Research
What we have measured
Nobody can show that AI search reads llms.txt. What can be measured is what the file does for an agent that reads it, and how much of the web has one. Each study below publishes its method, its version and its defects alongside its numbers.
Benchmark · phase 1 · September 2026
AI Read Cost: what llms.txt does to agent token spend
400 attempts across 20 domains. A small llms.txt made a site about twice as cheap to read. A large one made it about nine times more expensive. Size decides everything.
Registry · updated as the scanner runs
llms.txt adoption across the web
A ranked sample of the web, probed for llms.txt and re-probed on a schedule. Counts, not estimates, with every file's history kept.
Methodology
How AI Read Cost is measured
Routes, stopping conditions, tokens to answer, answer quality, and what we never claim. Rendered from the file the benchmark ships with.