llms.txt adoption: every study compared

People quote "10%", "5.6%", "28%" and "97%" for llms.txt as if they measured the same thing. They don't. Below is every llms.txt study we could find, read on its publisher's own page on Oct 9, 2026, sorted by what it measured: whether sites have the file, whether AI crawlers request it, and whether having it gets a site cited more. For what Google and the AI companies themselves say about the file, see Does llms.txt work?.

1. How many sites have an llms.txt file

Scans of top-sites lists first, newest first, then other samples. "Hit" is what each study counted as a real file. Figures as published, except where marked.

The scans of top-sites lists done in 2026 (Rankability, SEOmator, the DEV Community post, lintlab, Burridge, Humphrey, Thunderbit and ours) land between 5.1% and 11.4% of all domains checked. The two figures far above that, Ahrefs' 28% and Search Atlas' 32.2%, come from sites that use SEO software, which are much more likely than the average site to follow SEO advice. The Web Almanac's 2.1% covers the whole web in July 2025, when the file was newer.

2. Do AI crawlers ask for the file?

Server-log studies: who actually requested /llms.txt.

The pattern holds from one site to Ahrefs' 38,000: the same AI crawlers that fetch robots.txt thousands of times rarely or never fetch llms.txt. Most requests for the file come from SEO audit tools, scanners like the ones in section 1, and site owners checking their own file. The newest data shows some crawlers starting to look (Anthropic's and Google's in WISLR's follow-up, Meta's in EZY's), while OpenAI's crawlers fetched it 7 times in 12 weeks across EZY's 83 sites and not at all in WISLR's follow-up.

3. Does having one get you cited more?

Three studies, three methods (correlation and machine learning, a three-step control for domain authority, matched pairs), one answer: no measurable lift. All three are by companies that sell SEO or AI visibility software, and none is peer-reviewed.

Why the adoption numbers disagree

1. The denominator. A list of "top domains" is full of names that aren't websites: CDN hosts, ad servers, API endpoints. In our Tranco scan, 236 of the 1,000 domains didn't even resolve in DNS. Divide by all 1,000 and you get 10.2%; divide by the 679 that answered and it's 15.0%. SEOmator (8.6% vs 11.4%) and the DEV Community post (11.4% vs 15%) show the same gap. Rankability keeps every domain in the denominator, including 474 of its top 1,000 it couldn't decide either way; its 93 hits are 9.3% of all 1,000, but 17.7% of the 526 it could decide (our arithmetic).

2. Fake 200s. Many sites answer any unknown address with their normal HTML page and a "200 OK" status. Count every 200 as a file and adoption can more than double: SEOmator saw 20,171 domains answer 200 but only 8,598 serve a real file; Thunderbit 1,606 vs 586; Humphrey 1,050 vs 737; the DEV Community post 246 vs 114; we saw 120 HTML or empty 200s next to 102 real files. lintlab only checked the content type, not the body, so its 6.6% is not filtered the same way.

3. Files nobody wrote by hand. Platforms and plugins now create llms.txt for their users. Shopify alone made 18.2% of the files SEOmator found, and WordPress SEO plugins another 9.6%. Plugins made a third of the files in the hotel study and 12% in Humphrey's. That pushes adoption up for whichever list holds more shops and WordPress sites, and says nothing about whether anyone chose to publish the file.

4. The list. Tranco blends several rankings and includes infrastructure domains; the Majestic Million ranks domains by links; the Chrome UX Report (used by HTTP Archive and lintlab) counts sites real Chrome users visit; Cloudflare Radar ranks by DNS traffic. Ahrefs and Search Atlas measured their own customers. Each list holds a different mix of software companies (which adopt most: about one in five in our sample) and other sites.

5. The date. Adoption has grown fast. Burridge measured 1.04% of the top 10,000 in July 2025 and 5.61% in June 2026; Originality.ai's count rose 8.8x in the same year; Chris Green found 15 files in the whole Majestic Million in February 2025. A study from 2025 and one from late 2026 are measuring different webs. Rankability warns against reading its June and September 2026 figures as a trend, because its scanner and list changed in between.

Our scan

On Oct 9, 2026 we requested /llms.txt from every domain in the Tranco top 1,000 (list PY35J), trying www. when the bare domain failed. A hit had to return a 200 with a non-HTML, non-empty body that wasn't an error page, a JSON error, or a catch-all reply (we also requested a made-up filename to catch hosts that answer everything), and the final address after redirects still had to be a .txt file. Result: 102 of 1,000 domains (10.2%), or 15.0% of the 679 that answered over HTTP. Among the domains we labelled by type, software and developer companies were far ahead of news sites, social networks and infrastructure hosts. One network and one run, so sites that block bots or our location count as misses. We're extending the scan to the top 10,000 and will add the result here.

What this means for you

If you want to publish an llms.txt, it costs nothing and does no harm. Just don't expect it to change how AI assistants see you: the log studies show the main AI crawlers barely ask for it, and the citation studies show no lift. Letting the crawlers in (robots.txt and firewall rules) and ranking in ordinary search matter more, as our llms.txt guide explains. If a tool you pay for scores you down for a missing llms.txt, compare what each tracker actually measures in our price index.

How we built this page

We searched for every published study of llms.txt adoption, requests or citations, then read each one on its publisher's own page on Oct 9, 2026, checking every figure here against the page's text. Dates are the publisher's. Not included: roundups that only repeat other studies' numbers, studies of hand-picked groups of a few hundred sites or fewer, and studies whose figures we couldn't read on the page. If we've missed a study, or a figure here has changed at its source, email hello@aivisibilitynerd.com.