answers the query in the result itself. Instead of a
title followed by a short description, the reader sees a paragraph, list or
table taken from a page, with the page's title and link underneath.
Google's documentation on featured snippets, last updated on 2025-12-10 and
checked on 2026-10-12, adds where they appear: they "can also appear within
a related questions group (also known as "People Also Ask")," the
expandable list of related questions in many results.
## Who chooses the passage
Google chooses it, from the text of pages that rank for the query. There is
no markup that requests a featured snippet, which sets it apart from
[rich results](/ai-seo/structured-data/rich-results/), which depend on
structured data. A featured snippet comes from what the page says, not from
what it declares about itself.
That also means the [meta description](/meta-tags/meta-description/) does
not decide it. The description is the page's suggested summary; a featured
snippet is a passage Google picks because it answers the question.
## How a page becomes a candidate
A page becomes a candidate by answering a question clearly in its own text.
The same habits that help a passage be lifted by any answering system apply
here, as covered under [AEO](/ai-seo/aeo/):
**Answer at the start of a section.** A heading that names the question,
followed by a sentence that answers it, gives Google a passage that stands
on its own.
**Use the form the answer needs.** Steps read best as a numbered list,
comparisons as a table, definitions as a short paragraph.
**Make the passage self-contained.** A paragraph that depends on earlier
text loses its meaning when shown alone in a box.
None of these guarantees selection. Google picks one passage per query from
many candidates, and the choice can change as pages and queries change.
## Paragraphs, lists and tables
Featured snippets take the shape of the answer. A question that asks what
something is tends to get a paragraph; a question about steps gets a
numbered list; a comparison may get a table. The shape is chosen by Google
from what the page offers, so the page has to offer the right one.
That is a reason to use real HTML structure. A list written as a series of
paragraphs, or a table built from images or styled blocks, gives Google
nothing to lift as a list or table. An ordered list for steps and a table
element for a comparison are the forms it can read and show.
## What changes a snippet
A featured snippet is not held for good. It can move to another page when
that page answers better, when the query is phrased differently, or when
Google changes how it shows results for a topic. Pages also lose a snippet
when their own text changes: a rewrite that moves the answer lower, or
splits it across sections, can remove the passage Google was quoting. When
editing a page that holds one, keep the answering passage intact, and check
the result again after the change is live, since the snippet may take a while
to reflect it or may move to another page.
## How long the passage should be
Google does not give a figure. Its documentation says: "Google does not
provide an exact minimum length required to appear as a featured snippet,"
because the minimum "is variable based on a number of factors, including,
but not limited to, the information in the snippet, the language, and the
platform." Write the answer as long as it needs to be to be complete and no
longer; a target word count from elsewhere has no basis in Google's
documentation.
## Opting out
A site can keep its pages out of featured snippets with the directives that
limit what Google shows from a page; Google's documentation describes using
`nosnippet` to block all snippets and a lower `max-snippet` value to reduce
the chance of a featured one. Those directives belong to how pages are
served to crawlers, which [BotBridge](https://botbridge.io/) covers; this
site does not go into them.
## Featured snippets and AI Overviews
Featured snippets and [AI Overviews](/ai-seo/ai-overviews/) are different
features. A featured snippet quotes one passage from one page and links it.
An AI Overview writes its own summary from several sources and links them.
Both start from pages that answer questions clearly, which is why the same
writing habits serve both.
## Checking whether a page has one
Search the questions the page answers and look at the top of the results,
and in the People Also Ask list. Google's Search Console reports clicks and
impressions per query, which shows where a page appears, though it does not
label featured snippets separately. Record what you find with the date,
since featured snippets change often and a single check shows only one
moment in time on one device.
## FAQ
**What is a featured snippet?**
A featured snippet is a search result shown with the passage first and the
link second. Google describes it as a box "where the format of a regular
search result is reversed, showing the descriptive snippet first." It can
also appear inside the People Also Ask list of related questions.
**How do I get a featured snippet?**
There is no setting or markup for it; Google picks the passage from pages
that rank for the query. Make a page a candidate by answering the question
in the first sentence of a section, using lists or tables where the answer
needs them, and writing passages that make sense on their own.
**Do I need schema markup for featured snippets?**
No. Featured snippets come from the page's visible text, not from
structured data, which is what sets them apart from rich results. Markup
still helps describe the page for other purposes, but it does not request
or decide a featured snippet.
**Can I stop my page appearing in featured snippets?**
Yes. Google's documentation describes using the nosnippet rule to block all
snippets, or lowering the max-snippet value to make a featured snippet less
likely while keeping regular ones. These are directives about how a page
is served to search engines, covered by BotBridge rather than this site.
--- https://bridge-seo.com/meta-tags/meta-description/ ---
# Meta description length: no limit, only truncation
Google sets no limit on how long a [meta description](/meta-tags/) can be;
it truncates the snippet as needed, typically to fit the device width. That
is the wording of Google's own documentation on snippets, last updated on
2026-04-20 and checked on 2026-10-08. What matters is how much of a
description survives the cut, and that depends on the screen.
## Where the cut falls
The cut falls at a width, measured in pixels on the screen that shows the
result, so the same description can end in different places on a phone and
on a desktop. Characters are a rough proxy for that width and no more.
Letters differ in width. A line of wide letters such as W and M fills the
space sooner than a line of narrow ones such as i and l, so two descriptions
of identical character count can be cut at different words. A counter that
reports 150 characters says nothing about which of those characters a reader
will see.
Crawlers estimate the width for you. Screaming Frog's SEO Spider, for
example, flags descriptions over 985 pixels and over 155 characters, and its
documentation describes the pixel figure as an estimate that tries to match
the latest truncation points, which Google adjusts frequently. As checked on
2026-10-08, those are the flags it raises. Treat any such figure as a
reading of today's screen, with the date it was taken.
## A working target
Aim for roughly 120 to 155 characters. On a phone the important part stays
visible, and on a desktop the description fills its lines without running
long. This is a target for writing, not a limit: nothing breaks above it,
and nothing is rewarded for landing inside it.
The range is only useful alongside a preview of the actual text. Two
descriptions of the same length can be cut at different words, so the check
that settles it is reading the snippet as it will be shown, in a preview
that renders the width, before the page is published.
## Put the point first
Truncation removes the tail, so the tail is where anything optional
belongs. The opening words carry what the page is and why somebody would
choose it; detail that only helps a reader who has already decided can sit
at the end and be lost without harm.
Compare two descriptions for the same product page on example.com:
- "Welcome to our store, where we have been helping customers since 2010. Browse the Atlas 40 backpack, a 40-litre bag with a laptop sleeve."
- "Atlas 40: a 40-litre travel backpack with a padded laptop sleeve and side access. Fits most cabin-bag limits. Free returns."
The second survives any reasonable cut. The first spends its opening on
words that would fit on every page of the shop.
## Checking what is actually shown
The description you wrote and the snippet a reader sees are two different
texts, so the check has to look at the second. Search for the page in
Google, by a query it should answer or by its address, and read the result
as it appears; repeat on a phone, since the narrower screen cuts sooner.
For more than a handful of pages, a crawler does the measuring. Screaming
Frog's SEO Spider shows each description with its estimated width and a
snippet preview; standalone previews such as Portent's SERP Preview Tool
and the Mangools SERP Simulator render a single description the same way.
Their widths are estimates of today's screen, which is why a live result
remains the final word.
## When the description is not shown at all
A description of perfect length is still a suggestion. The engine can
choose text from the page instead, usually when the written description
does not match the query or says nothing specific, and length does not
change that. The parent section on meta tags covers why that substitution
happens and what makes it less likely.
## How to write one
**Describe this page, not the site.** The description competes with nine
other results that are broadly about the same subject. Something true and
specific about this page beats a sentence that could sit on any of them.
**Lead with what distinguishes it.** Where several pages differ in one
attribute, that attribute belongs at the front. Pages that share the same
description are a separate problem, covered under
[duplicated titles and descriptions](/meta-tags/duplicates/).
**Write plainly.** A description is read in a second by somebody scanning a
list. Short sentences and concrete nouns survive that reading; slogans do
not.
**Say what the click leads to.** Where the page lets a reader do
something, such as compare plans, download a file or book a slot, name that
action in the description. A reader choosing between results wants to know
what happens next, and a specific verb answers that faster than an
adjective.
**Keep it true to the page.** A description that promises what the page
does not deliver earns a click and a quick return, and it is also the kind
the engine is most inclined to replace.
## Where nobody wrote one
Large sites rarely have a description on every page; the ones nobody got
to are filled with whatever the engine assembles. Filling those gaps from
what each page says, without touching the descriptions you wrote, is what
[meta titles and descriptions written where missing](/ai-seo/meta-title-description/)
covers.
## FAQ
**How long should a meta description be?**
There is no required length: Google sets no limit and cuts the snippet to
fit the screen. A working target of roughly 120 to 155 characters keeps the
important part visible on a phone and fills the lines on a desktop. Check
the actual text in a preview, because width depends on the letters, not
only on their count.
**Does Google cut by characters?**
No. It truncates the snippet as needed, typically to fit the device width,
so the cut is a matter of pixels on a given screen. Character counts are a
convenient approximation, and two descriptions of the same count can be
cut at different words because wide letters take more room than narrow
ones.
**Is a long description penalised?**
Nothing in Google's documentation describes a penalty for length. A long
description is simply cut where the screen runs out, so the cost is that
readers may not see its end. That is why the point belongs at the front
and anything optional at the back.
**Should I count characters or pixels?**
Write against a character target, because it is easy to keep in your head,
and check against pixels, because that is where the cut falls. Crawlers and
snippet previews estimate the width for you; treat their figures as a
reading of today's screen, since the truncation point moves.
--- https://bridge-seo.com/meta-tags/meta-keywords/ ---
# Meta keywords: why Google ignores the tag since 2009
The meta keywords tag is a line in a page's ``,
``, listing a page's terms.
Google does not use it: its documentation says the tag "is not used by
Google Search, and it has no effect on indexing and ranking at all." It is
the one [meta tag](/meta-tags/) you can drop at no cost.
## What the tag is
The meta keywords tag sits in the page's head, invisible to readers, and
holds a comma-separated list of words. Google lists it in its own reference
of meta tags, in the form:
```html
```
It was meant to tell search engines what a page was about, in the page
author's own words. The idea was reasonable; the problem was that nobody
reading the page could see the list, so nothing kept it honest.
## Why Google stopped using it
Google stopped because the tag was abused. Its Search Central blog
explained the decision on 2009-09-21, in a post by Matt Cutts: "Because the
keywords meta tag was so often abused, many years ago Google began
disregarding the keywords meta tag." The same post put it plainly: "Google
doesn't use the keywords meta tag in our web search ranking."
The reasoning is in the same post. In the early days of search, it says,
keyword meta tags "quickly became an area where someone could stuff
often-irrelevant keywords without typical visitors ever seeing those
keywords." A signal anyone can fill with anything, unseen, stops being a
signal.
## What Google says today
The position has not changed. Google's current reference for meta tags,
last updated on 2025-12-10 and checked on 2026-10-12, lists the tag under
"Unsupported tags and attributes" and says: "The meta-keyword tag is not
used by Google Search, and it has no effect on indexing and ranking at all."
The 2009 blog post now carries a note of its own: some of its details may
be outdated, "However Google still doesn't support the keywords meta tag."
## Should you remove it
Removing the tag changes nothing in Google, and keeping it costs nothing
there either. The decision rests on other grounds.
**Maintenance.** A tag nobody uses still has to be kept accurate if it
stays, and a stale list is one more thing in the template to explain.
**Disclosure.** The tag is public. Anyone can read it in the page source,
which is why the 2009 post used the example of a competitor copying one
company's keywords; for Google, the answer was that it did not matter.
**Other systems.** Google's statements cover Google Search only. Some
internal site search tools may read the tag; if yours does, keep it for
that reason. This page makes no claim about how other public search engines
treat it.
For most sites, the sensible course is to stop adding the tag to new pages
and remove it when the template is next changed, rather than editing every
single page by hand, one at a time.
## Checking your own pages
Finding out whether your site still outputs the tag takes a minute. Open
the source of a few pages, one of each template, and search for
`name="keywords"`. If it is there, note where it comes from: a theme, a
plugin setting or a field in the content system. Removing it at that
source removes it from every page that template produces.
If the lists are long, stale or copied across many pages, that is a sign
they were filled once and never maintained, which is itself a reason to
drop them rather than tidy them.
## Why the advice lingers
Guides written before 2009, and templates built from them, still mention
the tag, and some editing screens still offer a keywords field. That keeps
the habit alive long after the reason for it disappeared. When a checklist
or tool tells you to fill in meta keywords, check whether it is talking
about the tag or about a private note for editors; only the second has any
use, and only inside your own workflow.
## What does the job instead
The work the tag was meant to do is done by things readers can see. The
[title tag](/meta-tags/title-tag/) and the
[meta description](/meta-tags/meta-description/) describe the page in
words that appear in results. Headings and the text say what the page is
about. And [structured data](/ai-seo/structured-data/) declares the type
of page and its main facts in a form programs read, under the rule that it
must match what is visible.
Each of these is checked against the page itself, which is exactly what the
keywords tag never was. If the words matter, put them where readers will see
them: in the title, the opening sentence and the headings. There they help
the reader decide, and Google reads them as part of the page rather than as
a list it has long since stopped trusting.
## A note on keyword lists in plugins
Some content systems and SEO plugins still ask for a "focus keyword" or a
keyword list when editing a page. That field is usually a writing aid for
the editor, used to check the text, and does not necessarily output a meta
keywords tag. Check the page source to see what is actually published.
## FAQ
**Do meta keywords still matter?**
Not for Google. Its documentation states that the meta keywords tag "is not
used by Google Search, and it has no effect on indexing and ranking at all,"
and Google has ignored it since at least 2009 because it was so widely
abused. Titles, descriptions, headings and the text itself do the work the
tag was meant to do.
**Should I remove meta keywords from my site?**
Removing them changes nothing in Google, and keeping them costs nothing
there. Remove them when convenient, usually at the next template change, to
avoid maintaining a tag nobody reads. Keep them only if an internal search
tool on your own site actually uses them.
**Does Bing use the meta keywords tag?**
This page does not answer that, because the sources it relies on are
Google's own statements about Google Search. Check each search engine's
current documentation before relying on the tag for any of them; Google's
position is clear and dates back to 2009.
**What replaced the meta keywords tag?**
Nothing replaced it directly. Its job is done by visible signals: the title
tag, the meta description, headings and the page text, plus structured data
that describes the page and must match what readers see. All of these can
be checked against the page, which the keywords tag never could.
--- https://bridge-seo.com/meta-tags/title-tag/ ---
# Title tags for SEO: length, truncation and rewrites
A title tag is the `` element in a page's ``: the name on
the browser tab and the first source Google lists for a result's title.
Google sets no limit on its length and truncates the title link to fit the
device width. It is one of the two [meta tags that shape a search
result](/meta-tags/).
## Where the title tag lives
The title tag lives once in the ``, as plain text between
`` and ``, and it describes the whole page. A page that
sells a backpack on example.com might carry this:
```html
Atlas 40 Travel Backpack, 40 Litres | Example Outfitters
```
The H1 is a different element with a different reader. The H1 sits in
the visible page and greets somebody who has already arrived; the title
appears in search results, on the browser tab, in bookmarks and in the
history list, where the reader has not arrived yet. The two often carry
the same words, and that is fine, but they are written for two different
moments.
## How long a title tag can be
There is no maximum. Google's documentation on title links says so in one
sentence: "While there's no limit on how long a `` element can be,
the title link is truncated in Google Search results as needed, typically
to fit the device width." That page was last updated on 2025-12-10 and
checked on 2026-10-09.
The cut is a matter of width, not of character count. A title made of wide
letters reaches the edge sooner than one made of narrow letters, and a
phone reaches it sooner than a desktop. Crawlers estimate where the edge
is. Screaming Frog's SEO Spider flags titles over 561 pixels and over 60
characters, and its documentation describes the pixel figure as an
approximation of the latest truncation points, which Google adjusts
frequently. Those were its flags when checked on 2026-10-09.
**A working target.** Aim for roughly 50 to 60 characters. That keeps the
words that matter inside the visible line on most screens without
squeezing the title into a label. It is a target for writing, not a rule:
nothing breaks above it, and nothing is rewarded for landing inside it.
## What truncation removes
Truncation removes the end of the title and leaves the beginning, so the
beginning is where the page has to be named. If the title opens with the
site name or a slogan, a narrow screen may show nothing else.
The mechanism is the same one that shortens descriptions, measured in
rendered width on the screen that shows the result; the full account of
why character counts are only a proxy sits under
[meta description length](/meta-tags/meta-description/) and applies here
unchanged. The practical rule for titles is shorter: name the thing first,
qualify it second, brand it last.
## Previewing a title before it ships
A preview is only useful when it renders width rather than counting
letters. Crawlers such as the SEO Spider show each title with its estimated
pixel width, and standalone snippet previews render a single title the same
way. Their figures are estimates of today's screen.
The check that settles it is the live result. Search for the page by a
query it should answer, read the title as it appears, and repeat on a
phone. If the visible part still says what the page is, the title works,
whatever its count.
## When Google shows a different title
Google builds the title link automatically and may not use your title tag
at all. Its documentation says that the generation of title links "is
completely automated and takes into account both the content of a page
and references to it that appear on the web."
The same page lists the sources it draws on: the content of ``
elements, the main visual title on the page, headings such as `
`,
`og:title`, other large and prominent text, other text on the page,
anchor text on the page, text in links that point to the page, and
`WebSite` structured data. The title tag is first on that list, not the
only entry.
The documentation also names the common reasons a title gets replaced:
- **Half-empty titles**, such as a site name with nothing before it.
- **Obsolete titles**, where the page shows this year and the title still says last year.
- **Inaccurate titles** that no longer describe what the page holds.
- **Micro-boilerplate**, where many pages share one title and differ only in a detail the title leaves out.
- **No clear main title**, where several headings carry the same weight.
- **A language or script mismatch** between the title and the page.
Each of these is a fault in the title rather than a whim of the engine,
and each is fixed in the page. The parent page covers the same behaviour
for titles and descriptions together; this list is the title-only part,
taken from Google's own wording.
## How to write a title tag
**Name the page, not the site.** The title competes with nine other
results about roughly the same subject. A concrete noun from this page
beats a phrase that would fit any of them.
**Put the distinguishing word first.** Where pages differ in one
attribute, such as a size, a model or a season, that attribute belongs at
the front. Titles that differ only at the end look identical once
truncated, which is how [duplicated titles](/meta-tags/duplicates/) are
born on large sites.
**Add the site name once, at the end.** Google's documentation suggests a
site name at the beginning or end, separated by a hyphen, colon or pipe,
and notes it may drop a site name that repeats what the result already
shows. At the end it costs nothing when truncated.
**Match the visible heading.** Google reads the main visual title and the
`
` as alternative sources. When they agree with the title tag, there is
nothing better for it to choose.
**Keep it current.** A title with a year, a price or a status ages faster
than the page. If it changes, change it in both places.
Compare two titles for the backpack page:
- "Example Outfitters | Quality Gear Since 2010 | Atlas 40 Backpack"
- "Atlas 40 Travel Backpack, 40 Litres | Example Outfitters"
The second keeps the product in view on any screen. The first spends its
opening on words every page of the shop already shares.
## Where nobody wrote one
A page without a title tag still gets a title link; Google assembles one
from the sources above, and the result is whatever the page happens to
offer. Filling that gap from what each page says, without touching titles
somebody wrote, is what
[meta titles and descriptions written where missing](/ai-seo/meta-title-description/)
covers.
## FAQ
**How long should a title tag be?**
Google sets no limit and cuts the title link to fit the device width. A
working target of roughly 50 to 60 characters keeps the important words
visible on most screens. Width depends on the letters, so check the
actual text in a preview or a live result rather than trusting a count
alone.
**Why does Google change my title?**
Google generates title links automatically from several sources, of which
the title tag is the first. It is more likely to replace a title that is
half empty, out of date, inaccurate, repeated across many pages, or in a
different language from the page. Fixing those faults in the page is the
lever you have.
**Is the title tag the same as the H1?**
No. The title tag sits in the `` and is shown in search results, on
browser tabs and in bookmarks; the H1 is the main visible heading of the
page. They often share words, and agreeing with each other helps, because
Google treats the H1 as an alternative source for the title link.
**Should the site name be in every title?**
Once, at the end, is enough. Google's documentation suggests placing it at
the beginning or end with a separator, and it may omit a site name that
repeats what the result already shows. At the end it is the first thing
truncation removes, which is the right order of loss.
--- https://bridge-seo.com/on-page/alt-text/ ---
# Alt text: how to write it for people and for search
Alt text is the description in an image's `alt` attribute that stands in
for the image when it cannot be seen, read by screen readers and search
engines alike. Google calls it "the most important attribute when it comes
to providing more metadata for an image." It is one of the
[on-page signals](/on-page/) that serves readers first.
## What alt text is
Alt text is a short text alternative written into the `alt` attribute of an
`` element. Google's image documentation gives a plain example:
```html
```
The attribute is for the content of the image in the context of the page,
not for its file name or its caption. Google's SEO Starter Guide, last
updated on 2025-12-10, describes it as "a short, but descriptive piece of
text that explains the relationship between the image and your content."
## Who reads it
Three kinds of reader depend on alt text, and they need the same thing.
**People who cannot see the image.** Screen readers read the alt text in
place of the image. Google notes that alt text "improves accessibility for
people who can't see images on web pages, including users who use screen
readers or have low-bandwidth connections."
**Browsers when the image fails.** If the image does not load, many
browsers show the alt text instead, so the page still makes sense.
**Search engines.** Google's image documentation, last updated on
2026-03-02 and checked on 2026-10-12, says: "Google uses alt text along with
computer vision algorithms and the contents of the page to understand the
subject matter of the image."
## How to write it
Describe what the image shows that matters to this page. Google advises
"creating useful, information-rich content that uses keywords
appropriately and is in context of the content of the page."
**Say what a reader would miss.** For a photo of a product, the alt text
names the product and the detail the page is about. For a chart, it states
what the chart shows, not that it is a chart.
**Keep it short.** A sentence or less is usually enough. Long explanations
belong in the page text, where everyone can read them.
**Skip "image of".** Screen readers already announce that an element is
an image; starting with "image of" repeats it.
**Do not stuff it.** Google warns: "Avoid filling alt attributes with
keywords (also known as keyword stuffing) as it results in a negative user
experience and may cause your site to be seen as spam."
Two versions for the same photo on a product page:
- "backpack travel backpack best backpack cheap backpack"
- "Atlas 40 backpack with the side zip open, showing the laptop sleeve"
The second helps every reader. The first helps none and is the kind Google
warns about.
## Images that need an empty alt
Not every image carries information. A decorative divider, a background
flourish or an icon next to text that already says the same thing adds
nothing for someone who cannot see it. The W3C's Web Accessibility
Initiative, in its images tutorial updated on 2026-04-08, says to "provide
a null text alternative (`alt=""`) when the only purpose of an image is to
add visual decoration to the page."
An empty `alt=""` is different from a missing attribute. The empty value
tells screen readers to skip the image; a missing attribute leaves them to
guess, often by reading out the file name. For groups of images that
together convey one piece of information, the same tutorial advises that
"the text alternative for one image should convey the information for the
entire group."
## When an image is a link
When an image is the only content of a link, its alt text becomes the
link's text. Google notes that "alt text in images is useful as anchor text
if you decide to use an image as a link." In that case, describe where the
link goes, not what the picture looks like: a logo linking to the home page
needs alt text that says so.
## Alt text in a CMS
Most alt text is written in a content system rather than in HTML. Google's
SEO Starter Guide notes that "your CMS may have an easy way to specify a
description for an image when you're uploading it to your site." That field
is the alt attribute under another name, and it is easy to skip.
Two habits keep it filled. Write the alt text when the image is added to a
page, not when it is uploaded to a library, because the right description
depends on the page that uses the image: the same photo of a backpack says
different things on a product page and in a packing guide. And check the
template for images it inserts on its own, such as author photos, logos and
icons, since those often ship with empty or missing alt attributes on every
page.
## Alt text and social cards
Images shared on social platforms through Open Graph tags have their own
description field, separate from any `alt` on the page. What a preview
card contains, and how it is built, is covered under
[preview images](/ai-seo/preview-images/). AI SEO enrichment fills missing
title, description, Open Graph, card and JSON-LD tags; it does not write
`alt` attributes, which describe images in the page's own content.
## FAQ
**What is alt text?**
Alt text is the description written in an image's alt attribute that
stands in for the image when it cannot be seen. Screen readers read it
aloud, browsers show it when the image fails to load, and Google uses it,
along with computer vision and the page's text, to understand what the
image shows.
**How long should alt text be?**
There is no official length, and this page does not invent one. Write the
shortest text that says what the image shows that matters to the page,
usually a phrase or a sentence. Longer explanations belong in the page
text, where every reader can see them.
**What is the difference between alt text and the title attribute?**
Alt text replaces the image for people and programs that cannot see it,
and it is required for meaningful images. The title attribute adds extra
information, usually shown as a tooltip on hover, and many readers never
see it. When an image is a link, Google uses the alt text as anchor text.
**Do decorative images need alt text?**
They need an empty one. The W3C's accessibility guidance is to use alt=""
when an image is purely decorative, which tells screen readers to skip it.
Leaving the attribute out entirely is worse, because a screen reader may
read out the file name instead.
--- https://bridge-seo.com/on-page/anchor-text/ ---
# Anchor text: link words that say where a link goes
Anchor text is the visible, clickable text of a link: in
`ghost peppers`, it is "ghost peppers". Google
describes good anchor text as "descriptive, reasonably concise, and
relevant to the page that it's on and to the page it links to." It is one
of the [on-page signals](/on-page/) that connect pages.
## What anchor text is
Anchor text is the part of a link a reader sees and clicks. Google's link
documentation, last updated on 2025-12-10 and checked on 2026-10-12, calls
it "the visible text of a link," and adds that "this text tells people and
Google something about the page you're linking to."
The anchor text sits between the opening and closing tags of an ``
element. Google's own example of a good link:
```html
ghost peppers
```
The same documentation shows the bad case: a link with no text at all,
which tells neither readers nor Google anything about where it leads.
## Why it matters
Anchor text matters because it is how both readers and search engines learn
what is on the other side of a link before following it. Google says links
are "a signal when determining the relevancy of pages and to find new pages
to crawl," and that "the better your anchor text, the easier it is for
people to navigate your site and for Google to understand what the page
you're linking to is about."
For a reader, the effect is direct. Someone scanning a page often reads
only the links, and a screen reader can list all the links on a page out of
context. A list of "click here, click here, read more" tells them nothing.
## What makes it good
Google's guidance sets three tests: descriptive, reasonably concise and
relevant to both pages.
**Descriptive.** The words name what the reader will find. Google's
"better" example: "For a full list of cheese available for purchase, see the
list of cheese types," with "list of cheese types" as the link.
**Reasonably concise.** A few words, not a sentence. Google's example of a
"weirdly long" anchor runs a whole clause into the link; the fix keeps only
the words that name the target.
**Relevant to both pages.** The link makes sense in the sentence it sits in
and describes the page it points to.
Google offers a simple check: "Try reading only the anchor text (out of
context) and check if it's specific enough to make sense by itself. If you
don't know what the page could be about, you need more descriptive anchor
text."
## What to avoid
**Generic phrases.** Google lists "Click here," "Read more," "website" and
"article" as bad anchor text: they could point anywhere.
**Keyword stuffing.** Google asks you to "resist the urge to cram every
keyword that's related to the page that you're linking to," noting that
keyword stuffing "is a violation of our spam policies." A link written for
a search engine rather than a reader usually reads badly to both.
**Links side by side.** Google advises: "Don't chain up links next to each
other; it's harder for your readers to distinguish between links, and you
lose surrounding text for each link."
**Empty links.** A link with no text, or an icon with no label, leaves the
anchor blank. As a fallback, Google can use a `title` attribute, but a real
text label is the reliable choice.
## When the link is an image
When an image is the whole link, its alt text becomes the anchor text.
Google's documentation says: "For images as links, Google uses the alt
attribute of the img element as anchor text, so be sure to add descriptive
alt text to your images." For such images, the alt text should describe
where the link goes; the [alt text](/on-page/alt-text/) page covers how to
write it.
## Links you do not write
Anchor text on other sites that link to yours is chosen by those sites, not
by you. It is worth reading, because it shows how others describe your
pages, but it is not something to engineer: links bought or arranged for
their anchor text fall under the same spam policies as keyword stuffing.
The part you control is the anchor text of links within your own site and
the links you place to others.
## Buttons, menus and icons
Not every link sits in a sentence. Navigation menus, buttons and icons are
links too, and their text is their anchor text. A menu item such as
"Guides" is fine where the menu itself gives context; a button reading only
"Go" or an arrow icon with no label is not.
For icon links, give the link a text label that screen readers and
crawlers can read, either visible or hidden for sighted readers. The test
is the same as for any anchor: read the label alone and ask whether it
says where the link goes. A row of icons labelled "link", "link", "link"
fails it as plainly as "click here" does.
## The words around the link
Google points out that context counts too: "the words before and after
links matter, so pay attention to the sentence as a whole." A short,
descriptive anchor inside a sentence that explains why the link is there
gives a reader everything they need to decide whether to follow it.
## FAQ
**What is anchor text?**
Anchor text is the visible, clickable text of a link, placed between the
opening and closing tags of an a element. Google says it tells people and
Google something about the page being linked to, and describes good anchor
text as descriptive, reasonably concise and relevant to both pages.
**Does anchor text matter for SEO?**
Google says that the better your anchor text, the easier it is for people
to navigate your site and for Google to understand what the linked page is
about. It also uses links as a signal of relevance. Write anchor text for
readers first; descriptive links serve both.
**What is exact match anchor text?**
It is anchor text that repeats the exact search phrase a page targets, such
as linking the words "best hiking boots" to a page about them. Used
naturally it is just a descriptive link; forced into every link it becomes
the keyword stuffing that Google's spam policies prohibit.
**Is "click here" bad anchor text?**
Yes. Google lists "Click here" and "Read more" as examples of anchor text
that is too generic, because they say nothing about the destination. Reading
the link out of context should tell you what the page is about; "click
here" fails that test.
--- https://bridge-seo.com/on-page/breadcrumbs/ ---
# Breadcrumbs: showing where a page sits on your site
Breadcrumbs are the short trail of links, usually near the top of a page,
that shows the path from the home page down to the current page, such as
Home › Guides › Backpacks. Google says a breadcrumb trail "indicates the
page's position in the site hierarchy." They are the
[on-page signal](/on-page/) that makes a site's structure visible.
## What breadcrumbs are
Breadcrumbs are a navigation aid: each step is a link to a level above the
current page, and the last step is the page itself. A reader who lands deep
in a site from a search result can see at once where they are and move up a
level without going back to the home page.
Google's documentation for breadcrumb markup, last updated on 2026-09-08 and
checked on 2026-10-12, describes the benefit in reader terms: the trail
"may help users understand and explore a site effectively. A user can
navigate all the way up in the site hierarchy, one level at a time, by
starting from the last breadcrumb in the breadcrumb trail."
## Visible trail and markup
A breadcrumb has two parts that should match: the trail a reader sees and
the structured data that describes it to search engines.
**The visible trail** is ordinary HTML links, usually in a `