AI Visibility
The llms.txt census: how many SaaS sites actually serve one
Some of the traffic reading your pages is not a person. It is an agent, reading on a buyer's behalf, working out in seconds what your product is and whether to mention it. llms.txt is the file written for that reader: a plain text map at your domain root saying what you are and where your important pages live1. What nobody had a current answer to is how many companies actually serve one.
What we checked
On 26 August 2026 we requested /llms.txt once from each of 170 distinct domains. The population was a fixed 152-company research cohort (118 companies founded from 2023, 18 established controls, and 16 companies selling AI-search tools) plus the 19 well-known SaaS and devtools companies we had checked in July. One domain was unreachable.
A file counts as present when the request returns a 2xx with a body that is neither empty nor HTML. The HTML test matters more than it sounds. A single-page app that answers every path with its index shell would otherwise appear to serve a file at every address we asked for. Any other HTTP status counts as missing.
How many serve one
| Population | Domains | Serve llms.txt |
|---|---|---|
| All domains checked | 170 | 45.9% |
| Well-known SaaS and devtools companies | 19 | 84.2% |
| Companies selling AI-search tools | 16 | 68.8% |
| Established controls | 18 | 38.9% |
| Companies founded from 2023 | 118 | 38.1% |
Adoption is a seniority gradient rather than a young-company habit. The best-known names sit at 84%, up from 79% when we checked the same 19 companies in July: one had added a file since, and none had removed one. The companies selling AI-search tools are at 69%, which is about what you would expect from vendors who sell the problem. Everyone else, young startups and established controls alike, sits at 38%.
Inside the young cohort the split runs by sector. Developer infrastructure adopts at 43.8%, business software at 30.0%, healthcare at 25.0%. Companies whose own audience reads machine-readable files are likelier to write one.
The files are not stubs
We expected junk. The guess written down before the run was that most existing files would be auto-generated link dumps with no positioning summary. Thin was defined in advance, so the result could not be graded after the fact: under 500 bytes, or containing no prose sentence anywhere in the file.
Of the 78 files we found, one was thin.
| Slice | Files | Opens with a heading | Carries a summary | Median size |
|---|---|---|---|---|
| All files found | 78 | 94.9% | 85.9% | 7,628 bytes |
| Companies founded from 2023 | 45 | 95.6% | 84.4% | 6,506 bytes |
| Established controls | 7 | 85.7% | 71.4% | 13,258 bytes |
| Well-known SaaS and devtools companies | 16 | 93.8% | 93.8% | 6,435 bytes |
The median file carries 26 links and reads as a written document. 86% include the blockquote summary the convention asks for, and every file we found contained at least one prose sentence. Sizes run from 387 bytes to 1.5 MB. Whatever else is true of this convention, the companies that serve a file did not generate it and forget it.
Serving it is not the same as being read
A published file can still go unread. In August we watched an autonomous agent walk our own site carrying the intent of signing up. It moved through sixteen steps in 41 seconds. It followed the link to our API documentation and ignored the llms.txt link sitting immediately beside it, then reported that it could not find the onboarding documentation that the file contains.
That is a single observation and it settles nothing by itself. What it points at is the half of the job an adoption figure cannot measure. If you publish the file, make it discoverable: link it from the page head, declare it in a response header, and check that robots.txt is not quietly contradicting you.
What we got wrong
The census was pre-registered so the prediction could be scored rather than fitted afterwards, and two of its three parts were wrong. We expected adoption below 30%, and every slice measured came in above it. We expected the files to be thin, and one of 78 was. The part that held was the sector split, developer-facing companies ahead of general and enterprise software, and it held in direction only.
What this does and does not say
- It measures adoption and file quality on one day. It is not evidence that serving llms.txt causes AI visibility, and no figure here should be read that way.
- Serving the file correlates with traits (age, sector, how online a team is) that also drive how often AI answers name a company. A census cannot separate those.
- Presence means a 2xx, non-empty, non-HTML response at /llms.txt. It does not mean an engine fetched the file, read it, or acted on it.
- Two domains answered 403, so the true count could be higher by at most two. One domain was unreachable and is excluded from the reachable-share figures.
- The population is a fixed research cohort plus 19 well-known companies. It is not a random sample of software companies, and nothing here claims to generalise beyond it.
The checker fetches your llms.txt the way a crawler would and tells you what came back. If you do not have one, it generates a starter file you can publish. Free, no signup.
Run the llms.txt checkerRead next
Organic growth in the AI era: why AI visibility matters
Organic search still carries most of the traffic. A second discovery channel is growing fast alongside it, and it behaves differently.
Read →JournalOne answer in sixteen: how often AI search names a young company
11,552 AI search answers, one fixed YC cohort. Young companies are named in 6.4% of answers, YC's graduates in 36.7%, and 48 of 118 were never named once.
Read →