We ran llms.txt for 90 days on nine sites. Nothing moved.

llms.txt is a proposed file at the root of a domain that lists a site's important pages in markdown, on the theory that a language model will read it and understand the site better. It costs an afternoon to write. It is currently being sold as an AI visibility service.

We had no strong prior against it. A cheap convention that works would be excellent news. So we ran it properly.

1. 01Nine B2B SaaS sites across six categories, all with an existing tracked prompt panel.
2. 02Baseline: 30 days of runs before any file was published.
3. 03Intervention: a complete llms.txt at the root, plus llms-full.txt where the site was small enough for it to be meaningful.
4. 04Observation: 90 days, three runs per prompt per engine per week, nothing else changed on any of the nine sites.

The last condition is the one most vendor case studies skip. If you publish a file, ship twelve pages and land a podcast in the same quarter, you have not tested the file.

Weekly naming rates moved, because they always move. None of the movement was distinguishable from the variance we had already recorded during the baseline period on the same sites. Two sites finished slightly up, three slightly down, four flat.

We also checked the simpler question underneath it: was the file fetched at all? Server logs across the nine sites showed retrieval of /llms.txt by identifiable AI crawler user agents at a rate low enough to be indistinguishable from general root-path probing 1.

It sells because it is legible. It is a file you can look at, a deliverable that can be shown in a status call, and it belongs to a genre of work that felt effective for twenty years. robots.txt mattered. sitemap.xml mattered. The shape is familiar, so the conclusion feels earned before anyone checks it.

What is different is who is on the other end. A crawler follows a documented contract. A model chooses what to cite from a retrieval layer that mostly returns pages other people wrote about you. There is no line in that pipeline where a file you host tells the model what to think.

The same quarter, on a subset of the same sites, we worked on third-party surfaces the panels showed the engines actually reading. That work is slower, less legible in a status call, and produced movement outside baseline variance. We will publish that panel with its own nulls attached.