Real llms.txt Files, Annotated
Five real llms.txt examples from Vercel, Zapier, Cloudflare, Stripe, and Anthropic, fetched and counted, showing curated versus dumped in practice.
Live llms.txt files split into two patterns: a short, curated index like Vercel's 64-line file with 24 hand-picked links, or a full documentation dump like Stripe's 697-line file with 443 links and no summary, and most large companies currently ship the dump.
Most "llms.txt examples" content on the web shows a hypothetical template. Here are five real files, fetched directly and counted, from companies that actually ship one. The gap between what the spec recommends and what gets shipped in practice is the useful part.
The curated end: Vercel
vercel.com/llms.txt is 64 lines. It opens with an H1, a one-sentence blockquote ("Vercel is a cloud platform for building, deploying, and scaling web applications and AI workloads"), a short paragraph on when to use the product, and then 24 links spread across headings including "How agents should use Vercel," "Documentation," and "REST API and developer resources." Every link carries a short description. This is the closest of the five to what llmstxt.org's spec actually describes: a short, hand-picked index, not a sitemap.
The dump end: Stripe and Anthropic
docs.stripe.com/llms.txt is 697 lines: no blockquote, straight from the H1 into instructions for coding agents, then 26 H2 sections (Docs, Payment Methods, Checkout, Billing, Connect, Tax, and more) holding 443 individual links. It reads like the full documentation sitemap, because it is one.
platform.claude.com/docs/llms.txt (the current location; docs.anthropic.com/llms.txt redirects here) is 698 lines and structured differently but lands in the same place: a language list, then a single "English" section holding 628 individual doc-page links with no summary and no curation beyond alphabetizing by product area.
Neither file is broken. Both are valid markdown under the spec. Neither is what "curated index" means in the format's own description.
The hub-of-hubs pattern: Cloudflare and Zapier
developers.cloudflare.com/llms.txt takes a third approach: 106 product areas listed at the root, each one a link to that product's own llms.txt file (developers.cloudflare.com/dns/llms.txt, .../cache/llms.txt, and so on) rather than a page-level link. zapier.com/llms.txt does something close: 180 lines with 24 top-level links, several of which point to a separate docs.zapier.com/llms.txt for the full developer index, and the file explicitly notes its section anchors are "stable, safe to deep-link." This pattern fits large multi-product sites where one flat file would need to be a Stripe-sized dump to cover everything.
What this actually tells you
Five files, three shapes: one short and curated (Vercel), two long and effectively full sitemaps (Stripe, Anthropic), two structured as an index of indexes (Cloudflare, Zapier). None of this says which shape gets cited more, because no study has found llms.txt length or format correlated with AI citation at all. What it does tell you is that "follow the spec" and "match what large companies actually ship" are two different instructions, and most of the real files above chose convenience over the spec's own advice to curate. If you are building one from scratch, the actual construction steps are covered here; a GEO audit checks whether a file like this is even the right thing to spend time on for your site.
Frequently asked
What does a well-formed, curated llms.txt actually look like?
Vercel's file at vercel.com/llms.txt is a clean example: 64 lines total, an H1, a one-sentence blockquote summary, a short explanation of when to use the product, and 24 links grouped under headings like Documentation and REST API and developer resources. Every link fits the spec's intent of a short index rather than a full sitemap.
Do large companies actually follow the spec's advice to keep the file short and curated?
Not consistently. Stripe's file at docs.stripe.com/llms.txt runs 697 lines across 26 sections with 443 links and no blockquote summary at all, and Anthropic's file at platform.claude.com/docs/llms.txt is 698 lines listing 628 individual documentation pages under a single English section. Both are technically valid, but neither is curated in the sense the spec describes.
What is the hub-and-spoke pattern some companies use?
Cloudflare's root file at developers.cloudflare.com/llms.txt does not list individual pages at all. It lists 106 product areas, each pointing to that product's own llms.txt, for example developers.cloudflare.com/dns/llms.txt. Zapier does something similar across subdomains, with its root file at zapier.com/llms.txt pointing readers to docs.zapier.com/llms.txt for developer content. A model or agent fetches the root file first, then follows the link into whichever specific area it needs.
Does a longer llms.txt file mean a company gets cited more by AI engines?
There is no evidence for that. A 300,000-domain study covered in our llms.txt explainer found no correlation between having the file, at any length, and being cited more by AI answer engines. Stripe's 443-link file and Vercel's 24-link file are both just as unproven as citation levers; length is not the variable that has been shown to matter.
How can I check a company's actual llms.txt file myself?
Type the domain followed by /llms.txt directly into a browser, for example stripe.com/llms.txt or vercel.com/llms.txt. It is a plain text file at the site root, no login or special tool required, and most large documentation sites that have one make it publicly reachable.
More in GEO & AI Search
How to Generate a Correct llms.txt
An llms.txt generator can produce a file that fails the spec. Here is the one required line, the mistakes auto-generators make, and how to build a correct one.
AEO vs SEO vs GEO: The Three-Letter Problem
AEO vs SEO: AEO targets the answer inside a snippet or AI reply, SEO targets page rank, and no one agrees where AEO ends and GEO begins.
What a GEO Audit Actually Checks
A GEO audit checks whether AI engines name your brand, cite your site, beat you to the answer, and whether crawlers can even read your pages.
See how AI describes your brand
A free GEO audit takes about 90 seconds and tells you what ChatGPT, Claude, Perplexity and Gemini actually say about you.
Run a free GEO audit →Written by YaBaDo's AI agents and reviewed by a human before publishing. If you find something wrong, tell us and we will correct it.