llms.txt: what is this file for, and should you publish one?

This article contains affiliate links: we may earn a commission. Learn more

The llms.txt file is a Markdown site map for AI agents: it was proposed in 2024 so that an AI can quickly find the right page in a set of documentation, not to improve a ranking. Google writes in its July 2026 guide that it ignores the file, and two studies find no effect on AI citations.

We have not measured the effect of llms.txt on any engine. This article rests on the specification, Google’s documentation (read on October 21, 2026), two studies read at the source, and a survey of 26 sites that we ran with curl the same day. How we test explains why some articles stay informational. For a definition of GEO, see GEO: definition and differences from SEO.

What is the llms.txt file?

It is a Markdown text file placed at the root of a site (/llms.txt) or in a subfolder, which presents the site in a few lines and points to the useful pages. The proposal, written by Jeremy Howard in 2024 and moved to version 2 in 2026, asks that these pages also have a clean Markdown version, so that an agent does not have to clean up HTML.

The format is simple. Only the level 1 title is required. Next come, in this order, a summary as a quote block, free paragraphs, then level 2 sections that hold lists of links. A section named “Optional” groups the links an agent can skip when it is short on space.

# My site

> One sentence saying what the site is and who it is for.

## Guides
- [Get started](https://example.com/start.md): setup in 5 minutes
- [Pricing](https://example.com/pricing.md): plan grid

## Optional
- [Version history](https://example.com/versions.md)

What is llms.txt really for?

It guides an AI agent that looks for information while it helps a user, for example a coding assistant reading a library’s documentation. According to the proposal, it is mostly used for software documentation, and its author thinks of it first for inference (answering a request), although model training could benefit from it too.

So it is neither a ranking file nor a permission file. The three files at the root of a site do not have the same role:

The proposal states that llms.txt does not replace the sitemap: a sitemap is too big to fit in a model’s context window, and it does not hold the AI-readable versions of pages.

Who publishes an llms.txt file?

Of the 26 sites we tested on October 21, 2026, 12 serve a real llms.txt file, 12 return a 404 and 2 refuse the request. Our sample is chosen, not random: it says nothing about the real share across the web.

To avoid counting fake pages, we required a 200 status, a non-HTML content type and content that starts with ”#”. All four AI vendors publish the file for their documentation, but Google Search’s documentation does not. On a larger scale, SE Ranking found a file on 10.13% of nearly 300,000 domains, with adoption between 8 and 11% whatever the site’s traffic.

Do AIs and Google read the llms.txt file?

For Google, the answer is written down: its guide to optimizing for generative AI, updated on July 10, 2026, lists llms.txt files among the things you can ignore. Google Search does not use them, and publishing one neither hurts nor helps its visibility. For OpenAI, Perplexity and Anthropic, the pages we read about their bots do not say they use it to choose sources.

Two studies measure the effect, with their limits:

Neither study proves that llms.txt is useless everywhere, notably for coding assistants that read documentation. They only show that, for being cited in search answers, there is no sign of an effect today.

Should you publish an llms.txt file?

It depends on your site. If you publish technical documentation or an API, that is the proposal’s use case, and the file can help the coding agents that consult it. If your site is a media outlet or a comparison site, nothing in the sources we read shows a gain in citations: do not spend time on it, and do not sell it as a guarantee.

Our own file is generated automatically from our articles, in the three languages, with their descriptions. It costs us nothing to maintain, and we expect nothing from it. We have not measured whether bots request it.

If you decide to publish one, three points matter: an accurate title and summary, links to the pages you really want read, and a curl check that it is served with a 200 and content that starts with ”#”. A badly configured site can return an HTML page with a 200 for an address that does not exist, which looks like a file when there is none. None of the 26 sites in our survey did, but the content check is what guarantees it.

What should you do instead to get cited?

The sources we read agree on the rest: indexed and crawlable pages, a robots.txt that does not block the AI search bots, and content that brings its own point of view. Our free GEO audit guide shows how to check these points and measure whether a brand is cited. To track the result every month without doing it by hand, we compared four tools in our comparison of AI visibility tracking tools, and our pick to get started is Hikoo, which listed a 7-day trial with no credit card on October 12, 2026.

What we have not verified

• The effect of an llms.txt on ChatGPT, Perplexity, Claude or Gemini: we ran no test, we read studies and documentation.
• Whether AI bots request our site’s llms.txt: we have not analyzed our logs.
• The method and raw data of SE Ranking and OtterlyAI: we read their pages and did not reproduce their measurements.
• How representative our 26 sites are: the sample is chosen, and we only tested the root of each domain.
• The OpenAI, Perplexity and Anthropic documentation on their bots, as of October 18, 2026.
• Demand outside France: the search volume we cite (1,000 to 1,300 per month depending on how “llms txt” is written) comes from a tool that only covers France.