What makes a page citable by an AI engine

The short answer: A page becomes citable when an engine can fetch it, finds the answer to the asked question near the top of it, and can lift that answer as a passage that stays true once it leaves the page. Schema, word count and page speed sit well behind those three. The largest public citation analysis we have found puts the typical cited paragraph at a depth of 0.36, which is the top third.

The word citable does a lot of quiet work. It sounds like a property of a page. It is closer to a property of a paragraph, judged by a system that will never show your layout to anyone.

The decision is made on a passage, not on a page

An assistant does not read your page and form an opinion of your company. It retrieves candidate documents, splits them, keeps a handful of passages that look like they answer the question, and writes from what it kept. Your page is raw material for that step. The citation is attached afterwards, to the source of a passage that survived.

Two things follow. The first is that a brilliant page whose answer sits in paragraph nine loses to a thinner page that answers in paragraph two, because the thinner page keeps appearing in the retained set. Answer-first writing, and why the opening paragraph carries the answer is the practical version of that.

The second is that your paragraph will be read with nothing around it. No H1, no navigation, no preceding sentence that established what "the platform" refers to. A paragraph whose meaning rests on whatever came before turns into a false statement the moment it is lifted, and a false lift is how a model ends up inventing things about you while technically quoting your own site.

What the citation data says actually predicts a citation

An observational study published in 2026 collected about 2.1 million citations from four production engines over six months, then feature-engineered the 10,042 pages those citations pointed at. It is the largest public analysis of the question we have found, and some of its results are uncomfortable for the standard checklist.

The strongest page-level predictor was not structure. It was alignment: the token overlap between the page and the actual prompts people were asking, with a standardised coefficient of +0.37. Nothing else came close. Page length helped a little at +0.13. Title similarity to the prompt helped at +0.09.

Then the null results. Real-user Core Web Vitals showed no independent effect once the domain was controlled for. Synthetic performance scores, the same. Structured data showed no significant independent effect, and FAQ blocks landed at +0.07, which the authors did not treat as reliable.

One format did stand out. Pricing pages carried a coefficient of +0.39, better than any other page type measured, and listicle-style reviews came in negative. That matches what we see in audits: buyers ask engines what something costs, and the page that answers plainly gets quoted. It is also the reason pricing pages deserve their own attention rather than a link to "contact sales".

Keep the ceiling in view. A model using only the domain explained about 32 percent of the variation in citation counts. Adding every page-level feature took it to 45 percent. Page work moves roughly a seventh of the outcome. The rest is whose site it is, which is slower and mostly not a content problem.

Three ways a good page fails to be citable

It never arrives. The passage game cannot start if the fetch does not finish. Bot rules written for scrapers, origin latency, a JavaScript render that an engine's fetcher does not wait for, or an edge rule that answers one region differently from another. That last one is worth a look on its own, because your CDN decides who gets served your page and a block that only fires for certain networks is invisible from your own desk.

It answers a question nobody asked. A page built around an internal category name has almost no token overlap with how buyers phrase things. This is the +0.37 signal in its negative form, and it is the most common thing we rewrite.

It answers, but not extractably. The answer exists, spread across three paragraphs, each carrying half of a condition. Nothing can be lifted without becoming either incomplete or untrue, so the engine keeps someone else's cleaner paragraph instead.

A check you can run this afternoon

Take one page and one real question a buyer would type. Then do four things.

Fetch the page the way a bot would, without a browser, and read what comes back. Use curl with a plain user agent and look at the raw HTML. If the answer is not in that response, no amount of writing fixes the page.

Find the paragraph that answers the question. Count characters from the start of the body to that paragraph, divide by the total, and compare with 0.36. A depth of 0.7 tells you the answer is parked behind the preamble.

Paste that paragraph alone into a blank document. Read it as a stranger. If a pronoun, an abbreviation or a "this" has no referent, fix it in place rather than in the sentence above it.

Check the claim against the rest of your site. One paragraph says 14 days, another says two weeks and the pricing page says a fortnight on a plan you retired. An engine that keeps two of those will hedge, and a hedge reads as absence.

The self-check extends this into ten passes, including the access ones that are easy to assume away.

Where schema still earns its place

The null result on structured data is narrow and people over-read it. It says markup did not independently raise citation counts once domain effects were removed. It does not say markup is pointless.

Markup still resolves identity. It still drives rich results in classic search, which feeds the index these systems draw on. It is cheap, it is machine-checkable, and a correct Organization or Product node is the fastest way to make the company named on the page and the company named in a knowledge graph the same company. Ship it. Just stop treating a FAQPage node as the reason you are not being cited.

Google's own guidance now says this plainly. Its generative AI guide, last updated 10 July 2026, states that you do not need new machine-readable files, AI text files, markup or Markdown to appear, that chunking content into small pieces is not a requirement, and that writing in a special way for generative features is not necessary. The companion page on AI features adds the part people miss: a page has to be indexable and eligible for snippets. nosnippet removes the preview text entirely, which takes the page out of those surfaces. max-snippet is a length cap rather than a switch, so check the directive and the number you actually set before assuming either is innocent.

Honest limits

Observational data shows association, not cause. The alignment coefficient could partly reflect that teams who know their buyers' language also do ten other things well. We have not run a controlled rewrite at a scale that would settle it.

Engine behaviour also drifts. The same study found the median age of a cited page differed by about three months between two of the four engines, which means a freshness rule tuned for one surface is wrong for another. And citation counts are not revenue. Nobody has an attribution path from a named mention to a signed contract that survives scrutiny, ours included. A sample audit report shows how we attach evidence to a finding so it can at least be argued with.

FAQ

Does adding FAQ schema make a page citable? On its own, no. In the citation study it was among the weakest effects measured and not a reliable one. Add it because it is correct markup, not as a visibility tactic.

How long should a citable page be? Length had a small positive effect, well below alignment. A short page that answers the question in the first screen beats a long page that answers it on the third. Write to the question and stop.

Should I create an llms.txt file? Google's guidance says no new machine-readable file is needed for its generative features. Other engines have not committed to it either. It costs little, and it is not a substitute for a page that can be fetched and read.

Why does a competitor with a worse page get cited instead? Usually the domain. About a third of the variation in citation counts was explained by domain alone, before any page feature entered the model. Page work is the part you control, and it moves a real but bounded share.

Does blocking AI crawlers protect the content? It also removes the page from the answers those systems write. That is a business decision about a paywalled archive, and rarely the right default for a product site.

Sources

• What Drives Citations in Production Large Language Models? An Observational Multi-Method Study of Two Million AI Citations Across Ten Thousand Web Pages, 2026: arxiv.org/html/2609.35077v1
• Google Search Central, optimizing your website for generative AI features, last updated 10 July 2026: developers.google.com/search/docs/fundamentals/ai-optimization-guide
• Google Search Central, AI features and your website, last updated 10 December 2025: developers.google.com/search/docs/appearance/ai-features
• Schema.org type definitions: schema.org/docs/full.html