Which passage of a page does an AI cite?
Published on
A developed paragraph that uses the words the AI searched with. With Claude, whose API precisely indicates the cited passage, a paragraph of 50 to 100 words is cited 3.7 to 4.8 times more often than a paragraph of 10 to 25 words, and a paragraph that uses at least one-third of the words from the model's search is cited 5 to 8 times more. The passage can be located anywhere on the page, except at the very bottom. The page is chosen based on the user's question; the passage, based on the AI's words.
For what type of question? Everything depends on the question for the choice of the page. For the choice of the passage within the selected page, the rules below held true with Claude across our corpora in French as well as in English.
Two vocabularies: that of the page and that of the passage
An AI searching the web makes two successive choices. It first chooses the pages, and this choice is made based on the user's question: the words from this question in the title and the page address increase its chances of being selected. It then chooses, within the page, the passage it cites, and this second choice is made based on its own words, those it wrote in its searches.
With Claude, a paragraph that uses at least one-third of the words from the model's search is cited 5.1 times more often across the ten or so sectors in French and 8.2 times in English. The words from the user's question also play a role, but less so: 2.4 to 3.4 times. These are two vocabularies in two different places on the page; confusing them, by putting the AI's words in the title, leads to losing on both fronts.
A developed and affirmative paragraph
Length is the strongest variable. With Claude, a paragraph of 50 to 100 words is cited 4.8 times more often than a paragraph of 10 to 25 words across the ten or so sectors in French, 3.7 times in English, and 5 to 6 times in our initial corpora. Short paragraphs, hooks, and isolated sentences are rarely the selected passage.
Formulation also counts. A paragraph containing a rule verb, such as "is", "allows", or "must", is cited 1.7 to 2.7 times more often: the AI selects passages that state facts. Prose usually wins over lists, up to 3.2 times in English, but not always: on software troubleshooting, long list items describing a step were selected more often than paragraphs.
Where the cited passage is located
Contrary to popular belief, the first paragraph is nothing special. In our initial corpora, the position of the cited passage on the page is almost uniform, with only one dip: the very end of the page, where legal notices, links, and secondary content are often found. In our subsequent corpora, position did not show any stable effect.
A page therefore does not need to concentrate everything in its introduction to be cited. Any developed paragraph can become the selected passage, provided it is self-sufficient: the AI cites an extract, not the entire page, and a paragraph that refers to "what precedes" loses its meaning once isolated.
What Gemini and publishers' documentation show
Anthropic's documentation specifies that each web citation from Claude contains up to 150 characters of the cited text, and that citations are always enabled for web search. This is what allows us to locate the selected passage without approximation. Google, for its part, documents a passage ranking system that identifies sections of a page to judge its relevance.
With Gemini, the API does not return the position of the cited passage: it must be reconstructed, and each reconstruction method introduces a bias. Our reconstructions show the same direction as with Claude, with a paragraph of 25 to 100 words carrying the search words being selected more often, but the magnitude of the effect cannot be measured with certainty. The only measurement without this bias is that of Claude.
What Google says, and how to read it
Google indicates, in its guide on optimization for its generative AI features updated on July 10, 2026, that there is no obligation to break content down into small pieces: its systems know how to understand the nuances of multiple topics within a page and show the user the relevant passage. It adds that it is not necessary to write in any particular way for AI search, because these systems understand synonyms and the general meaning of what is being searched.
These claims and our measurements do not cover the same systems: Google is talking about its search, we are measuring what Claude cites through its API. They agree on the essentials. It is not a matter of writing for a machine or chopping up a page, but of writing complete paragraphs that state things clearly and use the actual vocabulary of the subject; this is also what readers expect.
What this changes for a business
To be cited, it is not enough to be the right page: you must contain the right passage. Concretely, this means writing paragraphs of several sentences, which clearly state what is true, and which use the words that AIs use to search on your topic. These words cannot be guessed: you get them by asking the questions to the AIs and noting their searches.
This is the method followed for this section: each question was asked to ChatGPT, Claude, and Gemini, their searches were recorded to write the paragraphs, and the titles carry the question, not the search words.
What we do not know
The exact location of the passage is only available with Claude; with Gemini it is reconstructed, and with ChatGPT we have not measured it. Our measurements focus on the models' API and describe associations between paragraph characteristics and citations.
Sources
- Anthropic — documentation for Claude's web search tool, .
- Google Search Central — ranking systems guide, .
- Google Search Central — "Optimizing your website for generative AI features on Google Search", .
- FaireDuBruit measurements, from September 4 to 16, 2026 — see "How we measured".