llms.txt: what is this file for, and who reads it?

Among the tips for “showing up in ChatGPT”, one file comes up often: llms.txt, presented as the robots.txt of AI. It exists, it takes ten minutes to write, and it can be useful. But it remains a proposal, and no major AI engine has announced reading it to choose the sites it cites.

Here is what it contains, where it comes from, and what is really known about its readers.

Where llms.txt comes from

The file was proposed on September 3, 2024 by Jeremy Howard, on the llmstxt.org site. A second version came out in August 2026, revised after two years of use.

The starting point is simple. A language model reading a web page wastes a good part of what it can take in on menus, scripts and banners. llms.txt offers it a shortcut: a short text file, written in Markdown, that presents the site in a few lines and points to the useful pages.

The author designed it for on-demand reading: an agent looks for information for someone, reads the file, then follows the links. He sees it less as a gateway for training models. The format was born around software documentation, but it aims to suit any site.

The format, in five parts

The file goes at the root of the site: your-domain.com/llms.txt. It is written in this order:

1. A first-level heading with the site’s name. It is the only required part.
2. A quote that sums up the site in one or two sentences.
3. Details, if needed, in paragraphs or lists, without subheadings.
4. Second-level sections, each with a list of links: the page name, its address, and a short note after a colon.
5. A section titled “Optional”, for secondary links an agent in a hurry can skip.

For a bakery, it would look like this:

Version 2 adds two ideas. Each page can offer a Markdown version at the same address, ending in “.md”. And a page can point to the llms.txt that covers it, through a tag in its code or an HTTP header.

What is known about its readers

• Developer tools use it. According to llmstxt.org, thousands of sites publish one, documentation platforms generate it automatically, and coding agents use it reliably. Anthropic, OpenAI and Perplexity publish one for their own developer documentation.
• AI companies don’t announce reading it on other sites. The pages where OpenAI, Perplexity and Anthropic describe their bots explain how to allow or block them through robots.txt. They say nowhere that these bots consult the llms.txt of the sites they visit; OpenAI and Perplexity only point there to the file of their own documentation.
• Google says it doesn’t need it. For its AI Overviews and AI Mode, Google states that there is no need to create new machine-readable files, text files for AI, or special markup.
• Chrome added it to Lighthouse, its audit tool, presenting it as an emerging convention, optional for now. The audit only flags a server error when the file is requested. A site without llms.txt is rated “not applicable”: it loses nothing.

What isn’t known: whether consumer assistants, ChatGPT, Gemini, Perplexity or Claude, open this file when they answer a question about a shop or a practice. None has announced it, and no public measurement makes it possible to say.

Should you publish one?

For the one-page site of a shop, a craftsperson or a practice, the stakes are low either way.

• It replaces nothing. Neither indexing by Google, nor a robots.txt that lets search bots through, nor clear facts written on the page. Those are what get a site cited: see how to get cited by AI.
• It costs little. A ten-line file, no tool needed.
• It mostly helps documentation-heavy sites: software, technical sheets, online courses. An agent gains from quickly finding the right page among hundreds.
• It must tell the truth. If it exists, it is kept up to date like the rest. A file announcing old opening hours spreads one more error.

If you write one

1. Create a text file named llms.txt, saved in UTF-8.
2. First line: a hash sign, a space, then the name of your business and your town.
3. Next line: a greater-than sign, a space, then your business summed up in one sentence.
4. Add a section, for example “Practical info”, preceded by two hash signs, with one link per line: a dash, the page name in square brackets, its full address in parentheses, then a colon and a short description.
5. Put it at the root of the site, then open your-domain.com/llms.txt to check that it displays.

Put nothing in it that the page doesn’t already say: an agent comparing the two must find the same facts.

The 29webs website publishes its own, at 29webs.com/llms.txt. It lists its pages of advice, trades, styles and features, each with its description.

In short: llms.txt is a serious proposal, useful to agents that read documentation. No consumer assistant has announced using it to cite sites, and Google says it doesn’t need it. Publish one if you like, but only after the essentials: an indexed page that clearly states the facts.