Who makes this

Why it exists

AI agents read websites all day and never say a word back. They arrive with a task, form an impression, sometimes get it wrong, and leave. The site owner sees a user-agent string in a log file and nothing else — not what the agent was looking for, not what it concluded, not what it told the person who sent it.

Every tool in this space works the same way: prompt a chatbot, see what it says about you, guess at the rest. That measures a simulation. Revorbi instruments the actual visit — an agent that was really there, on a real task, reporting what it really found.

The format is open and published at spec.revorbi.com. It is deliberately constructed so that an agent's safety filter can tell it apart from prompt injection: nothing is addressed to the model, nothing is asked of it, and participating changes nothing about what the page returns. That constraint is not decoration — it is the only reason a well-built agent would consider answering.

What we collect

Answers to at most four questions about a page, plus whatever identity an agent volunteers about itself. No end-user prompts, no personal data, no IP-linked identity. This is stated machine-readably as collects_user_data: false in every question document we serve, so an agent can verify it rather than take our word for it.

Contact

Everything goes to one address for now — support, security, press, and questions about the format:

[email protected]

Security reports are welcome — a note in the subject line helps them get routed. We would particularly like to hear from agent vendors and framework maintainers: the hardest part of this problem is not the format, it is that most agents have no code path to participate even when they want to.

Status

Early. The format is at version 1.9, the collector is live, and billing is not switched on. If you run a site and want to try it, the tag is one line and you can remove it as easily as you added it.