The llms.txt spec, who actually serves one in 2026, the honest evidence on whether crawlers read it, and the anatomy of a real trilingual implementation.
There are two conversations about llms.txt and they rarely meet. In one, it is the new robots.txt, the file every site needs before AI search eats the web. In the other, it is a cargo-cult artifact that no crawler reads. Both conversations quote the same three sources and neither usually involves someone who actually runs the file in production and watches the logs.
I run one. kescoda.com has served a trilingual llms.txt and a full-corpus llms-full.txt since summer 2026, regenerated automatically at every build. This article covers the three layers in order: what the proposal actually says, what the adoption and crawler evidence honestly shows as of August 2026, and how my implementation works, including what I can and cannot measure yet. My file has been live for weeks, not months, so you will find first data and a measurement method here, not miracle graphs.