llms.txt made no measurable difference
A ‘zero find’ is still a find. This is the figure we would have looked for ourselves before spending an afternoon on the file.
No, not in our data. 202 of the 1,912 companies had a llms.txt file. They were used as a source in 43.6 per cent of cases, compared with 43.3 per cent for the 1,710 without it. The difference of 0.3 percentage points is far from statistically significant (p = 0.94). When it came to being mentioned by name, the figure even pointed in the wrong direction: 41.8 per cent with the file compared with 47.6 per cent without — nor is this statistically significant (p = 0.16).
The three figures that support the finding
Broken down by Denmark, Norway and Sweden
The countries point in different directions — yet another sign that we are seeing random variation rather than an effect.
| Country | With llms.txt | Without llms.txt |
|---|---|---|
| Denmark | 22.2 % (n=45) | 39.4 % (n=378) |
| Norway | 54.4 % (n=90) | 41.9 % (n=843) |
| Sweden | 43.3 % (n=67) | 48.7 % (n=489) |
What you can use the figure for
We sell a tool that, amongst other things, can create llms.txt files. It would have been easier for us not to measure this. But a dataset of which only the favourable parts are published is simply advertising disguised as data — and so it’s of no use to anyone.
The argument for creating the file is therefore not that it works today. It is that it takes half an hour, can do no harm, and that you’ll be ready if the search engines start using it. If you need to prioritise your time, the benefit lies elsewhere: in the fact that the page can be read and retrieved at all.
What the figure doesn’t reveal
We set out the reservations on this page, not in a footnote at the bottom.
- A null finding does not prove that the effect is zero — only that it is too small to be detected across 1,912 companies. A small effect may lie below our threshold.
- The file may also have an effect that we do not measure here: it is possible that it influences what the AI writes about you, without changing how often you are cited as a source.
- It is a correlation, not a cause. We check whether the company had any chance at all of being mentioned — not the size of the company or how well-known the brand already is. A large, well-known company often has both better technology and a better reputation.
- The citations are measured using Perplexity, which is the only search engine that displays actual sources. Other search engines may behave differently.
How the figure was arrived at
1912 Companies from 15 industry panels in Denmark, Norway and Sweden, as measured between 2026-07-26 and 2026-08-24. The technical score is calculated deterministically without any AI assessment. The outcome is measured based on correct AI responses to unbranded questions.
The test used is the Mann-Whitney U test, with corrections for ties and continuity; this has been cross-checked against a permutation test with 10000 rounds and a fixed starting value, so that the result can be replicated.
Mann-Whitney U: z = 0.08, p = 0,94, rank-biserial effect size 0.003 — effectively zero. The finding is robust to both citation and mention outcomes.
Make the most of the time when it’s on the move
The free test reveals what is actually holding your business back from AI-generated responses.
Assess my business for free