llms.txt examples: 10 real SaaS sites and what they get right

These are ten SaaS llms.txt files fetched on 13 October 2026. Each note describes the HTTP response from that day: the heading, the blockquote if there was one, and how the links are organized. A company can edit or remove the file tomorrow. Nothing here claims that publishing the file caused a citation. No answer engine has published llms.txt as a ranking requirement.

The proposal at llmstxt.org asks for a Markdown file, usually at /llms.txt: an H1, an optional blockquote that defines the project, then H2 sections of links. A note after a colon tells a model which link answers the question in front of it. llms-full.txt is a separate convention, a concatenated export, and it is not in the proposal. CiteScore’s own check is smaller than either idea. The free check and the $39 audit only ask whether /llms.txt returned a non-empty text body.

What was fetched

Counts below are from those responses. “Noted” means a bullet whose Markdown link is followed by a colon and a note.

1. Vercel — a short file a person can finish

https://vercel.com/llms.txt returned text/plain, 65 lines, 25 Markdown links. All 25 have a note. The H1 is “Vercel”. The blockquote is one sentence: “Vercel is a cloud platform for building, deploying, and scaling web applications and AI workloads.” A paragraph under it, with no heading, tells an agent to follow the linked indexes. Later H2s split documentation, the REST API, and an Optional section. The Optional links are secondary machine files: an agent catalog, a product taxonomy, and a documentation graph. The documentation section points at a full export, https://vercel.com/docs/llms-full.txt.

That is the shape to copy when you have a real product and a handful of canonical URLs. The file is written for a coding agent — CLI, MCP, API auth — more than for a buyer asking what Vercel costs. That is a choice of audience. The definition in the blockquote still works for both.

2. Stripe — one product per section

https://stripe.com/llms.txt returned 402 lines and 308 Markdown links. The H1 is “Stripe”. The blockquote starts “Stripe is a technology company that provides financial infrastructure for businesses.” H2s follow the catalog: Payments, Connect, Checkout, Billing, Radar, Issuing, and others — 31 sections. Almost every bullet is a titled link with a colon and a note. Docs links often end in .md.

A second file, https://docs.stripe.com/llms.txt, returned text/markdown, 710 lines, and 513 Markdown links, 505 of them ending in .md. It has no blockquote. It does include an H2 titled “Instructions for Large Language Model Agents: Best Practices for integrating Stripe”, and the prose under it tells an agent which guides to prefer and which API object to default to. The product file is a map of the company. The docs file is a map of the integration. They answer different questions, and Stripe publishes both.

The limit is size. Four hundred lines is a catalog. The notes on the product file are marketing sentences, so a model still has to choose. A smaller SaaS can copy the pattern — H1, definition, one H2 per product, a note on every link — and stop at the pages a buyer actually needs.

3. Cloudflare — an index of indexes

https://developers.cloudflare.com/llms.txt returned 145 lines and 113 Markdown links. Every link has a note. The links point at another llms.txt, one per product (Cache, DNS, Workers-related products, and the rest), grouped under nine H2s such as Application performance and Application security. The blockquote says so directly: each product links to its own file, and that file is the recommended way to explore that product. A sentence sits between the H1 and the blockquote, so the summary is not the first line after the title.

This is the right pattern when one host documents dozens of products. Inlining every page would make a file no model will read in full. Pointing at a second index keeps the top file small. The tradeoff is that an agent has to fetch twice before it sees a doc page.

4. Anthropic — coverage, and a search tool beside the list

Requesting https://docs.anthropic.com/llms.txt on 13 October 2026 returned a 301 to https://platform.claude.com/docs/llms.txt, which returned a 308 to /llms.txt. The file that answered is https://platform.claude.com/llms.txt: 834 lines, H1 “Anthropic Developer Documentation”, no blockquote, 761 Markdown links, 760 of them ending in .md. The file says the English docs are “760 pages” and that those pages are the ones included. Other languages are named with their page counts, and the file tells the reader to visit the site for that content. It also points at an MCP server on https://platform.claude.com/docs/mcp for search, instead of pretending the flat list is the only way in.

Coverage is the thing this file gets right. The gaps are the ones a strict reading of the proposal would flag: no blockquote, and the body uses H3s under an “English” H2. Many links have no colon-note. It is a generated tree of the docs, useful as a corpus index, thin as a one-screen definition of Claude. It describes the developer platform, not the consumer chat app.

5. Supabase — the full export is on line 3

https://supabase.com/llms.txt returned 56 lines and 42 Markdown links. There is no blockquote. The first sentence after the H1 is the useful one: the complete documentation is at https://supabase.com/llms-full.txt. An agent that wants everything does not have to discover that URL. The rest of the file is a short section list (Auth, database, storage, and so on), then per-language reference files, including JavaScript, Dart, Swift, Kotlin, and Python, then pricing and a Management API OpenAPI spec.

Most bullets have a title and no colon-note, and the titles repeat “Supabase -”. The file is easy to scan because it is short, and because the full export is labeled. It is a weak definition of the product. The homepage still has to say what Supabase is.

6. Perplexity — the API, summarized in the blockquote

https://docs.perplexity.ai/llms.txt returned 180 lines and 170 Markdown links. 163 end in .md. 155 have a note. The H1 is “Perplexity”. The blockquote names the product this host actually documents: “Perplexity API documentation for building with the Agent API, the default for web-grounded AI and multi-provider applications, plus Search and Embeddings APIs.” The H2s are “OpenAPI Specs” and “Optional”.

The notes are specific enough to choose a page (“Generate an API key and make your first call”). The structural miss is that the long link list sits before the first H2. The proposal puts file lists under H2 sections. This file is the API docs, not the consumer Perplexity search product.

7. Zapier — a map of the product site

https://zapier.com/llms.txt returned 180 lines, 116 Markdown links, and eight H2s. The blockquote is a navigation note, not a product definition: it points at a Surfaces section and a Topics section and says the anchors are stable. The definition lives under “What is Zapier”, in prose. Surfaces points at https://docs.zapier.com/llms.txt for the developer docs and names the help center and community hosts. “Notes for automated fetchers” tells an agent to fetch /llms-full.txt for expanded content, warns that the file is not a pricing spec, and lists the anchors that will not be renamed without a coordinated update.

That is the rare non-docs example in this set. It treats llms.txt as a map of a marketing site with several subdomains. It also uses H3s and horizontal rules, which the proposal does not ask for. The useful habit is the one in the fetcher note: say what the file is not, and say which other file to open next.

8. Linear — docs grouped by the job

https://linear.app/llms.txt returned 198 lines and 136 Markdown links, 134 ending in .md. The H1 is “Linear”. The blockquote defines the product: “Linear is a purpose-built tool for planning and building products.” Two H2s, Documentation and Developers, contain H3 groups a user would recognize: Getting started, Intake, Projects and initiatives. The links do not have colon-notes. The title of the link is the only hint.

Grouping by job is the part worth copying. A flat alphabetical dump makes a model read titles until one sounds relevant. “Intake” and “Projects” are already the questions. Adding a colon-note would make the same file easier to route, and the file does not do that today.

9. Notion — organized the way a buyer asks

https://www.notion.com/llms.txt returned 74 lines and 48 Markdown links. All 48 have a note. None of the URLs end in .md. The H1 is “Notion”. The blockquote starts “Notion is the connected workspace where better, faster work happens.” H2s follow a buyer’s path: Core Products, Solutions By Team Types, Solutions By Role, Getting Started. A separate file at developers.notion.com/llms.txt is the developer docs. The marketing file and the API file are not the same map.

The notes are short product descriptions, which is what a buyer-facing file needs. Under the blockquote, the file also prints a “Key terms” line of keywords. That line does not help a model choose a URL. The sections do.

10. Resend — the definition, then the request

https://resend.com/llms.txt returned 162 lines and 57 Markdown links. The H1 is “Resend”. The blockquote is the definition: “Resend is the email API for developers. Send transactional and marketing emails at scale with a simple, modern API.” The next H2, “When to use Resend”, is written as tasks, and the tasks name the HTTP call. Sending a transactional email is POST https://api.resend.com/emails. Batch send is named with its limit. Later sections point at a setup guide, the CLI, an MCP server, and an OpenAPI spec. A larger docs index lives at https://resend.com/docs/llms.txt. The root file is the curated one.

This is the example to copy if you do not have 700 documentation pages. Say what the product is. Say when to use it. Point at the three URLs an agent will actually need. Most of Resend’s bullets have no colon-note. The “when to use” prose carries that information instead, which works because the list is short.

A file fetched the same day that misses the menu

https://cursor.com/llms.txt also returned text/plain on 13 October 2026, 497 lines, H1 “Cursor Documentation”, no blockquote. The links are bare URLs. There is no title and no note, so a model learns the path and nothing about the page. One changelog bullet is the literal string https://cursor.comhttps://cursor.com/changelog.md — the host is doubled. Other changelog links in the same file are well-formed. Markdown URLs are the useful habit in this file. Missing titles, and a broken host, are the miss. It is still a real SaaS file. It is not one of the ten above because the ten are there for a choice worth copying.

What to copy, and what to leave

• One sentence that defines the product. Vercel, Resend, Linear, and Stripe’s product file all do this in a blockquote. Anthropic’s developer file and Supabase’s file do not, and the homepage still has to carry the definition.
• A note on the link. Vercel, Stripe’s product file, Cloudflare, and Notion annotate every bullet. Linear and Supabase mostly do not. The note is the routing instruction.
• Markdown when you have it. Anthropic, Perplexity, and Linear point at .md pages. Notion’s marketing file points at HTML. Both are legal. Markdown saves the model a scrape.
• Do not inline a catalog you would not read. Cloudflare links to per-product files. Supabase and Zapier point at a full export from the short file. Stripe’s docs file is large on purpose and then adds a section of instructions. A generated dump of every URL, with no notes, is the Cursor failure mode.
• Say which host is which. Stripe, Notion, Resend, and Zapier publish more than one file. The short file should name the others.

Leave the keyword line. Leave the claim that the file itself gets you cited. The llms.txt checker page is the exact HTTP check CiteScore runs, including what a missing file does to the score. The generator copies the H1, blockquote, and annotated-link shape into a file you can edit. It does not upload anything.

Check the file, then score the page

The free check fetches /llms.txt on the host you enter and reports whether the body is non-empty text. It also reads robots.txt for GPTBot, PerplexityBot, and Google-Extended, plus the sitemap and JSON-LD on the homepage. It does not query ChatGPT, Perplexity, or Google AI Overviews, and it does not grade the writing inside the file.

The $39 audit scores one public URL. A found llms.txt is a positive mark on the crawler dimension only. The page the file points at still has to answer the question. The AI Citation Readiness Checklist is the unpaid list of those checks, including the six that apply to the file itself.