Not invented examples. Every file below was fetched from a live domain, stored, and counted by a script, so any figure on this page can be checked against the file it came from.
An llms.txt file opens with an H1 naming the site, a blockquote summarising it, then H2 sections of links with a sentence about each. The 10 real files below were fetched on 13 August 2026, and they run from 2.1 KB to 206 KB.7 of them follow the format exactly. The template further down is the shortest file that conforms and is still worth serving.
Conforms to v2. Replace everything in brackets and delete nothing else. The comment lines are prose in this page rather than in the file, because the format has no comment syntax and a line you meant as a note is a line an agent reads as content.
# [Company name]
> [One or two sentences: what you do, who for, and what it costs to start.
> Put every figure next to the thing it describes. This is the passage most
> likely to be quoted on its own, so it has to be true without the rest.]
[Optional paragraph of detail. Anything an agent needs to interpret the links
below: what you call things, which product names are current, where the
authoritative version of a number lives.]
## Products
- [Product name](https://example.com/product): What it does, who it is for, and
the one thing that distinguishes it from the next link.
- [Second product](https://example.com/second): Same again. A sentence, not a
slogan.
## Pricing
- [Pricing](https://example.com/pricing): Entry price, what the tiers are, and
the date these figures were last checked.
## Facts worth quoting
- [Flat statement that is true on its own, with the figure inside it.]
- [Second statement. Written the way you would be happy to see it pasted into
an answer verbatim.]
## Optional
- [Company blog](https://example.com/blog): Secondary. Skip when context is tight.
## Notes for AI systems
Figures on this page were last checked on [date]. The live figure a customer is
charged is always the one shown at https://example.com/pricing.
Serve it at /llms.txt as text/plain or text/markdown, not as HTML. Under v2 you can also put one at a path, so /docs/llms.txt covers everything under /docs/, and where two files could apply an agent uses the more specific one. The specification page has the rest of the format.
Ordered smallest first. Bytes, sections and links were counted by a script against the stored body of each file, not read off the page.
| Domain | Size | Sections | Links | Explained | Follows v2 |
|---|---|---|---|---|---|
| anthropic.com | no file | — | — | — | — |
| sentry.io | 2.1 KB | 6 | 27 | 0 of 27 | departs |
| wordpress.com | 6.9 KB | 7 | 24 | 13 of 24 | yes |
| shopify.com | 12.2 KB | 12 | 74 | 71 of 74 | yes |
| zapier.com | 12.4 KB | 8 | 80 | 24 of 80 | yes |
| cloudflare.com | 16.5 KB | 13 | 88 | all | yes |
| github.com | 28.1 KB | 12 | 117 | all | yes |
| hubspot.com | 30.0 KB | 27 | 201 | all | departs |
| stripe.com | 63.5 KB | 30 | 288 | 285 of 288 | yes |
| vercel.com | 206 KB | 21 | 1830 | 196 of 1830 | departs |
| visibility100x.com | 42.0 KB | 11 | 115 | all | yes |
All fetched on 13 August 2026 with a browser user agent, following redirects. Explained counts list items carrying the colon and notes the proposal suggests. Our own file is in the table and is not excluded from anything said about the set.
One line on why each file is here, and one on how we read it. The second is our opinion and is written as one.
The apex domain returns 404. The real file is on the documentation subdomain, at docs.anthropic.com/llms.txt, where it runs to about 57 KB.
The most useful row on this page. A checker that only tries the apex will tell you a company has no llms.txt when it has a substantial one, and this is the normal arrangement rather than an oddity.
The smallest genuinely usable file in the set. Six sections and twenty seven links in about two kilobytes.
Departure: Opens with a blank line before the H1, and no link carries the colon and notes the proposal suggests. Every entry is a bare title and a URL.
Proof that a short file is a real file. What it gives up by skipping the notes is the one thing an agent most needs, which is a reason to pick one of the twenty seven links over the others.
The best summary blockquote in the set, because it does something almost nobody does: it dates itself, and it says what the document is for.
The line worth stealing is the one carrying a date inside the summary, so a reader reaching the file cold knows how old the claims in it are. It also spends a whole section distinguishing itself from wordpress.org, which is entity disambiguation done in the right place.
The closest thing here to a textbook business file: conforming, sectioned by what a reader wants to do, and almost every link explained.
If you are copying one shape, copy this one. Sections named for jobs rather than for the internal org chart, and a sentence per link written for somebody who does not already know the product names.
The only file in the set that treats itself as an addressable document, with stable section anchors it invites agents to deep link to.
A genuinely original idea: the summary tells an agent that anchors such as the MCP section are stable and safe to link. It also points at per surface files on its subdomains, which is the v2 subpath model used deliberately rather than by accident.
Every single link carries its notes, and the Optional section is used the way the convention intends.
The most disciplined file here. Eighty eight links and eighty eight explanations is not an accident, it is a generator with a required field. The Optional section holds the pointer to the full text variant, which is exactly what secondary means.
Opens with a How to use section that hands an agent the documentation APIs instead of making it scrape pages.
The most agent aware file in the set. It tells the reader which endpoint returns markdown and gives worked curl examples, which is the file understanding that its reader has a job to do rather than a page to render.
Twenty seven sections and two hundred links, all explained, from a company that also sells an AI visibility grader.
Departure: No summary blockquote. The opening paragraph does the same job but is plain text, so a parser looking for the blockquote the proposal describes finds nothing.
Thorough to a fault. At twenty seven sections it has stopped being an editorial selection and started being a catalogue, which is the failure mode the proposal warns about when it asks for a file small enough to sit in context.
Large, conforming, and organised by product line. The upper end of what still reads as an index.
Technically correct and near the limit of useful. Two hundred and eighty eight links means an agent has to search the file rather than read it, which the v2 proposal explicitly allows, so this is a defensible choice rather than a mistake.
The largest file here by an order of magnitude, and the clearest illustration of the index against corpus distinction.
Departure: Starts with three blockquote lines before any heading, so the first H1 appears on line five rather than first. Of 1,829 list links, only 196 carry notes.
This has stopped being an index. At 206 KB it is most of a context window before the agent has fetched anything, which inverts the point of the format. The pointers it opens with, to a full text variant and to a JSON cross link graph, are genuinely interesting ideas buried in a file too large to reach them in.
Ours, included so the standard we are holding other files to is one we are also measured against.
Conforming, generated from the same configuration the pricing pages render from so a price cannot go stale in it, and every link explained. It is also larger than we would like at 43 KB, which is the same criticism we make of Stripe above and it applies to us too.
7 of the 10 files here follow the v2 format exactly. The 3 that do not are worth looking at, because the departures are not careless. They are what happens when a real publishing pipeline meets a format written for a simpler case, and two of them are arguably improvements the specification has not caught up with.
The missing blockquote. HubSpot opens with a plain paragraph rather than the summary blockquote. A human reads it identically. A parser looking for the element the proposal describes finds nothing, and the summary is the passage most likely to be quoted on its own, so this is the departure with the clearest cost.
Content before the H1. Vercel puts three blockquote lines above its first heading, pointing at a full text variant, a product taxonomy and a cross link graph. The information is genuinely useful and there is nowhere in the format to put it, so it went above everything. The cost is that the one required element in the whole specification is no longer the first thing in the file.
Links without notes. Sentry lists twenty seven links and explains none of them. The file is valid, since the proposal marks the notes optional, and it is much less useful than it would be with one sentence per link. This is the most common weakness in the wild and the cheapest to fix.
Files that have become catalogues. HubSpot at twenty seven sections and Vercel at over eighteen hundred links have both crossed from index into corpus. The proposal asks for a file small enough to sit in context with the detail behind the links. Ours is 42.0 KB and the same criticism applies to us, which is why it is stated here rather than only about other people.
When we crawled 1,048 of the most visited sites in August 2026, 121 served a real llms.txt, which is 11.5%. A further 162 returned HTTP 200 for the file while serving an HTML page, because their server answers every unknown path with the site rather than a 404. A checker that trusts the status code would have reported adoption at 27%, more than double the truth.
So the test is not whether the request succeeds. It is whether what comes back is markdown beginning with an H1 rather than HTML beginning with a doctype. Two more checks worth doing by hand: confirm the content type is text rather than HTML, and try your documentation subdomain as well as your apex, because that is where most of the real files in our crawl turned out to live.
The full study publishes the method, the denominator and the CSV, so any of that can be recomputed rather than taken on trust.
Related reading: what llms.txt is and what v2 changed, whether the file does anything yet, and why Google says you do not need one for AI Overviews.
An H1 with the site name, a blockquote summarising what the site is, then H2 sections holding lists of links, each link followed by a colon and a sentence saying what is behind it. The template on this page is the shortest thing that conforms to the v2 specification and is still worth serving. Every real file below follows that shape to a greater or lesser degree.
You can copy the structure, and you should. What you cannot copy is the editorial work, which is the whole value: deciding which twenty of your pages matter and writing a sentence about each. A file copied wholesale with the nouns swapped lists somebody else's priorities under your domain name.
Every URL in the table on this page returns one, and each was fetched on 13 August 2026. Add /llms.txt to almost any developer documentation domain and you will find another. Note that the file often lives on the docs subdomain rather than the apex, which is why anthropic.com returns 404 while docs.anthropic.com serves a large one.
Long enough to name the pages that matter and short enough to sit in a context window beside the question being asked. In this set the smallest usable file is 2.1 KB and the largest is 206 KB, and the large one has stopped being an index. The proposal is explicit that detail belongs behind the links, fetched when needed.
The specification marks the notes after the colon optional, and in practice they are the most valuable part of the file. 4 of the 10 real files here explain every link. The ones that do not leave an agent choosing between two dozen titles with nothing to choose on.
llms.txt is an index of links. llms-full.txt inlines the whole text so nothing has to be fetched afterwards. Cloudflare links its full variant from the Optional section of its index, which is the tidiest handling of the pair in this set. For most business sites the full variant is unnecessary.
Fetch it and read the body rather than the status code. A large number of servers answer 200 for any unknown path by returning the homepage, so a status code check reports a file that is not there. Our llms.txt checker reads the response for that reason, and reports robots.txt and sitemap status alongside it.
The audit reads your llms.txt body rather than its status code, checks each AI crawler separately, and tells you whether the pages it points at survive without JavaScript. Free, no card.
Free account, no card. Paste your URL and get a real, scored report of your AI and search visibility.
Measuring rankings in