Does anyone fetch llms.txt? What this site recorded in 90 days

What is in the 249-request count?

The exact path matters. /llms.txt contributed 249 records; /llms-full.txt contributed another 27 and is excluded from that headline. The latter name can be probed whether or not a useful document exists there, so neither path’s count establishes a successful response.

The method split keeps five HEAD requests separate because HEAD responses have no document body. Even a GET observation occurs before this dataset can establish the returned status or useful content. Treating all requests as file readers would overstate the evidence.

Which names did clients claim?

The named labels include bingbot, Googlebot, GPTBot, OAI-SearchBot and ChatGPT-User. They come from User-Agent classification. They are not authenticated provider identities, and some records may be diagnostic or synthetic checks.

The small named totals are useful for choosing a verification sample. They cannot establish that an AI provider adopted llms.txt, or that a search engine used its linked content. There is no session-to-answer linkage in this observer.

Why a request is a weaker result than a reader

Several links are missing: independently identifying the client, confirming a useful response, establishing that its contents were consumed, and connecting that consumption to a later answer or referral. This study only observes the request stage. A client could request a missing path, receive a cached error or fetch a file without using it.

Absence of an observed provider label would also be ambiguous. Retrieval could use a different identifier, a search index, a cache or an intermediary, and the best-effort observer could lose a delivery. The broader snapshot describes the selected routes and capture limits.

The window was not a randomized deployment experiment. Other content, links, publication and search conditions can change during ninety days. Comparing traffic before and after an llms.txt edit would need a design that addresses those changes; these counts do not isolate the file’s effect.

What I would maintain, and what I would measure next

I would keep a published file accurate, concise and connected to canonical public resources. Broken links and stale claims make it a poor guide for any client. That maintenance decision can stand on its small cost and editorial purpose without claiming a measured citation benefit.

For a stronger test, define the desired outcome first: delivered file, subsequently fetched linked document, answer citation or referral. Capture the fields needed to join those stages within the site’s privacy constraints, verify client identity where possible, and retain a comparison. The server-log guide is the starting point for the attribution step.

Until that evidence exists, the result is precise and limited: 249 recorded requests to this exact path, including five HEAD requests, with reading and downstream effects unknown.

Inspect the path and method totals

The CSV preserves exact path, method, observation class, claimed label, verification state, count and first/last event times. Sum only /llms.txt for 249; include /llms-full.txt to obtain 276 across the two names. Blank claimed labels remain unmatched.

The snapshot bundle includes the extraction SQL, daily denominator and method. No raw visitor records or private network identifiers are published. The source snapshot is frozen because live ingestion and retention continue.