Does the llms.txt file help to be cited by AI?
Published on
Nothing shows that the llms.txt file helps to be cited by AI. It is a 2024 proposal, for which no major AI publisher documents its use to select sources, and Google wrote in July 2026 that it is useless to create text files for AI. We measured it: among the sites retrieved by searches from ChatGPT, Claude, and Gemini across around ten business sectors, 17% publish an llms.txt, and the cited sites do not publish it significantly more often than the others. For ChatGPT, it is even slightly the opposite.
For what type of question? It all depends on what is expected from the file. To be cited in an AI's response to a question, nothing shows that it helps. It was proposed to help AI tools read a site, for example technical documentation, which we did not measure.
What the llms.txt file is
The llms.txt file was proposed by Jeremy Howard on September 3, 2024. The idea is to add a Markdown-formatted file to the root of a site, readable by language models, which presents the site. The format requires a title with the site name, then generally a short summary and lists of links to important pages, with an optional section for what a tool can ignore when it is short on space.
This file is often presented in GEO advice as a way to guide ChatGPT, Claude, or Gemini to the right pages. In the documentation from OpenAI, Anthropic, and Google that we read, none say they use it to choose their sources. It does not replace the robots.txt file, which tells bots what they are allowed to read.
What Google and studies say
Google responds in its guide on optimization for its generative AI features, updated on July 10, 2026: it is useless to create text files for AI, since its search does not use them. As early as April 2025, John Mueller from Google observed that to his knowledge, no AI service had said they used this file, and that one can see in his server logs that these services do not even look for it.
SE Ranking published an analysis in November 2025 of nearly 300,000 domains, of which approximately 10% published an llms.txt. Conclusion: having this file did not make a domain more likely to be cited by AI models. However, the publisher deems it low-risk to adopt in anticipation of future use.
What we measured
We took the sites retrieved by searches from ChatGPT, Claude, and Gemini across around ten business sectors queried in French on September 10, 2026, and verified on September 16 whether each publishes a valid llms.txt with its title in the expected format. Before the measurement, we had written down our prediction: less than 10% of sites would publish one, and the gap between cited and non-cited sites would not exceed 3 points for any AI.
The first prediction fell: 17% of sites publish an llms.txt, more than expected. The gaps remain small and do not go in the same direction: for Claude, 16% of cited sites versus 12% of non-cited sites; for ChatGPT, 14% of cited sites versus 16% of non-cited sites. A statistical permutation test shows that these gaps are compatible with chance. For Gemini, whose API does not distinguish discarded pages, 23% of cited sites publish one, with no point of comparison.
Why the file cannot get a page cited
For ChatGPT and Claude, a page is first found by a search that the AI writes itself, then chosen from the results. What matters at these two stages, in our measurements, is the page's rank in the results, the words of the question in its title and its address, and paragraphs that answer using the vocabulary of the search. A separate file, which the search does not read, does not intervene at any of these stages.
This site itself publishes an llms.txt file, which lists its pages. We do not claim that it gets us cited: it is just one more summary, not a measured lever.
What this changes for a business
If your site does not have an llms.txt, that is not the reason for your absence in AI. Start with what can be measured: AI search bots' access to your site, pages that answer your clients' questions using their words, and a presence in the sources that AI consults.
Publishing an llms.txt costs almost nothing and does no harm. But a provider presenting it as a citation lever is selling a promise that neither Google, nor studies, nor our measurements confirm.
What we do not know
Our measurement covers the websites retrieved across about ten sectors in French, verified six days after collection: a website may have added or removed its file in the meantime. It compares cited and non-cited websites, which describes an association, not an effect. We did not measure the file's usage by agents reading documentation.
Sources
- Jeremy Howard — "The /llms.txt file", proposal, .
- Google Search Central — "Optimizing your website for generative AI features on Google Search", .
- Search Engine Journal, Roger Montti — Google compares llms.txt to the meta keywords tag, .
- SE Ranking, Yulia Deda — llms.txt and citations by AIs, .
- FaireDuBruit measurements, from September 4 to 16, 2026 — see "How we measured".