Your next customer might be an agent

AI agents are starting to buy, book and sign up on behalf of people, and most sites cannot support it. Buttons go unlabelled, errors exist only as a red outline a model cannot read, and cart state changes in ways nothing on the page announces. The tools that claim to test for this check whether the right metadata files are present, which says nothing about whether an agent can finish the job.

AgentAudit tests the only thing that matters, which is completion. A real agent is sent through the checkout, booking or signup flow, and what happens is scored out of a hundred. The run happens twice, once with the agent seeing nothing but the page, and once with the site's own tools exposed to it, so the gap between those two numbers is a measurement rather than an opinion about how agent friendly a site feels.

Where the agent stalls, the report does not stop at naming the step. It writes the fix, checks that fix against the real page, and then runs the live journey again with the fix applied.

Only a passing re-run gets stamped verified, which is the discipline most scanners skip. Anything can produce a list of suggestions, and a suggestion that has never been executed against the page it describes is a guess with formatting.

The uncomfortable part for most sites is that nothing is broken in the human sense. Everything works for a person who can see, infer and guess, and the agent simply has none of those, so it exposes every place where meaning lived in a visual convention rather than in the page itself.