How to Write an llms.txt File

Why it matters

A sitemap lists every URL; llms.txt does the opposite — it is a deliberately curated shortlist of the pages that actually carry your knowledge, written for an AI reader. Placed at /llms.txt, it is meant to let a crawler skip the navigation, marketing, and boilerplate and go straight to the concepts, methods, proof, and tools you most want cited. Be clear-eyed, though: it is an open proposal, and no major AI provider has confirmed it reads the file — Google explicitly lists AI-specific text files like llms.txt as not required.

So why do it? Two reasons that don't depend on any crawler consuming the file. First, writing a good llms.txt forces you to identify and articulate your canonical pages — the same curation discipline that sharpens everything else in the framework. Second, if the convention is picked up by emerging AI tools, a substantive file is ready to help them where a near-empty stub would not. Set expectations honestly with clients: it is cheap insurance and a useful exercise, not a confirmed source of AI visibility — the fundamentals in Phases 2–6 are what actually earn citations.

What good looks like

A strong llms.txt reads like a well-organised table of contents for a machine — Markdown, grouped by content type, every link annotated:

• Lives at the domain root: https://yourdomain.com/llms.txt, served as text/plain or Markdown, returning 200.
• Opens with an # H1 (your organisation/site) and a one-sentence description of what you publish.
• Groups links under typed ## headings — we use Concepts, Methods, Proof, Tools (our convention; the proposal only requires ## sections of links, plus an optional ## Optional section for secondary pages).
• Each link has a short annotation saying what the page is and why it matters.
• Covers your canonical pages — as a rule of thumb, 20 or more on a content site — and stays consistent with your sitemap (Phase 5.5).

Step by step

1. Create the file at the domain root Save a plain-text/Markdown file as llms.txt and serve it from the root so it resolves at https://yourdomain.com/llms.txt. Like robots.txt, it must be at the root — not in a subfolder.
2. Open with an H1 and a one-line summary Start with # Your Organisation and a single sentence describing what you publish and who it's for. This is the orientation line an AI reads first — make it a direct, complete statement (the Phase 2.1 direct-answer habit).
3. Group links by content type Use ## headings to sort entries into Concepts (definitions, frameworks), Methods (how-to, process), Proof (research, case studies, data), and Tools (calculators, templates, products). Typed grouping tells the AI what role each page plays.
4. Annotate each link Write each entry as a Markdown link followed by a short description: - [Page title](https://…): one line on what it is and why it's canonical. The annotation is what lifts a link from a URL to a usable signal.
5. Cross-reference the sitemap and keep it current The high-signal pages in llms.txt should match the priorities in your sitemap.xml — if they diverge, the AI gets a partial map of your authority (Phase 5.5). Revisit the file whenever you publish or retire a canonical page.

Worked examples

# AI Knowledge Signal
> The AI Knowledge Signal Publication Framework: a 6-phase method for structuring content so AI systems can find, trust, and cite it.

## Methods
- [Configure robots.txt for AI crawlers](https://aiknowledgesignal.io/framework-guides/phase1-crawlability/robots-txt.html): separate training, search and user-fetch agents; declare the sitemap.
- [Implement JSON-LD schema](https://aiknowledgesignal.io/framework-guides/phase5-ai-usability/json-ld.html): mark up entities so AI can disambiguate them.

## Concepts
- [Framework How-To Guides](https://aiknowledgesignal.io/framework-guides/): all 30 subdomains across the six phases.

## Tools
- [Free AI visibility audit](https://aiknowledgesignal.io/): score any URL across five dimensions.

## Optional
- [Newsletter](https://aiknowledgesignal.io/newsletter): monthly GEO and AI-search research digest.

Common mistakes