llms.txt (llmstxt.org, v2) is a proposal: a Markdown file at /llms.txt, with only the H1 line required, for agents to read when needed while helping a user; the proposal page does not state which AI answer apps (ChatGPT, Claude, Gemini, Perplexity) read it.
The llms.txt page describes a proposal, not a standard: a Markdown file at /llms.txt where only the title line is required, meant for agents to read when they need it. The page names no AI answer app that reads it.
- A site publishes /llms.txtMarkdown; only the H1 line is required
- An agent needs information while helping a userUsed on demand, mainly for inference rather than training
- The agent reads the fileThe proposal page names no AI answer app that does this
Source: llmstxt.org, v2 of the proposal, read on 22 October 2026. What the page says; it is not evidence of what any app does.
Why I went back to the source
Earlier this month I watched training crawlers fetch the llms.txt on my lab site, while AI search crawlers and the fetchers that act for users never touched it. Before telling anyone to write one, I wanted to know what the file actually promises. On 22 October 2026 I read llmstxt.org. The page says: "This is v2 of the proposal".
What it says
- It is "a proposal to standardise on using an
/llms.txtfile to provide information to help agents use a website". - It "would mainly be useful for _inference_ rather than _training_". The file is "used on demand, when an agent needs information about a topic while assisting a user".
- The file sits at the root of a site or at any subpath.
- Format: an H1 with the name of the site, which is "the only required section". After that, optionally, a quoted summary, some paragraphs, and
##sections that list links as[name](url). A section called Optional holds what an agent can skip.
What it does not say
Who reads it. The page says thousands of sites publish an llms.txt, that AI companies publish one for their own developer docs, that Chrome Lighthouse checks for the file, and that there is an MCP server agents can use to fetch it. It does not say that ChatGPT, Claude, Gemini or Perplexity read the file when they answer a question.
There is a small irony here. The proposal is aimed at agents at answer time. On my site the readers I could see were training crawlers, which took the file as one more URL in a crawl.
What you can do
- Treat llms.txt as optional. Make sure your important content is in HTML pages listed in your sitemap first.
- If you write one, keep it small: the H1, one line of summary, and links to your key pages.
- Put each URL on its own line, with nothing glued to it. Some crawlers read the file as raw text, and a
）or。right after a URL becomes part of the address.
My scanner's advice on llms.txt says the same: it is an optional step, and it is not established that having the file changes what an AI answers.
What this does not prove
- A proposal page is not evidence of what any app does.
- The page not naming a reader does not prove that no app reads the file.
- My own log observations come from one site, where AI search crawlers barely visit at all, so they cannot settle the question either.
More on what AI crawlers read on my lab site: Does AI read JS, JSON-LD and llms.txt?
Sources
The measurements behind this are on Parallel World
What this does not prove
the proposal page is not evidence of any app's behavior; the page not stating which apps read the file does not prove that no app reads it.