We asked whether llms.txt works. It is worth 1.37 points, and a quarter of that is us counting the file.

RBy Ronnie Huss, founder of SearchScore

In July we registered the question on SearchScore Labs before running it: do sites publishing an llms.txt achieve a materially higher AI visibility score than sites without one? The registration committed us to publishing the result either way, including a null, and noted that a positive result would only motivate a controlled follow-up rather than prove anything on its own.

Sites with an llms.txt average 49.14 against 43.78 for sites without, a gap of 5.36 points with a 95% confidence interval of 5.17 to 5.55. On 29,692 sites against 73,182, that interval is narrow and the direction is not in doubt.

Whether 5.36 counts as materially higher is the interesting argument, and there are two reasons to think it does not.

A quarter of the gap is the scorer noticing the file

The llms.txt check is 8 of the 105 points in AI citability, which carries a 0.18 weight, so publishing the file moves the score by 1.37 points on its own. That is 26% of the measured 5.36. For Organisation schema, the equivalent figure is 8%.

It does not tell good sites from bad ones

A useful marker should be rare among weak sites and common among strong ones. Organisation schema does that almost perfectly. llms.txt does not.

Look at the bottom row. Among sites scoring under 50, 22% publish an llms.txt and only 5.6% publish Organisation schema. A fifth of the least ready websites on the index have already done the thing the industry spent the year recommending, and it did not make them ready.

That is what a weak signal looks like. The file is trivially cheap, it is increasingly emitted by templates and site generators without anyone deciding to publish it, and a signal that costs nothing and requires no judgement stops carrying information about judgement.

The effect is flat, which is its own tell

When we ran the same split for Organisation schema, the sites carrying it led by 39 points on platform optimisation and 29 on topical authority, categories that have nothing to do with schema. That spread is what made it readable as a marker of care. llms.txt produces no such pattern.

Every llms.txt gap sits between 2 and 8 points. Nothing stands out, because nothing much is being revealed. Sites with an llms.txt are slightly better at everything, in the way that any site whose owner has read one article about AI search is slightly better at everything.

Two different jobs, and only one of them is search

Our data measures whether a site publishes the file. It says nothing about whether anything reads it, and on that the public record now points in two directions at once, which turns out to be the useful part.

We have since run that audit ourselves and written up what the Agentic Browsing category actually contains: three of its six checks are WebMCP and none of them can currently cost a site a point.

Google Search says an llms.txt is not needed. Its AI optimisation guidance of 15 May 2026 tells site owners so plainly, for AI Overviews, AI Mode and every other generative AI Search feature. No major provider has confirmed that its assistant parses the file at inference time.

Google Chrome says something else. Lighthouse 13.3 added an Agentic Browsing category, and one of its audits checks for a machine-readable summary at the domain root. Absence is marked Not Applicable rather than failed, because publishing one is optional, and a server error when fetching it is flagged. It is explicitly not a ranking signal. Adoption is running ahead of both positions: more than 844,000 sites publish one and Yoast has built support into its WordPress plugin.

Those two positions are not in conflict. They are about different jobs. Answering a question is retrieval, and there the file is unused. Navigating a site to complete a task is agentic browsing, and there a map of what the site is and which pages matter is worth having. The file is becoming a standard for the second job while remaining irrelevant to the first.

That is how to read the 5.36 points above. The gap is real and measured, and it is measured on a score built for AI search visibility, which is the job the file does not do. Nothing here says llms.txt is pointless. It says the case for it is agent readiness, and our number is not evidence for that case because it is not measuring it.

What this cannot tell you

This is correlational, as the registration said it would be. It cannot tell you what happens when a given site adds an llms.txt tomorrow, and 1.37 points is the only part we can promise. It also cannot see whether any AI engine reads the file, because the score measures publication rather than consumption. That is the more interesting question and this design does not touch it.

The United Kingdom

Among the 5,482 GB domains in the panel, 33.5% publish an llms.txt against 27.3% across the weighted panel, and the gap there is 4.04 points with an interval of 3.34 to 4.74. Slightly more common, slightly less valuable, which is the same pattern we found for schema and the normal life cycle of a signal that spreads.

What to do with this

Publish one, and do it for the right reason. It takes ten minutes, it is a genuine 1.37 points, and 91% of the sites that cleared the AI-Ready bar have one. But the case that is strengthening is agent readiness rather than search: Lighthouse now audits for it, Yoast ships support, and more than 844,000 sites publish one. Our llms.txt guide covers the format.

What the data here rules out is treating it as an AI search strategy. A fifth of the worst sites on the index already have one, and the engines answering questions are not reading it.

If you are choosing where to spend an afternoon, the schema result is the better buy, and both of them together are worth less than the thing neither signal reaches: content an engine can lift a clean, checkable sentence out of. That is where the field is still open, and it is the only part of this that cannot be done by adding a file. The free audit will show you where your own site stands.

Method

Source is the SearchScore SAVI index panel, wave 1, dated 1 August 2026: 102,873 domains audited on scorer geo-145-2026-07 with method quick+geoOnly, weighted to the population. Cells are 29,692 sites publishing an llms.txt and 73,182 without. Confidence intervals are 95%, computed on weighted means using Kish effective sample size.

That 1.37 points is what the check itself is worth, and it comes from the scorer rather than an estimate: the llms.txt check is 8 of the 105 points available in the AI citability category, and that category carries a 0.18 weight in the overall score. Detection is presence of a fetchable /llms.txt, not its contents or quality, which is a real limitation: a file listing nothing useful counts the same as a good one.

To confirm the extract sits on the same basis as the published index, we recomputed the figures the SAVI Q3 report states. Weighted mean score reproduces at 45.24 against 45.24 published, and llms.txt prevalence at 27.3% against 27.6%. Our measurement standards set out what these numbers are fit for.

Pre-registered 8 July 2026 as EXP-0005, before the query was run, with a commitment to publish regardless of direction. The hypothesis is confirmed in direction. Whether a 5.36-point gap, a quarter of which is the score paying out for the file, meets the registration’s word materially is a judgement we are putting in front of the reader rather than settling quietly in our own favour. Our reading is that it does not.