Does llms.txt actually do anything?

What is settled and what is not

The strongest evidence is a denial rather than a silence. In June 2025 Google’s John Mueller wrote that no AI system currently uses llms.txt, and likened it to the keywords meta tag — a comparison worth sitting with, because that tag was widely adopted, trivially cheap, and did nothing.

That is one company speaking for its own systems at one point in time, and it is not the same as proof that nothing anywhere reads the file. But it is the clearest statement any vendor has made, and it points the opposite way to most of the advice being written.

The convention itself is well specified and trivially cheap to satisfy, which is why the honest recommendation is still to publish one. A generated index costs nothing to maintain and cannot hurt. A three-week project to hand-craft one spends real time against an unmeasured return, and there are measured returns available in the same budget.

The counter-argument, stated fairly

Pages published during 2026 argue the Mueller framing is now out of date, and it would be dishonest to leave that out. The claims we have seen fall into two groups, and they are not equally strong.

Crawler-log evidence — that agents request the file — is the better kind, because a request either happened or it did not. Outcome evidence is where the argument thins: figures like "sites with llms.txt earn twice the revenue" compare sites that differ in a dozen other ways, and the direction of causation is exactly what is at issue. A store organised enough to publish one is organised in other ways too.

Our position is narrow. A crawler fetching a file is not the same as a model using it, and until a vendor documents the second, adding one is a cheap bet rather than a technique.

What the adoption number does not mean

High adoption is regularly quoted as evidence that the file matters. It is mostly evidence that a platform switched it on. Of the stores we found serving one, the large majority served a templated file generated for them, near-identical from brand to brand.

So an adoption statistic in this category is measuring a vendor default, not an industry decision. Any comparison that treats it as a maturity signal is reading a platform release note.

The failure that matters more

Auditing readiness through this file is where it actively misleads. We requested it alongside the agent-commerce discovery path on the same 147 hosts, and the results diverge: twelve stores that accept a full agent checkout publish a file that never mentions it, seven of them serving no file at all.

A store can therefore be entirely transactable by an agent and score zero on an llms.txt audit. If the question is what a machine can do with your store, the file is the wrong place to look, and the discovery path answers it directly.

A measurement trap worth knowing

If you are counting adoption yourself, check the content-type rather than the status code. One host in our sample returns HTTP 200 at that path and serves two megabytes of HTML — its catch-all page. Any survey counting 200s records it as an adopter, and every published adoption figure derived that way is inflated by some unknown amount.