Does llms.txt do anything? Three studies say mostly no

Two pieces on this site have already asked who publishes an llms.txt file. Our adoption study found 14 of 60 major sites serve one, concentrated among SEO vendors and developer platforms. Our tracker comparison showed five trackers disagreeing on the adoption rate because none measured the same population. Neither piece asked the harder question: once a file exists on a server, does anything read it, and does having one change what an AI system does with the site. Three separate checks published in 2026 answer that question, and the answer is mostly no.

The usage gap Ahrefs found underneath its own adoption number#

Ahrefs’ analysis of 137,210 domains running its Web Analytics or Bot Analytics product found 28% serve a valid llms.txt file, roughly 38,360 domains, a figure already used in our tracker comparison. The same report goes further, checking server logs for whether anything actually requested those files. In May 2026, 97% of the roughly 38,360 valid files received zero requests. Ahrefs is explicit that even the remaining 3% is a ceiling on real usage, not a floor: a fetch does not mean the content was read or acted on, since “many bots may have fetched the llms.txt file without ever acting on what’s inside.” A file that nothing requests cannot be doing anything for the site that published it, which makes the adoption question and the usage question genuinely different findings, not two ways of stating the same one.

Two small tests, two honest limitations#

Search Engine Land tracked 10 sites across finance, B2B SaaS, ecommerce, insurance and pet care, measuring AI traffic and crawl frequency for 90 days before and after adding llms.txt. Eight sites showed no measurable change. One declined 19.7%, a drop the piece attributes to unrelated causes. Two grew, 25% and 12.5%, but the article’s own analysis credits that growth to new functional content the sites published alongside the file, not to the file itself: “llms.txt documented those efforts. It didn’t drive them.” A 10-site sample cannot settle the question for the whole web, and the piece does not claim it does.

The one positive figure in this space comes from an independent researcher who tracked citations across four of his own properties over 90 days and measured an 11% overall lift after publishing llms.txt, concentrated almost entirely on Claude and Perplexity with no movement on ChatGPT Search or Google AI Overviews. He then undercut his own headline number in the same piece: “my +11% observation came from a window in which several changes shipped together, so the design does not support attributing it cleanly to llms.txt.” That sentence is the reason this figure belongs in this piece at all. A researcher willing to name the hole in his own methodology, on a corpus he controls, is more useful than a cleaner-looking number from someone who did not check.

Why we care#

Adoption and impact are two different claims, and only one of them has evidence behind it right now. Fourteen sites in sixty publishing a file, per our adoption study, tells you nothing about whether the file did anything once published. The usage and traffic data above is the actual answer to that second question, and it is a weak one: near-total non-use in the large sample, no effect in most of the small traffic test, and a positive citation number its own author will not stand behind without a caveat.

If a vendor pitches llms.txt as a traffic or citation lever, ask which of these three studies they are relying on. None of the three supports a clean causal claim. The Ahrefs number shows the file mostly sits unread. The Search Engine Land test shows no effect in 8 of 10 real sites. The one number that moved comes from a researcher who told you himself not to trust it as isolated evidence, which is the standard every figure in AI search coverage should be held to before it goes in a deck.