When an AI assistant lands on your website, it doesn't see your nav, your hero animation, or your carefully art-directed homepage. It sees markup — and it has a token budget. llms.txt is a small, standardized file that hands the model a clean, high-signal map of your site so it spends that budget on the right things.
It sits at the root of your domain — https://yourbrand.com/llms.txt — and it is plain Markdown. Think of it as a robots.txt for meaning instead of crawling: where robots.txt says what a bot may fetch, llms.txt says what matters and where to find it.
Why it matters for AI visibility
Models and agents increasingly fetch a page (or a handful of pages) and reason over the text. If your most important, most quotable facts are buried under scripts and boilerplate, they may never make it into the model's working context — and you don't get named in the answer. A good llms.txt is a shortcut to the facts and pages you most want cited.
It's also a clear, cheap signal that you've thought about being machine-readable at all — which most of your competitors haven't.
The format
The llmstxt.org spec is deliberately minimal:
- An H1 with the name of the site or project (the only required line).
- An optional blockquote (
>) with a short summary. - Optional free-text paragraphs with more detail.
- Zero or more
##sections, each a list of Markdown links —- [Name](url): optional note.
The key discipline is in step 4: sections should be link lists, not prose. Each bullet points the model at a real URL it can fetch.
A worked example
# Acme
> Acme makes precision torque tools for aerospace assembly lines.
> Trusted by 40 OEMs across 12 countries.
Founded 2009, HQ in Geneva. Sold direct and through certified distributors.
## Key pages
- [Products](https://acme.com/products): the full torque-tool catalog with specs
- [Pricing](https://acme.com/pricing): list pricing and volume tiers
- [Documentation](https://acme.com/docs): calibration and integration guides
## About
- [Company](https://acme.com/about): history, certifications, leadership
- [Contact](mailto:hello@acme.com): sales and support
That's a complete, valid llms.txt. A model reading it knows what Acme is, what it's known for, and exactly where to go for products, pricing, and docs — in a few hundred tokens.
Common mistakes
- Prose instead of links. A
##section that's a paragraph (or plain bullets with no URLs) gives the model nothing to fetch. Use- [Name](url)lists. - Marketing fluff. "World-class, best-in-breed solutions" is invisible to a model. State concrete, checkable facts: what you make, who you serve, a proof point with a number.
- Letting it drift.
llms.txtis content, not config — when pages move or facts change, update it. A stale map is worse than none. - Skipping the summary. The blockquote is the single highest-leverage line; it's often the sentence a model paraphrases when describing you.
Where it fits
llms.txt is one signal among several — structured data, a clean entity identity, quotable claims, and crawler access all matter too. It's the easiest of them to ship: one file, a few hundred words, an afternoon. It won't single-handedly make you the answer, but it removes a stupid reason for being skipped.
If you want to see how your site reads to an AI today — including whether your llms.txt, schema, and claims are doing their job — run a free Legible report. It scores your machine-readability across eight dimensions and hands you the fixes.