The file everyone added that nothing reads

Adding llms.txt costs an hour and carries no risk, which is exactly why it spread so fast and why nobody stopped to check whether it works. Two years in, the honest position is that it is a reasonable thing to have and a bad thing to expect results from. Here is the distinction, and what actually moves the needle instead.

What each file actually does

The middle row is the point of this article. "No confirmed consumption" is not the same as "provably ignored" — it means no operator has documented using it and no independent test has isolated an effect, which is the state of the evidence rather than a verdict.

What it was meant to solve

The proposal was sensible. A model reading your site has a limited context window and a lot of navigation, boilerplate and markup to wade through. llms.txt offers a curated map in plain Markdown: here are the pages that matter, described accurately, without the furniture.

The analogy everyone reaches for is robots.txt, and it is the wrong one. robots.txt works because it is a permission file and crawlers have commercial and legal reasons to respect permissions. llms.txt is a recommendation file, and nothing obliges a model to prefer your summary of your site over its own reading of it.

Why the evidence stays thin

To show that llms.txt works you would need a model to be measurably more likely to cite a site that has one, holding content quality constant. Nobody has produced that, and the reason is structural: you cannot see inside the retrieval step, so any measured difference could as easily come from the content the file points at as from the file.

The adoption pattern makes it worse. The sites that added llms.txt early skew towards the technically careful — the same sites with clean structure, good schema and solid content. Whatever advantage they show is thoroughly confounded, and it always will be.

So should you have one?

Yes, with correctly sized expectations. It takes an hour, it cannot hurt, and if consumption becomes real you are already there. Treat it as a cheap option on a future standard, not as an SEO tactic with an expected return. Ours has been in place since 2026 and we would not remove it.

What deserves the attention that llms.txt gets is the boring adjacent work: complete structured data, facts stated consistently across your site and off it, and content shaped so a single paragraph answers a single question. That is the mechanism models demonstrably use, and it is all in our GEO guide.

The pattern worth recognising

This is the third or fourth time a zero-cost file has been sold as an AI-visibility lever, and it will not be the last. The test to apply is simple: can anyone show the mechanism, or only the correlation? If a tactic has no documented consumer and no isolated result, it belongs in the "cheap, harmless, unproven" column — which is a real column and a fine place to be.

The uncomfortable version is that the work which does move AI visibility is slow, editorial and hard to sell as a tip. How long it takes to get cited at all sets a realistic expectation, and what crawlers actually send back sets a realistic ceiling on what citation is worth.

Sources

Checked September 2026. Note that the primary claim here is the absence of evidence, which no single source can prove — these document the state of the discussion.

• WikiBusines — llms.txt in 2026: hype, data, and what to do instead ↗

  The sceptical case: broad adoption without demonstrated consumption by major models, and the recommendation to invest in structured data instead.

• Digital Strategy Force — Does your site need llms.txt to get cited by AI search in 2026? ↗

  Reviews what would be required to demonstrate an effect and why the confounding described in the second section makes it difficult.

• New Target — llms.txt and the fight over AI access to your website content ↗

  Background on the proposal itself and how it differs from robots.txt — a recommendation rather than a permission.