llms.txt is a proposed convention: a Markdown file at the root of a site that tells language models what the site contains and where the good material is. The idea is reasonable. Robots.txt worked, so a machine-readable orientation file for models sounds like the obvious next thing.
The question is not whether it is a good idea. It is whether anything reads it. On current evidence, no.
What the platforms say
Google's guidance on generative AI features puts llms.txt in its list of things site owners can ignore, and the wording is unusually flat: you do not need to create machine-readable files, AI text files, markup or Markdown to appear in Google Search including its generative capabilities, because Google Search itself does not use them. It adds that keeping one will neither harm nor help visibility, because Search ignores it.
Google's John Mueller has said separately that no AI system currently uses llms.txt. Beyond Google, no major provider has publicly committed to consuming the file in a production system.
What the logs say
Documentation tells you a stated position. Request logs tell you behaviour. In May 2026 Ahrefs published a study of 137,210 domains, of which around 38,000 carried a valid llms.txt. Of those, 97 percent recorded no requests for the file during the month.
That is a vendor study rather than independent research, and it measures what one crawl vendor could observe rather than the whole web. It is still the largest published sample anyone has put behind the question, and it points the same direction as the documentation.
Our own test, which has not reported yet
We publish an llms.txt on this site and log every request to it with user agent and date, precisely so this question can be answered from observation rather than opinion. The experiment is running. It has not produced a result we would stand behind yet, and we are not going to characterise a month of one low-traffic site's logs as evidence about the web.
When it reports, it reports either way. A negative result is the more likely one and it will be published as such.
What to do
If you already have one, leave it. It is a small file, it is accurate, and the cost of maintaining it is close to zero. If a convention does take hold later, you are already there.
If you do not have one, this is not the gap in your AI visibility. The gap is almost always content that reads like everyone else's, or pages that are not eligible to be surfaced in the first place.
What should stop is paying for llms.txt as a deliverable. It appears on AEO service scopes as though it were a lever. On the evidence available it is not a lever, it is a courtesy to a future that may not arrive.
What this means for an operator
Do not buy llms.txt as an AEO deliverable. If you want to know whether it does anything for you, log requests to the path yourself, which takes an afternoon and settles it.