The failure happens on the assistant's side of the conversation.
The customer is told it could not filter your products, and moves on. You never see the visit at all, and nothing in your analytics registers a problem.
How it works
By default we send our own honest user agent, never a vendor's. Where we do measure a named agent, the result says what your site serves a request *claiming* that name — we send the token from our own address, unsigned, so it is not a measurement of what the vendor's own fetcher receives.
No JavaScript is executed, because the major AI crawlers do not execute it either. Anthropic states it plainly: the web fetch tool does not support sites rendered dynamically with JavaScript. An emulation with a headless browser would systematically over-report what an assistant can actually see.
Then we measure the action space. A text agent has exactly two actions — follow a URL it has seen, or submit a form whose fields are in the HTML — so every control your pages offer either reduces to one of those or does not exist for the agent. We count the ones that do not, and check whether a link covers the same ground. Every finding carries the page it came from.
What we measure — and what we do not
Measured
Whether an assistant that reaches your site can navigate and use it: whether your pages serve it anything readable at all, and whether the controls they offer reduce to something an agent can act on — a link it can follow or a form it can submit.
Not measured
Not whether assistants reach you at all. Merchant-feed enrolment bypasses the website entirely, index membership is only partly observable, and a refusal for our scanner is never stated as proof that a named assistant is refused. Not your content either: we do not read your prices, stock states or opening hours, and we never grade whether one is there.
Caveats
Sampling, not audit: from a single vantage point and locale.