Open ten llms.txt files at random, and nine of them are a sitemap wearing a trench coat.
A heading, a wall of URLs, no descriptions, no ordering, nothing a machine could not have worked out by crawling the site itself.
The tenth is different, and the difference is instructive. Stripe’s file opens by telling agents not to trust memorized package version numbers and to check the registry instead.
Zapier’s includes routing rules and verification steps.
Vercel’s links to a product taxonomy listing canonical names, aliases, and deprecations. These are not indexes. They are briefing documents for a reader who arrives with confident, outdated assumptions.
Those distinctions are more important than the debates usually surrounding this file. The debate is whether llms.txt improves AI visibility, and the honest answer as of mid-2026 is that no major provider has committed to reading it in production, and the large-scale studies looking for a citation lift have not found one. Publishing the file will not get you cited.
What it will do, if you write it properly, is shape how coding agents, IDE assistants, and MCP integrations describe your product, a channel that’s vital enormously if developers are your buyers. It also forces an audit of your documentation that most teams never otherwise run.
These nine files are worth studying because each solves a specific problem well. All were live and publicly accessible when this article was researched earlier this year.
What llms.txt Is For
Before copying anything, be clear about what you are buying with the effort. The claims and the evidence have drifted apart, and knowing which is which prevents both overinvestment and reflexive dismissal.
What is established. Documentation platforms including Mintlify, Fern, GitBook, and Vercel generate the file automatically, which is why the adopter list reads like a developer-tools directory. Coding agents and IDE assistants do consume these files.
Perplexity has indicated it may retrieve the file as a hint about which pages matter.
What is not established. No major model provider- OpenAI, Google, Anthropic, Meta, or Mistral has publicly committed to using llms.txt in production retrieval.
Google has stated that AI Overviews require no special file or markup. Independent studies, including a 300,000-domain analysis by SE Ranking and a review of hundreds of thousands of citations by Trakkr, found no measurable citation advantage for sites that publish one.
Both things are true at once, and the resolution is straightforward. If AI coding agents interact with your product, llms.txt is a product surface and deserves proper editorial attention. If your buyers are marketers or operations leaders, it is cheap hygiene with an added side effect: writing it forces you to decide which pages are most relevant, an exercise most content teams benefit from regardless.
The adopter list is worth reading with that framing in mind. The companies most often cited as proof the format works Anthropic, Stripe, Cloudflare, Vercel, Supabase are almost all developer platforms shipping documentation, and most through docs tooling that generates the file automatically. That is a very different situation from a marketing site hoping the file will earn recommendations.
Growth-onomics treats it exactly that way in technical scoping: worth shipping, never sold as a visibility lever, and never prioritized above crawlability, content structure, or the third-party sources that inevitably carry citation weight.
Quick Comparison
| # | Company | Pattern Worth Copying | Best Suited To |
| 1 | Anthropic | Explicit scope and coverage declaration | Large multilingual docs |
| 2 | Stripe | Agent instructions that pre-empt model errors | Versioned SDKs and APIs |
| 3 | Supabase | Disciplined minimalism | Teams wanting low maintenance |
| 4 | Cloudflare | Sharded per-product index files | Broad product portfolios |
| 5 | Vercel | Taxonomy file for naming and deprecations | Products with renamed features |
| 6 | ElevenLabs | Teaching conventions instead of listing everything | Docs with predictable URL patterns |
| 7 | Zapier | Task routing with verification steps | Products agents operate directly |
| 8 | HubSpot | Prose positioning for non-developer buyers | Marketing-led SaaS |
| 9 | Pinecone | Company and entity facts in the file | Brands fighting entity confusion |
1. Anthropic
File: docs.anthropic.com/llms.txt
What They Did
The file opens by declaring what the documentation set is: a root URL, then a full list of available languages with page counts for each, marking which are included inline and which require visiting the site. Only after establishing that scope does it move into content.
Why It Works
Most files begin listing links without telling the reader what universe those links belong to. Stating that English covers 553 pages while each translated set covers 201 pages immediately tells an agent whether what it is reading is complete, partial, or a sample; the difference between confident summarizing and unnecessary hedging.
What to Copy
Open with scope. Say how large the documentation is, what languages exist, and what portion the file itself represents. It costs 4 lines and prevents a reader from assuming your index is exhaustive when it is not. If you serve multiple regions or languages, this is also the cheapest place to make that fact machine-readable.
2. Stripe
File: docs.stripe.com/llms.txt
What They Did
Before any links, the file addresses a specific failure mode: models pinning outdated package versions from training data. It instructs the reader to check the npm registry or PyPI for current versions and to install with the latest tag rather than a memorized number.
Why It Works
This is the most valuable pattern in the entire category, and almost nobody copies it. Stripe identified a recurring, costly error that AI assistants make about their product and corrected it at the source. Every link that follows uses a .md suffix, giving agents clean Markdown rather than rendered pages.
What to Copy
List the 3 things AI assistants most often get wrong about your product: a deprecated method, an outdated pricing tier, a superseded authentication flow, and correct them in the first paragraph. Your support team already knows what these are.
3. Supabase
File: supabase.com/llms.txt
What They Did
Almost nothing, deliberately. One H1, a single line pointing to the full documentation in one file, then a clean list of guide links with consistent naming and .md endpoints. No preamble, no marketing.
Why It Works
Restraint is a legitimate strategy when your documentation is already well-organized. The file inherits the structure of the docs rather than reinventing it, which means it does not drift out of date every time a guide is added, and it stays maintainable by a team that has better things to do.
What to Copy
If your documentation architecture is genuinely good, expose it rather than curate around it. The failure mode this avoids is the elaborate hand-written file that nobody updates after the first quarter, which is how most ambitious implementations end. A file that stays accurate for 2 years beats a richer one that goes stale in 3 months.
4. Cloudflare
File: developers.cloudflare.com/llms.txt
What They Did
The root file is an index of indexes. Products are grouped by category, and each entry links not to a documentation page but to that product’s own llms.txt file, with a one-line description of what the product does.
Why It Works
Cloudflare’s documentation is far too large to flatten into one useful file. Sharding lets an agent load only the branch it needs, keeping context windows manageable while still allowing full depth. The root file stays short enough to be read in full, which is the whole point of the format.
What to Copy
If you have more than a handful of distinct products, build a hierarchy rather than a list. One root file describing the portfolio, one file per product, each maintained by the team that owns it.
5. Vercel
File: vercel.com/docs/llms.txt
What They Did
Two pointers sit above the content: one to the full documentation text, and one to a product taxonomy in JSON containing canonical product names, aliases, and deprecations. The index that follows is nested, mirroring the documentation hierarchy.
Why It Works
The taxonomy file is the standout idea. Products get renamed, features get folded into other features, and models keep referring to the old names for years afterward because that is what their training data contains. A machine-readable list of current names, historical aliases, and deprecated terms addresses that directly.
What to Copy
If you have renamed anything in the last 3 years, publish the mapping. This is one of the few interventions that reliably improves how accurately AI assistants describe a product, and almost no company does it.
6. ElevenLabs
File: elevenlabs.io/docs/llms.txt
What They Did
An “Instructions for AI Agents” section teaches 3 conventions rather than enumerating every resource: append .md to any page URL for clean Markdown, append /llms.txt to any section URL for a scoped index, and connect to a documented MCP server endpoint for direct integration.
Why It Works
Teaching a pattern scales better than listing endpoints. An agent that learns the .md convention can retrieve any page, including ones published after the file was last updated. It also converts a static file into an entry point to a live interface rather than a snapshot.
What to Copy
If your documentation follows a predictable URL pattern, document the pattern. 3 lines of convention can replace 300 lines of index, and it never goes stale.
7. Zapier
File: docs.zapier.com/llms.txt
What They Did
The file includes a “For AI agents” section stating what the platform gives an agent, followed by an explicit routing rule: use one integration path for an AI client, another for a code project, a third for terminal work and numbered steps that end with a verification action before reporting success.
Why It Works
Zapier’s product is operated by agents, not just described to them, so the file is written as an operating manual rather than a catalog. The instruction not to stop at the index, and to confirm a verify step before claiming success, addresses the specific way agents fail: reporting completion prematurely.
What to Copy
If AI agents actually use your product rather than merely recommend it, write for the operator. Routing rules, prerequisites, and verification steps belong in the file, and the linking between a root map and a local index is a clean way to separate marketing scope from developer scope.
8. HubSpot
File: hubspot.com/llms.txt
What They Did
Rather than a documentation index, HubSpot publishes a prose description of the platform: what it is, what the underlying CRM does, how the products fit together, and which teams each serves, written in third person and readable end to end.
Why It Works
Most llms.txt guidance assumes an API product with developer documentation, which describes a minority of B2B SaaS. HubSpot’s file answers the question a general-purpose assistant is actually likely to have: what is this company and what does it sell, rather than an integration question nobody is asking about a marketing platform.
What to Copy
If your buyers are not developers, do not force your file into a documentation shape. Write a clear, factual, third-person description of the product and its main use cases, and link to the pages that answer the questions your sales team hears most.
9. Pinecone
File: pinecone.io/llms.txt
What They Did
Alongside the product summary, the file includes company-level facts: what the product does, the scale of its customer base, where the company is based, and who backs it, organized under plain “About” and “Why” headings.
Why It Works
This is entity information in a machine-readable place. For companies with a common name, a similarly named competitor, or a category that keeps getting confused with an adjacent one, stating the facts clearly gives systems something unambiguous to work from. It complements Organization schema rather than duplicating it.
What to Copy
Include a short factual block: what the company is, what category it competes in, headquarters, and one or two verifiable scale indicators. Keep it to facts you would be comfortable seeing quoted verbatim, because that is exactly what may happen.
What the Best Files Have in Common
Nine files, one shared shape. The strong ones do most of the following.
They open with a summary. One or two sentences in third person stating what the company does and what the file covers. An agent reading only the first paragraph should come away with an accurate description.
They write accurate descriptions. Every link gets a sentence explaining what is on the page and why an agent would want it. Descriptions carry specifics: actual tier names, actual constraints because they may be used as context without the URL ever being fetched.
They correct known errors. The highest-value lines in any of these files are the ones addressing what models get wrong: outdated versions, renamed products, deprecated methods.
They stay curated. 20 to 50 well-chosen links beat an auto-generated dump of every URL on the domain. If everything is important, the file communicates nothing.
They are maintained. A file describing last year’s product is worse than no file, because it looks authoritative while being wrong. Whoever owns the documentation should own this, with a review whenever something material ships.
They avoid the common failure modes. No walls of undescribed URLs, no marketing superlatives an agent cannot verify, no attempt to hide content from crawlers via this file; llms.txt is a hint, not an access control, and robots.txt remains the mechanism for that.
Practical sequencing is equally important. This is a 30-minute task that belongs after crawlability, content structure, and entity consistency are handled, not before.
In Growth-onomics engagements, it sits inside the content and structural refinement step for that reason: genuinely useful, cheap to ship, and never a substitute for the work that actually moves citations.
Conclusion
The interesting thing about llms.txt in 2026 is not whether it works. It is that the companies writing the best files are not doing it for visibility at all. Stripe is preventing bad code. Zapier is guiding agents through a workflow. Vercel is correcting naming confusion. Each solved a specific issue their users had, and the file is where the fix lives.
B2B SaaS teams can learn a lot from that. Ship the file, because it costs almost nothing and the exercise of writing it clarifies what your most critical pages are. Write it for a specific reader with a specific job rather than for an imagined ranking algorithm. Correct what models get wrong about you, describe what you actually sell in plain language, and keep it current.
Then set your expectations correctly. It will not lift your citation rate, and no credible reading of the current evidence suggests otherwise. The work that moves AI visibility remains what it has been: content that answers genuine questions, a site machines can crawl and parse, and third-party sources that describe you accurately.
To see where your AI visibility gap lies, and whether a file like this is anywhere near the top of your list, the Growth-onomics team can run the audit and tell you honestly.
FAQs
Does llms.txt improve AI search visibility?
There is no evidence that it does. As of mid-2026, no major model provider has publicly committed to using llms.txt in production retrieval, and Google has stated that AI Overviews require no special file. Independent analyses, including a 300,000-domain study by SE Ranking and a large citation review by Trakkr, found no measurable difference in citations between sites with and without the file. The credible reasons to publish one are different: AI coding agents and IDE assistants do consume it, and writing it forces a useful audit of what your most important pages are.
Where should the file be located?
At the root of the domain it describes, example.com/llms.txt, served as plain text or Markdown with a 200 status. Companies with substantial documentation commonly publish a second file on the docs subdomain, as Zapier does with a root map on the main domain and a local index on the docs subdomain. If you use a documentation platform such as Mintlify, Fern, GitBook, or Vercel, the file is likely generated already, so check before authoring one manually and make sure the two do not contradict each other.
What is the difference between llms.txt and llms-full.txt?
The first is an index: a short, curated set of links with one-line descriptions, meant to be read in full. The second is the concatenated content of those pages in a single file, intended for agents that want to ingest everything in one request. Full files get very large; some run to hundreds of thousands of tokens, which is why sharding by product, as Cloudflare does, is often the better pattern for broad portfolios. Publish the full version only if documentation is genuinely your most important content.
Should non-technical B2B SaaS companies publish one?
Yes, but not in documentation form. If your buyers are marketers, operations leaders, or finance teams, a developer-style index answers questions nobody is asking. Follow the HubSpot pattern instead: a clear third-person description of what the company does, how the products fit together, and links to the pages answering the questions your sales team hears most, pricing structure, security posture, integrations, and use cases. Treat it as a factual briefing document rather than an API map.
Who should own the file internally?
Whoever owns the documentation, in most cases, because that is where the file goes stale fastest and where the accuracy risk sits. Marketing should own the summary and the non-documentation links: pricing, security, use cases, since those are the pages buyers ask about. What does not work is treating it as an SEO deliverable owned by someone with no visibility into product changes, because a file describing last quarter’s product is worse than no file at all. Add it to your release checklist rather than a content calendar.
Can llms.txt stop AI crawlers from using my content?
No. It is a hint about which content is worth reading, not an access control mechanism, and nothing in the format prevents a crawler from ignoring it entirely. Access is governed by robots.txt directives and user-agent rules, plus server-level blocking if you need enforcement. Treat llms.txt as an invitation and robots.txt as the door. Confusing the two is a common and consequential mistake, because a team believing the file restricts access may leave content exposed that they intended to protect.
