llms.txt: What the Proposal Does and Does Not Do (Part 1 of 3)

llms.txt is a proposal for publishing a concise Markdown overview of a website, with links to useful content. It gives a reader or a tool that chooses to fetch it a curated starting point.

Publishing the file does not establish that ChatGPT, Claude, Gemini, or Perplexity will discover it, read it first, or use it when recommending your business.

Correction, October 1, 2026: this article originally called llms.txt “the file AI reads first” and said it removed model guesswork. Those claims were not supported. We have revised the title and explanation while keeping the original URL.

A short overview can be useful when you explicitly give it to an agent, or when a documentation tool supports the format. You choose the summary and links, and the reader can follow them for detail. That is a publishing use case; it is not evidence of a search ranking benefit.

llms.txt does not grant crawler access, guarantee indexing, or control an assistant’s answer. Keep crawler access rules in robots.txt, and keep important facts available on the public pages people actually read.

Google says AI Overviews and AI Mode have no additional technical requirements beyond Search eligibility. Its guidance says new AI text files and special structured data are not required. A missing llms.txt is therefore not evidence that your site cannot appear in those features.

The proposal uses Markdown. Its format starts with an H1 naming the site or project; that is the only required section. The optional parts provide context and routes to further information.

• An H1 with the site or project name.
• An optional blockquote summary.
• Optional explanatory text beneath the summary.
• Optional H2 sections with Markdown links and short notes.

Use accurate descriptions and public, working links. Prefer clean text or Markdown editions where available. The current proposal also allows a file at a subpath, such as /docs/llms.txt, to cover content beneath that path.

Our S&P 500 audit, published July 26, 2026, detected files on 71 of 500 company websites. Fifty-four had both an H1 and at least one Markdown link. That was our practical structure threshold, not a claim that links are mandatory under the proposal.

The audit measured file presence and structure. It did not test whether assistants read those files or whether publishing one caused more mentions, citations, or recommendations. Inaccessible websites remained a separate outcome from confirmed absence.

If a concise content overview would help your documentation users or an agent workflow you can test, publish one and maintain its links. If your goal is search visibility, first check that your important pages are accessible, accurate, and easy to find.

Check the file separately from your brand’s actual visibility. A structural checker can report what is published; it cannot prove that a provider uses it or that it changes an answer.