llms.txt example: a practical guide to machine surfaces
See a real llms.txt example, annotated line by line, and learn how Markdown, sitemaps and crawler rules fit together without Google ranking promises.
An llms.txt file is a short guide to useful content on your website. Start with a specific need: helping a documentation reader find the right reference, or giving an agent a maintained set of product instructions. Build it from pages you already trust and can keep current.
What llms.txt does, and where it fits
The llms.txt proposal describes a Markdown overview with links to further material. Its current version allows a file at the site root or within a directory, covering that path. A project heading is required; a summary and grouped links add context. This is an optional publishing convention.
Limits: llms.txt is not a Google ranking factor. Google's AI optimization guidance says these files neither help nor harm Search visibility or rankings. Publishing one does not establish that an assistant fetched, used, or cited it. Before-and-after answer samples cannot prove that the file caused a change.
Choose the reader before choosing the format. A developer looking for an authentication example needs a different starting point from a buyer checking supported integrations. Write down the few questions your file should help answer, then select pages that address them directly.
Keep established mechanisms in their own roles
- robots.txt: Publish crawler preferences using the Robots Exclusion Protocol. The protocol is not access authorization. Keep private material behind actual access controls.
- Sitemaps: List discoverable URLs. Google's sitemap guidance explains their discovery role. Use one to organize search discovery across the site.
- Structured data: Describe the visible page. Google's structured data documentation explains supported markup formats. Keep the markup consistent with what a reader sees.
- Markdown: RFC 7763 registers
text/markdown. Use it when serving a Markdown representation, and check that the response contains the intended content. - Cite-As: RFC 8574 defines a relation for expressing a preferred reference URI. Consider it when a resource has a separate permanent identifier.
- IndexNow: Send URL-change notifications to participating engines. The protocol documentation identifies 200 as successful submission and 202 as receipt with key validation pending. Label these as submission responses.
Keep a small inventory showing which mechanism serves which job. For example, assign the sitemap to URL discovery and the documentation index to guided reading. This prevents two files from becoming competing lists with different owners, review dates, and descriptions of the same product.
A real llms.txt example, annotated line by line
The FastHTML documentation file is a real published example. The following three selected lines retain their original wording and order; the intervening introduction and other entries are omitted. They show the project name, a resource group, and one destination.
# FastHTML
## Docs
- [FastHTML concise guide](https://www.fastht.ml/docs/ref/concise_guide.html.md): A brief overview of idiomatic FastHTML apps
- Line 1: The single hash introduces the project name. On your own site, use the name readers will also see on the documentation homepage.
- Line 2: Two hashes open a group of related resources. Choose a heading that helps someone decide which group to inspect.
- Line 3: The list item combines a readable label, destination URL, and short note. Here it points to a Markdown guide. For your file, describe the question that each destination resolves.
Inspect the full file before adapting the pattern. It also includes project context and other resource groups. Copy the organization only where it fits your material. A booking service might group setup instructions and cancellation policies; a software library might group tutorials and reference pages.
Keep annotations outside the production file unless they help its reader. In this article they explain the syntax; on your site, the useful content is the description of your own resources. Replace every sample destination with a real page and open each link before publishing.
Choose the pages before generating the file
Begin with a short selection that you can inspect manually. Include the primary product explanation, the setup guide, the reference page for common questions, and any conditions that affect use. Leave out duplicate announcements and pages whose claims you have not checked.
For each entry, compare its label with the destination's opening paragraph. If the label promises an installation guide but opens a sales page, fix that mismatch. If several entries answer the same question, choose the clearest maintained source instead of adding them all.
Assign an owner to the file and include it in the content review for product changes. When an integration is removed, revise its documentation and the index together. Keep a short change note so the next reviewer understands why a page was added or replaced.
Maintain Markdown and HTML together
The proposal also describes Markdown alternatives to HTML pages and links that help clients discover them. Generate both representations from the same source where possible. Compare headings, examples, dates, and limitations after a content update, rather than assuming a successful build preserved their meaning.
Check tables and code samples carefully. Open the Markdown URL directly, read its response, and follow its links. For a small pilot, ask a colleague to find a specific instruction using only the index. Record unclear labels and missing destinations, then repair those problems before expanding the file.
Review answers alongside the publishing work
Keep a dated record of the file, selected pages, and test questions. Review complete responses when checking whether your product was described accurately. Attach an observed link to the answer that returned it, and keep server requests in a separate technical record.
SearchSeal checks tracked questions on the AI tools listed on its current pricing page, saving answers to tracked questions with brand mentions, position, rivals, and returned source links. Explicit citation status remains unknown. Optional, read-only Google Search Console and Bing Webmaster connections add demand per page to help rank page fixes alongside those saved answers.