Do AI Assistants Read Your Whole Page? What Citation Data Shows About Which Part Gets Used

Many content teams follow a simple rule: the longer and more complete the page, the more chances it has to be cited by ChatGPT or Google's AI answers. The rule treats the page as one unit that is either cited or not.

Three recent datasets look inside the cited page instead, and they point to a different picture. AI search systems use a limited slice of each page they cite, usually a few hundred words. The sentences they use come disproportionately from the top of the page. That changes where your most important claims should sit.

Why it matters which part of a page gets used

When an AI assistant searches the web before answering, it retrieves pages and pulls passages from them to base its answer on. This process is called grounding: the answer is "grounded" in text taken from real sources. The cited link at the bottom of the answer shows which page was used. It does not show which sentences were used.

For a brand, that difference matters. A page can be cited while the answer repeats a generic definition from your introduction and skips the line that says who your product is for, what it costs or why it is different. Citation tracking counts the URL as a win either way.

So the first question is how big that slice is.

How much of a page does an AI answer actually use?

The clearest measurement comes from DEJAN, a search consultancy run by Dan Petrovic. In a December 2025 analysis, DEJAN ran 7,060 queries through Google's Gemini grounding with Google Search, the developer tool that lets Gemini base answers on search results. It then recorded exactly which text was taken from each source. This is not AI Overviews itself, but it shows how Google's own grounding tool selects text.

The median source contributed about 377 words, meaning half contributed more and half less. Across all sources for a single query, the median total was about 1,900 words.

Page length barely changed this. The amount selected from a page levelled off at around 540 words, however long the page was. As a result, the share of each page that got used fell sharply as pages grew:

Petrovic's own reading is that sources compete for "share of a fixed pie". His dataset and full method are not public, but the pattern in the table is clear: on a long page, most of the text is not part of the answer.

What this means: Adding length to a page does not add proportionally to what the AI uses. Past a few hundred words, extra text mostly shrinks the share of the page that gets used.

Which part of the page does that slice come from?

DEJAN measured how much text was used, not where it came from. Two other datasets answer that question, on different platforms and with different methods. Both point toward the top.

The first is from SEO consultant Kevin Indig, published in his Growth Memo newsletter in February 2026. The full study is paywalled, so the figures here come from Conversion's coverage of it. As reported, Indig analysed about 18,000 citations for position, drawn from about 1.2 million ChatGPT responses. ChatGPT does not say which sentence it used, so the cited passage was matched to the page by comparing meaning.

About 44% of those citations came from the first 30% of the page's text. If citations were spread evenly, that top section would get about 30%. So the opening drew roughly one and a half times its fair share. The middle 40% of the page drew 31%, and the final 30% drew about 25%.

The second dataset is from Daniel Shashko at Bright Data, a web data company, published in March 2026. It relies on a useful detail. Citation links in Google AI Mode and the consumer Gemini app often include a "text fragment", a piece of the URL that points to the exact sentence being cited. Shashko decoded those sentences and located them on each page.

Across 1,719 matched sentences, the median cited sentence sat 31% of the way down the page. Three-quarters sat within the top half.

Two platforms, measured two different ways, point to the same place: the upper part of the page.

That does not mean the bottom is ignored. In Indig's data, about a quarter of ChatGPT citations still came from the final 30% of the page. The top is over-represented, not the only part that counts.

Does that mean shorter pages win?

Not simply. Studies that look at whole pages rather than passages disagree on length.

SE Ranking, an SEO software company, studied about 216,000 pages in late 2025, as reported by Search Engine Journal. Pages over 2,900 words averaged 5.1 ChatGPT citations, against 3.2 for pages under 800 words. But the strongest predictor was how many other websites linked to the domain, so longer pages may simply sit on stronger sites.

A 2026 report by AirOps, a content platform, with Indig points the other way. It found that focused pages of 500 to 2,000 words were cited most consistently. Pages over 5,000 words were cited less often than pages under 500. Indig summarised the lesson as building "the page that is the best answer to one question."

The passage-level data helps reconcile this. A long page may be retrieved for more questions, but each answer still uses only a slice of it.

Either way, both findings depend on an earlier step: being retrieved at all. Before answering, ChatGPT runs a web search and gets back a ranked list of pages. In the AirOps data, a page's rank in that list was the strongest predictor of citation. Pages at the top of ChatGPT's search results were cited 58% of the time. Pages around tenth place were cited 14% of the time. As we covered in whether Google ranking still gets you cited in AI answers, getting into the results comes first. Where a sentence sits on the page only matters once the page is in the running.

What do the cited sentences look like?

If an answer takes a small slice of a page, the slice is made of sentences that make sense on their own.

In Shashko's AI Mode and Gemini app data, cited sentences averaged about 10 words, and the longest was 17. He notes this ceiling might be a display limit, so it does not prove AI avoids long sentences. It does show that the lifted text is short.

Indig's ChatGPT data, as reported, points the same way. About 36% of cited passages used definitive language, such as "X is Y", against about 20% of passages that were not cited. They also named more specific entities, such as products, companies and places.

What this means: A sentence that depends on the paragraph around it is hard to reuse. A plain claim that names the brand and states one fact can be lifted as it is.

What should teams change, and what should they not expect?

Taken together, the evidence suggests a few changes for your most important pages:

• Put brand-defining facts in the first screen. State the category, who the product is for, the price model and the main differentiator near the top.
• Write those facts as standalone sentences. One claim per sentence, with the brand or product named, so the line still makes sense when quoted alone.
• Keep each page focused on one question. Padding a page with loosely related material lowers the share of it that any answer uses.
• Check which sentence AI answers repeat about you. Ask the assistants your buyers use and compare the answer to your page. Our guide to auditing inaccurate AI answers about your brand covers how to run those checks.

Be clear about the limits, too. All of these studies are observational and come from practitioners or vendors, not peer-reviewed research. No one has tested whether moving a sentence higher causes it to be cited more often. Shashko states that his data could not separate position from relevance. Introductions may get cited simply because they hold definitions and summaries.

Google's own guidance on AI features also says no special optimisation is needed to appear in AI Overviews or AI Mode. So treat "key facts first" as good writing practice backed by consistent patterns, not as a ranking rule.

The core point holds either way: a citation tells you the page was used, not that your message was. If the sentence you most want repeated is buried at the bottom of a long page, it has to compete with everything above it for a few hundred words of space.