What words does an AI use when it searches?
Published on
Not your customer’s words. An AI searching the web reformulates the question into several searches, using its own words: source names, document types, years, and often English. On our questions asked in French on September 15, 2026, 95% of ChatGPT’s responses included at least one search in English, 75% for Gemini, and 32% for Claude. ChatGPT launches about ten searches per response, Gemini three, Claude two. These words cannot be guessed: you have to ask the AIs and record their searches.
For what type of question? It all depends on the question. When it names a brand or a source, the AI includes this name in its searches and looks for it. When it does not name anything, the AI chooses the words itself, and therefore the sources.
One question becomes several searches
Publishers document this. In May 2025, Google presented the 'query fan-out' technique of AI Mode, which splits the question into sub-themes and simultaneously launches a multitude of searches; its documentation for sites indicates that AI Overviews and AI Mode can use it. Anthropic specifies that Claude decides on its own to search, that the search can be repeated several times within the same response, and that a simple factual question generally requires one to three searches, while a comparative search requires ten or more.
We read these searches directly in the models' responses. On our questions from September 15, 2026, ChatGPT launched a median of ten searches per response, with spreads ranging from four to twenty-four. Gemini launched a median of three, Claude two. Each search brings back its own results: a page can be found by a search that the user never typed.
Searches in English, even for a question in French
The language of the question does not determine the language of the searches. On our questions asked in French, about 56% of ChatGPT’s searches were written in English, 42% for Gemini, and 14% for Claude; 95% of ChatGPT’s responses contained at least one search in English. The subject of these questions, visibility in AI, is heavily documented in English, which pushes this figure upward.
The phenomenon is not limited to this subject. Across about ten business sectors queried in French, nearly half of the pages brought back by the three AIs were in English. Conversely, on the same questions asked in English, Claude did not bring back any French pages. A company that only publishes in French therefore leaves aside a portion of the searches launched on questions that were nevertheless asked in its language.
Source names, document types, and years
ChatGPT often writes the name of the source it wants to consult in its search: a publisher, an administration, a media outlet. In our corpora, a large majority of its search queries thus name a publisher, from 72 to 91% depending on the corpus, compared to 3 to 40% for Claude. About half of ChatGPT’s searches also specify a document type, such as a guide, official documentation, or regulation. On our September 15 questions, 15% of its searches even targeted a specific site using the "site:" operator.
Years also appear: 18% of ChatGPT’s searches and 17% of Claude’s contained a year, compared to 1% for Gemini. These words tell us what the model is actually looking for: a source it knows, a document of a certain genre, dated information. A page that carries none of these signals is less likely to appear in the results of these searches.
Where these words matter on a page
The words in the AI’s searches matter for the choice of the cited passage, not for the choice of the page. In Claude, a paragraph that repeats at least one-third of the words from the model's search is cited 5 to 8 times more often than a paragraph containing none. In the title, however, the words that the AI adds to the question do not help: they reduced citations in Claude on our first corpora and had no effect on the subsequent ones.
Ahrefs, on ChatGPT 5.2 in April 2026, observes that the titles of the cited pages are close to these internal sub-questions. The two measurements do not focus on the same thing: a proximity of meaning for Ahrefs, the presence of words for us. The resulting practical rule is simple: the user's question in the title, the vocabulary of the AI's searches in the paragraphs.
Google sets a clear limit to this practice. In its generative AI guide updated on July 10, 2026, it writes that it may be tempting to create separate content for each variation of how people search, including searches reformulated by AI, but doing so primarily to manipulate rankings or AI responses violates its policy against large-scale content abuse. It adds that it is not necessary to write in a particular way for AI, as its systems understand synonyms. Knowing AI searches serves to better answer a real question, not to multiply pages.
When AI understands something other than the question
Recording searches also reveals misunderstandings. On the question 'Is what we read about GEO true?', the three AIs understood something other than optimization for AI engines: Gemini and ChatGPT searched whether GEO magazine is reliable, Claude searched for an organization. On 'Does a page title matter for being cited by an AI?', ChatGPT searched whether a title is protected by copyright, and cited Légifrance.
A word with multiple meanings can therefore send the AI off on another topic, and with it the sources it consults. These errors are not visible in the final response, which seems coherent; they are visible in the searches. This is one more reason to measure before writing.
What this changes for a business
Writing for AIs based on classic SEO keywords is not enough: the AI does not search using your client's words, and not necessarily in their language. You need to know, question by question, the searches that ChatGPT, Claude, and Gemini actually launch, then take them into account in pages that truly answer, rather than multiplying pages by variation, and provide an English version when searches pivot to English.
This is the method followed for this section: each question was asked to the three AIs, and their searches were recorded before writing. The pages that FaireDuBruit distributes for its clients are also published in English, for searches that are not made in French.
What we do not know
Our language tracking relies on simple automatic detection, which leaves some searches undecided; the percentages are therefore orders of magnitude. The observed searches are those from the models' API, which may differ from those of their consumer interfaces.
Sources
- Google, Elizabeth Reid — AI Mode update and 'query fan-out', .
- Google Search Central — 'AI features and your website', .
- Anthropic — documentation for Claude's web search tool, .
- Ahrefs, Louise Linehan — 'Why ChatGPT cites pages', .
- 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".