Does llms.txt Work? A 90-Day Watch on 3 Sites, Zero Fetches So Far

Does llms.txt actually work? llms.txt is a proposal (llmstxt.org) — not an adopted standard, and no major engine has committed to reading it — so we stopped arguing and instrumented: the file is served on our 3 production sites, and every request to /llms.txt is counted by user agent inside a 90-day observation window. This page is the running log. The count of fetches by documented AI engine crawlers currently stands at 0, and we publish the protocol alongside the number because a result with its method attached beats a claim without one.

What did we deploy, and where?

We deployed the llms.txt proposal's file (llmstxt.org, 2024) — not an adopted standard, and no major engine has committed to reading it — on all 3 of our production builds: an insurance lead-gen brand, an auto-finance authority rebuild, and our own magnitude network [our data]. Each file follows the proposal's format: H1 site name, blockquote summary, H2 sections of annotated links to the site's most important pages.

The files are generated from each build's existing page metadata, so the marginal cost was minutes, not hours. Each is served at the domain root as plain text and was verified fetchable with a curl request. Setup mechanics, if you want to replicate this, are in our llms.txt how-to.

The observation window runs 90 days from the oldest build's mid-July 2026 launch, closing in October 2026. We change nothing about the files mid-window — the point is to hold the surface constant and watch who comes.

What do the server logs show so far?

The headline so far: 0 requests for /llms.txt from any documented AI engine crawler, on any of the 3 sites [our data]. Not few — zero. The per-site picture, current through August 19, 2026:

The context that makes the zero meaningful: this is not a case of crawlers ignoring our sites. The same logs show documented AI crawlers — GPTBot, ClaudeBot, PerplexityBot — fetching HTML pages on the builds [our data]. The engines are present, active, and indexing — and none of them has asked for the file that was supposedly written for them.

The requests /llms.txt does receive come from SEO audit tools and unidentified user agents [our data]. Which is its own small finding: today, llms.txt's actual audience is the SEO tooling industry checking whether you have one.

Why publish the log before the window closes?

Because committing to the protocol in public is the honest version of this experiment. The sites are named by role, the window is declared, and the counting method — filter requests by path, group by user agent, accept only documented crawler names — is stated before the result is complete. This page updates as the window accrues, and the day-90 count gets recorded here whether it stays at 0 or not.

We also added permanence: requests to /llms.txt are logged by user agent on all 3 builds indefinitely, so if any engine ever starts fetching the file we will know from our own logs rather than from a vendor's announcement. That is the correct posture for an unadopted proposal — hold the option, watch the data, spend nothing.

How does the zero-so-far square with the public debate?

A zero is consistent with both of the loudest public positions — which is why the debate persists. Google's John Mueller compared llms.txt to the keywords meta tag (Search Engine Journal): a self-declared signal engines have little reason to trust. Our logs support the practical half of that skepticism — nothing has consumed the file.

Search Engine Land's counterpoint argues the comparison undersells a file that is cheap, spam-neutral so far, and potentially useful if adoption arrives. Our experience supports that too: the deployment cost us minutes and harmed nothing. Both positions survive contact with our data because they are claims about different things — one about present consumption, one about option value. The present-consumption answer, on our sites, is zero to date.

What our log does contradict is the sales pitch. Content promising that llms.txt improves AI visibility — and tools charging for llms.txt generation on that premise — are asserting a mechanism we have never observed firing: the engines' crawlers are documented, they visit our HTML constantly, and they do not request the file. See the complete AI crawler list for what they do request.

What is the generalizable rule?

The rule we took from this: deploy cheap surfaces freely, but only spend real budget on behavior you can see in your own logs. llms.txt passes the first test — minutes of setup, zero maintenance — and has yet to register on the second, with 0 engine fetches across 3 sites so far [our data]. So it earns a place in the build script and no place in the strategy.

The same log-first discipline is how we evaluate every tactic in the complete GEO guide, and it points at where AI-crawler attention actually goes: your HTML pages, your robots.txt, and your server capacity — decisions worth real thought, because they involve crawlers that demonstrably show up. If a tactic's entire evidence base is other people's blog posts, treat it the way we now treat llms.txt: as an option to hold for free, never as a result to pay for.