We looked for the file everyone tells you to add. Almost nothing has ever fetched it.

llms.txt is a proposed standard: a Markdown file at the root of your site telling AI models what it contains and where the important pages are. It has been recommended in essentially every AI-visibility guide published since early 2025, and it appears as a line item in agency proposals at four figures.

In June 2026 Ahrefs checked whether anything reads it. They searched 137,210 sites for the file, threw out the ones answering with an empty page dressed up as a real file, checked that what came back was the right kind of document, and then went to their own records to see who had come to read it.

A little under one site in three had published one. And 97 in every 100 of those files were not opened by anybody, not once, in the whole of May 2026. Of the few that did get visits, almost all came from automated programs of every kind, and only one in five from anything recognizably to do with AI.

Their caveat is worth stating: the sites they sampled lean technical, people who follow this sort of thing, so one in three is the ceiling rather than the average across the web. And they counted who came to read, not who then did anything with it.

An independent tracker reaches the same conclusion from the other direction. Across more than three million sites, Originality.ai watched these files go from 4,088 in June 2025 to 36,120 in May 2026: nine times as many in a year, and still nobody reading them.

And the plainest statement available, from Google’s John Mueller in June 2025:

“FWIW no AI system currently uses llms.txt.”

No vendor has contradicted him since. Google’s own documentation for site owners goes further, in a sentence worth quoting whenever someone quotes a price at you for this:

“You don’t need to create new machine readable files, AI text files, or markup to appear in these features.”

Is the file harmful? No. It costs an hour, it breaks nothing, and if a major model starts consuming it tomorrow, you were early. Our objection is not to the file. It is to it being sold as work, and to it occupying the slot in a budget where something with evidence behind it should sit.

So here is the hour better spent, and this is the point of the article.

Take the same sixty minutes and do this instead: open your site in a browser with JavaScript disabled. If the page goes blank or loses its text, you have found a real problem, the kind that stops any program that does not run that code from seeing anything at all. That is gate one, item four on our list, and unlike this file it can be measured, it has consequences, and it is frequently broken.

If the page survives, spend the hour on the second thing: take your main service page and move the answer to the top. Whatever the page is for, say it in the first sentence, in the words a customer would use. Across 15.7 million AI citations, four in five reused paragraphs give the answer in the opening sentence.

Neither of those is exciting. Both are supported by measurement, and both take the same hour.

Why this pattern keeps happening. This file is attractive for the same reason that packages of structured data and keyword counting were attractive: it is a thing, and a thing can be delivered, invoiced and shown to a client. Rewriting a paragraph so it stands on its own is much harder to put on an invoice, and worth considerably more.

When someone quotes you for AI visibility work, the useful question is not what they will deliver. It is which of the things on the list has been measured, and by whom.