Does llms.txt Work? Google Says No, and the Data Agrees

No. Google does not support llms.txt and has said large language models cannot use it to decide which site to surface. Ahrefs analyzed 137,000 domains and found 97% of llms.txt files received zero requests in May 2026. The file is harmless and mildly useful as a site map, but it is not a citation or ranking lever.

We should declare an interest before going further. JAS publishes a free llms.txt generator. What follows is not good for it, and we have corrected the claims on that page.

What Google actually said

Gary Illyes stated that Google does not support llms.txt and has no plans to. John Mueller compared it to the long-retired keywords meta tag, noted that no AI service had ever asked for the file, and described maintaining separate Markdown purely for bots as a bad idea. On Google's own podcast he said plainly that large language models cannot use the file to differentiate which website to surface.

Statements from search advocates are worth weighing carefully, since they describe Google's behavior rather than every AI system's. So the more interesting development is that somebody finally measured it.

What the measurement showed

Ahrefs looked at 137,000 domains and reported the following for May 2026.

Findings from the llms.txt data with the figure reported for each

Finding	Figure
llms.txt files receiving zero requests	97%
Domains that had published the file at all	28%
Share of requests from citation-producing fetchers	about 1%

That last row is the one that matters most. The crawlers whose visits actually turn into citations, OpenAI's search bot, PerplexityBot, and Claude's crawler, made up roughly one percent of requests. The file is not being read by the systems people are creating it for.

Why did so many people believe otherwise?

Because llms.txt was never validated. It was proposed, it sounded eminently sensible, and an entire layer of guides, checklists, agency services, and tools grew on top of it before anyone measured whether a single AI system read it. We were part of that layer.

The idea had everything a bad practice needs to spread quickly. It is easy to implement, it is free, it feels responsible, and there was no obvious way to tell it was doing nothing. Absence of results looks exactly like patience.

The pattern worth taking away

This matters beyond one file. AI visibility is young enough that plausible ideas travel faster than evidence, and the phrase "structured for AI" currently sells a great deal of work that nobody has tested.

So ask anyone selling you an AI visibility fix, including us, for the measurement. Not the reasoning. The measurement. If the answer is that it stands to reason, that is not a measurement, and it is exactly the position llms.txt occupied for over a year.

What is worth doing instead

What the evidence supports is duller than a new file, which is probably why it spread less.

1. Check that the citing crawlers can reach you. They are a small, named set, and whether your robots rules allow them is a real, checkable thing.
2. Make your answers extractable. A clear answer near the top of a page, self-contained enough to be lifted without losing meaning.
3. Be consistent everywhere you are described. Models reach for a single source when they cannot find corroboration.
4. Be specific. Concrete numbers, scope, and constraints are quotable in a way that positioning language is not.

Keep the llms.txt file if you have one. It costs nothing, it is a tidy map of your site, and some agents you point at it directly will read it. Just move it down your list, below the things that have been measured.

Why we built the tool anyway, and what we changed

We shipped an llms.txt generator because the file was widely recommended, trivial to implement, and appeared to be an emerging standard. Our page described it as something that helps you get understood and cited by AI answer engines.

That claim was not supportable and we have removed it. The page now states what the evidence shows, cites the Ahrefs figures, and points readers toward checking whether the crawlers that actually cite can reach them.

We are describing this in public for a reason. The most common failure in this field is not being wrong. It is being wrong quietly and letting a claim sit on a page long after the evidence has moved.

What llms.txt is genuinely useful for

Dismissing it entirely would be its own overcorrection. There are real, modest uses.

• A curated map for agents you point at it directly. If you are building or configuring something that reads your site, handing it a clean index is helpful.
• An internal clarity exercise. Writing a one-line summary of your site and its most important pages is genuinely useful, and many businesses find they cannot do it easily.
• Low-cost optionality. If adoption ever arrives, you already have the file.

What it is not is a lever. Keep expectations proportional to a file that 97% of the time nobody requests.

How to evaluate the next llms.txt

Something else will be proposed, will sound sensible, and will spread before anyone measures it. A few questions filter most of them.

1. Who consumes this, by name? If the answer is "AI systems" rather than named crawlers or products, that is a warning.
2. Has anyone measured uptake? Requests in server logs are the honest test.
3. What would disprove it? A practice that cannot fail cannot be evaluated.
4. What does it displace? The cost is rarely the implementation. It is the measured work not done instead.

None of this requires cynicism. It requires holding new practices to the same standard we would apply to a paid channel, which is a standard the SEO field has historically applied unevenly.