llms.txt : à qui ce fichier parle vraiment
A markdown file placed at a site's root, `/llms.txt`, summarising what the site contains and pointing at the pages that matter, written to be read by a model rather than a browser. It reads like a table of contents: a few lines of context, then annotated links. A variant, `llms-full.txt`, appends the whole content. **What it is not: an SEO lever.** That is the confusion to undo first, because it circulates everywhere as a grade of 'A, badly underrated'. The 2026 figures say otherwise: - Google updated its AI optimisation guide on **15 June 2026** to state that the file has **no effect**, positive or negative, on rankings or on generated summaries. The position is not new: back in July 2025, John Mueller publicly compared it to the `keywords` tag, abandoned twenty years ago. - **10.13%** adoption across a 300,000-domain survey by SE Ranking. After eighteen months of discussion, one site in ten. - **408** requests targeting the file across more than **500 million** AI bot visits observed over ninety days. A search bot does not go looking for it. - Of the **fifty domains most cited** by AI, exactly **one** has it. If it were the lever announced, the correlation would show up there first. **What it really is: a table of contents for an agent you point at it yourself.** Its value sits downstream, not upstream. When a coding agent or an assistant has to work on a library, handing it a `llms.txt` address saves it guessing a documentation structure from navigation pages. It is a service to whoever reads your documentation on demand, not a signal to whoever indexes the web. Hence the rule: a library or service vendor has a good reason to publish one, a local shop site has none. **The deciding test**: the file is useful if someone can hand its address to an agent. If nobody exists to make that gesture, it is useless, and the hour spent writing it was owed elsewhere.