The measurement programme

Episodic studies, not a running benchmark.

Each study asks one question of a fixed corpus in a fixed window, publishes every figure with its denominator, and stays citable afterwards. There is no continuous benchmark behind this site, and this page will not imply one until there is.

Published

AIR · 29 August 2026 · 680 origins · 47,600 requests · repeated 12 September

/robots.txt is the exception. Everything else is refused at twice the rate.

A differential study of what an origin serves a declared AI agent against what it serves a browser, across two networks and five client identities. It separates how much of an agent’s disadvantage comes from being an agent from how much comes from asking for a path the web was not built to serve — and finds that about half of it is not about AI at all. Repeated on 12 September on the same pages: four profiles moved by under a point, the declared agent by four, and the agent-native surface — MCP manifests, OpenAPI descriptions, Markdown on request — turned out to be nearly empty.

Read the findings · The long-form report · Every figure as JSON

Not run

What this programme cannot yet tell you.

No trend

One measurement window. There is no second run of the same corpus, so nothing here can say whether access is getting better or worse.

No score

The scoring model is specified on /method and is not in production. No origin in any study has been scored or ranked.

No operability data at scale

The second pillar is defined and measured per page, but no study has yet reported it across a corpus. It carries no weight for exactly that reason.

Every figure in a published study is registered with its numerator, denominator and egress arm before it appears, and a build check re-derives each one from the stored response rows. If one is wrong, the corrections process is how to say so.