The StartingUp Summer keeps taking the long view on a year in which visibility changed ground. After laying out the definition of GEO and detailing how to appear in LLM answers, we examine llms.txt, the file every checklist recommends: with the evidence, this time.
Drop a simple text file at the root of your site, and suddenly make it readable by ChatGPT, Claude or Perplexity: the promise of llms.txt fits in one sentence, and it has gone around the web-marketing world. It's readily presented as "the robots.txt of the LLM era," one more must-have to tick off before the end of the quarter. In French-language coverage, the subject is almost always treated with the same uniform enthusiasm.
llms.txt is a Markdown file placed at a site's root to give AI a map of its useful content. Is it useful? To this day, no major generative engine commits to reading it, and according to an Ahrefs study from June 2026, 97% of these files receive no requests at all: publishing one is a near-zero-cost bet, not a visibility strategy.
llms.txt files: published, but read?
Voir les données
| Share of observed llms.txt files | |
|---|---|
| Never requested by a crawler | 97 % |
| Requested at least once | 3 % |
Source : Ahrefs study, server logs, June 2026
This enthusiasm deserves a closer look. Because llms.txt is at once a serious proposal, championed by a respected figure in machine learning, and a file that nothing proves, as we write this, the major AI crawlers actually read. Both facts are true at the same time, and that's exactly what makes the subject interesting.
Where llms.txt comes from
On September 3, 2024, Jeremy Howard, co-founder of Answer.AI and creator of fast.ai, publishes a one-page proposal. His starting observation: models' context windows are too narrow to swallow an entire site, and modern HTML is a pain for a machine to read. Navigation, scripts, banners and chrome drown the useful content in considerable noise. His idea: give LLMs a map of the site, written in Markdown and plain language, at an agreed-upon address, /llms.txt.
The conceptual shift is elegant. robots.txt tells robots what they must not visit; llms.txt tells them what deserves to be, and in what order. One forbids, the other invites. These are two opposite logics, and that detail will matter further down.
llms.txt and llms-full.txt: a deliberately minimal format
The format comes down to four elements: a top-level title (the only mandatory part), a short summary as a blockquote, then sections listing annotated links. An "Optional" section flags what a model can skip when context is tight. Concretely:
The proposal also includes a maximalist variant, llms-full.txt: no longer the map, but the entire library, meaning all of the site's useful content concatenated into a single Markdown document, ready to be loaded into a model's context window. Useful for technical documentation, quickly excessive for everything else.
Who has adopted llms.txt
Adoption is real, but highly skewed: it comes first from technical documentation. Anthropic publishes the file for Claude's documentation, Stripe and Zapier for their APIs, Cursor for its editor; and documentation platforms like Mintlify generate it automatically for all their customers, which explains a good share of the files in circulation. On the measurement side, an SE Ranking study covering 300,000 domains put adoption at around 10%, and 8.7% of the world's 1,000 largest sites were publishing the file as of June 2026. For a proposal with zero commitment from any engine, that's a lot.
Which leaves the question this momentum would rather we forgot: publishing, fine. But who reads it?
The point the checklists skip: (almost) nobody reads llms.txt
Google first, through John Mueller, in June 2025:
For what it's worth: no AI system uses llms.txt as of today.
John Mueller (Google), on Bluesky, June 2025, our translation
Mueller specifies that this is "obvious in the server logs" and compares the file to the meta keywords tag, which engines stopped reading twenty years ago. Gary Illyes, also from Google, confirms in July 2025 that Google doesn't support llms.txt and has no plan to. And to our knowledge, no major provider (OpenAI, Anthropic, Google, Meta) has publicly committed to consuming it. Anthropic publishes its own, but promises nowhere to read yours: the symbol sums up the situation well.
Server logs confirm this at scale. An Ahrefs study published in June 2026, covering 137,000 domains, reaches an unambiguous number: 97% of llms.txt files received no requests at all over the month studied. And among the remaining requests, the top readers are... SEO audit tools (21.7%), far ahead of any individual AI bot. A notable share of the "AI" visits also comes from development agents like Claude Code, not from the crawlers that feed public-facing answers.
| robots.txt | llms.txt | |
|---|---|---|
| Born | 1994, web community consensus | 2024, an Answer.AI proposal |
| Status | Official standard (RFC 9309, 2022) | Community proposal, no standardization |
| Logic | Exclude: what robots must not visit | Invite: what models should read |
| Respected by major crawlers | Documented, verifiable in the logs | Not demonstrated to date |
| Consequence if ignored | Uncontrolled crawling, real risks | None observed |
The comparison used as a sales pitch is therefore exactly the one that doesn't hold up. robots.txt is respected because it's standardized, old, implemented everywhere, and anyone can verify that respect in their own logs. llms.txt has no standard, no track record, no mechanism of incentive: it's a convention waiting for readers.
Should you create an llms.txt file? A near-zero-cost bet, not a strategy
Our answer comes in three parts.
- Yes, if you have documentation or a structured offering. The only demonstrated use today is agents and assistants that come fetch the file on demand when they're pointed at your site: exactly the use case Howard imagined. Documentation sites, tools and services lead the way.
- Yes, but generated, never hand-written. A manual file will drift from the site within three months and then serve stale content. Wire its generation into your CMS or your documentation generator, and don't put anything in it you wouldn't stand behind publicly.
- No, if you're expecting visibility from it. No gain has been demonstrated in generative engines' answers to date. The file replaces neither SEO nor the groundwork of GEO: at best, it can only make content that already deserves to be read easier to read.
One warning to close: llms.txt forbids nothing. To control AI bots' access to your content, the tool remains robots.txt and your server configuration. Confusing the two is like believing a menu posted on the counter closes the restaurant's door.
Machine readability is a strategy, not a file
If llms.txt disappeared tomorrow, the question it raises would remain intact: is your content readable by the machines that, more and more, decide whether or not to cite you? Content accessible without executing JavaScript, a clear structure, dated and sourced claims, data a machine can extract without guessing: that's the work that pays off, file or no file.
So go ahead and drop the file: ten minutes, near-zero cost, a reasonable bet. But drop it for what it is: the most visible and least decisive gesture in a machine-readability strategy. The day major crawlers announce they consume it, you'll be ready. In the meantime, everything that matters plays out elsewhere: in the quality and structure of what you publish.
