Adoption is not usage: reading the llms.txt studies properly

Three numbers that look contradictory

If you have followed the llms.txt evidence, you have seen adoption reported as 10.13%, 28%, and 51.8% within a few months. The last figure comes from our fixed panel snapshot dated 3 August 2026.

Those numbers are all defensible. They are answers to three different questions, and the disagreement is entirely in the sampling frame. Untangling it is worth doing once, because it also explains why “is anyone using llms.txt” cannot be answered by an adoption figure at all.

What each one actually measured

SE Ranking, roughly 300,000 domains. Their study reports llms.txt on 10.13% of the domains in their dataset, and finds no relationship between having the file and how often a domain is cited in major LLM answers. When they removed llms.txt as a variable from their predictive model for AI citations, model accuracy improved. Among the 50 most AI-cited domains, one had the file.

Ahrefs, 137,210 domains. Their study reports 28%, drawn from domains using Ahrefs Web Analytics that received traffic in May 2026. They state plainly that this population skews more technical and SEO-aware than the web at large and that the figure should be treated as an upper bound. They also note they did not check whether the files were well-formed against the specification.

Our panel, 219 selected hosts. The llmtxt.info adoption panel fetches and parses /llms.txt weekly across a fixed list deliberately weighted toward documentation platforms and developer tools. On 3 August 2026, 113 of 218 reachable hosts served a qualifying file, or 51.8%. This is a longitudinal measurement of a chosen cohort, not an estimate of the web.

Put side by side, the ordering is exactly what you would expect: a broad crawl of the general web, then a technically skewed analytics population, then a hand-picked list of the sector where the convention originated. The numbers rise with how selected the sample is. That is a consistent picture, not a contradiction.

The distinction that matters

Here is the part that adoption figures cannot reach.

SE Ranking measured publication and its correlation with citations. Ahrefs measured publication and then, separately, server logs, covering every request to /llms.txt across their population and classifying it by user agent.

That second measurement produces a number an adoption crawl cannot: of the roughly 38,000 files Ahrefs accepted, 97% received no requests in May 2026, from any client, bot or human. Ahrefs checked for HTTP 200, Markdown rather than HTML, and common error signals, but did not test full conformance with the proposal.

So the two large studies stack rather than compete:

• 10% to 28% of sites publish the file, depending on the population.
• Of the files published, about 3% are fetched by anything.
• Of the fetches, 19.5% come from named AI bots, and 1.1% from AI retrieval bots specifically.

Compounding those, the share of published files that are fetched by an AI search crawler in a given month is small enough that it is more honest to describe it as rare than to put a percentage on it.

None of this is an argument that the file is worthless, and we set out the case both ways on benefits and limitations. It is an argument that “X% of sites have adopted it” is not evidence of use, and should stop being quoted as if it were.

How to read the next one

More studies will follow. Four questions separate the useful ones.

What is the sampling frame? “Domains in our analytics product” and “domains in a general web crawl” are different universes. A figure without a frame is not a figure.

Publication, or consumption? Checking whether /llms.txt returns 200 is cheap and tells you about supply. Reading server logs tells you about demand. Almost every published number so far is a supply number.

Fetched, or read? Even log data has a ceiling problem, which Ahrefs state explicitly: a request proves bytes were retrieved, not that a model used them. Every consumption figure is an upper bound on real consumption.

Was validity checked? Ahrefs did not verify their files against the specification, and said so. A file that returns 200 but does not parse is counted as adoption in most methodologies. Our validator exists partly because that gap is invisible in aggregate statistics.

If you want to know what happens on your own site rather than in aggregate, the answer is in your access logs, filtered to /llms.txt. That is a smaller and far more decision-relevant dataset than any of the three above, and per no AI agent goes looking for your llms.txt, what it mostly measures is whether anything has been told your file exists.