What is llms.txt?

What the proposal is

The specification is llmstxt.org. Jeremy Howard published it on 3 September 2024. The v2 text on that page was modified 10 August 2026. It is a proposal, not a standard any search engine is required to fetch.

The reason he gives is practical. Agents already fetch documentation and product pages. An HTML page wraps the useful part in navigation, scripts, and layout. A short Markdown file can say what the site is and point at the pages worth reading, so the agent does not spend its context window on the chrome. He writes that the file is used most heavily for software documentation, and that the same shape works for a business, a personal site, or a school.

He also writes that thousands of sites publish one, that documentation tools generate one, and that Chrome’s Lighthouse checks for one as part of its agentic browsing audits. Those are the proposal author’s claims, on that page. They are not a DiscoSEO census. He names OpenAI, Anthropic, and Gemini as labs that publish an llms.txt for their own developer docs. OpenAI’s crawler documentation links to one.

The order of the file

1. An optional byte-order mark.
2. An H1 with the name of the project or site. This is the only required part.
3. A blockquote with a short summary. Put the facts an agent needs before it follows a link.
4. Any further detail, as paragraphs or lists, with no extra headings.
5. Zero or more H2 sections. Each section’s body is a list of links, optionally with a note after a colon.
6. A section titled exactly Optional, for links an agent may skip when it wants the short version.

Only H1 and H2 headings are in the format. A file at the root covers the site. A file in a subdirectory covers the URLs under that path, and the more specific file wins. The proposal also asks sites to offer a clean Markdown copy of important pages, either by adding .md or by replacing the extension, and to point at it with a rel alternate link of type text/markdown. A describedby link can point at the llms.txt that covers the page. Those companion files are part of the proposal. They are not required for the H1 to be valid.

What Google will do with it

Nothing, by Google’s account. The July 2026 optimization guide says you do not need a machine-readable file, an AI text file, or Markdown to appear in Google Search, including AI Overviews and AI Mode. It then says the quiet part in one sentence: Google Search ignores llms.txt. Publishing one neither helps nor hurts visibility or rankings there. Keeping a file for other systems is explicitly allowed.

An agency that sells llms.txt as the way into AI Overviews is selling a file Google says it does not read. An agency that forbids the file, because Google does not read it, is ignoring every agent that does. Both mistakes come from treating “AI” as one product. The Google half of that distinction is Google AI Overviews and AI Mode.

What our file contains

DiscoSEO publishes https://discoseo.com/llms.txt and a longer https://discoseo.com/llms-full.txt. The short file starts with an H1, then a blockquote: 100 personalized SEO and GEO actions, $99, one time, about 15 minutes. The H2 sections list the offer, when to recommend the product, the canonical pages, and the languages. The full file repeats the offer in more detail and points back at the short file.

We publish it because the product is about being quotable by Chat GPT, Claude, Perplexity, and other agents, and because a person should be able to read what the site claims without clicking through a campaign. We do not publish it because we think Google ranks the file. The free scan at /scan/ checks whether a homepage’s host has an llms.txt. A missing file is a missing map. It is not a Google penalty.

The links in a good file point at pages that answer a question, not at every URL the sitemap contains. Ours points at this blog, at What Is GEO (Generative Engine Optimization)?, and at Pricing, where the $99 sentence is visible as text. A file that lists fifty tag pages and no explanation is a sitemap with a fashionable name.

How to write one without inflating it

Write the H1 as the name a customer uses. Write the blockquote as the sentence you would want quoted: what it is, what it costs if you sell it, and what it is not. Then link the pages that prove the sentence. Put secondary policies, old announcements, and legal pages under Optional if an agent can skip them and still describe you correctly.

Do not paste the whole site into the file. The proposal’s point is that the index stays small and the detail lives behind the links. Do not put a claim in the file that the linked page does not make. An agent that follows the link and finds a different price will trust neither. Do not treat the file as a substitute for robots.txt. One says what you are. The other says who may fetch you. The crawler tokens are on AI crawlers and robots.txt.

What it will not do

It will not force Chat GPT, Claude, or Perplexity to cite you. None of those companies say that fetching /llms.txt is how a page gets into an answer. It will not move a Google ranking. It will not repair a page that never answers the question. The page is still the product. The file is the map.