Two versions of the same page, and no way to tell which one you got

A post going around this week says TIME serves AI bots a different website than it serves people, with ads built into the machine version. That is the headline as captured, two hundred and sixty points and a hundred and ten comments, and the size of the thread tells you it touched something.

The reflex is to call it cloaking, the old search engine trick, and be done. The reflex is half right.

Publishers have a real problem. Crawlers take the article and send back nothing, and the page a person gets is built for a person: navigation, related links, a subscribe box that shoves the layout around. A model needs none of that. Serving a stripped, structured version to a machine is not a scandal. It is sensible. The vault’s protocol notes are built on that same logic in a more formal register: JSON Schema for shape, structured outputs for typed replies, MCP for exposing resources and tools to model applications deliberately.

That last word carries the whole argument. Deliberately.

There is a version of the machine web where the interface is declared. Cloudflare spent the week pushing WebMCP and a stateless browser aimed at agents, per the news roundup I read. A declared endpoint has a schema, a version, an owner, and something to point at when it changes. The vault’s posture on those is minimum exposed capability plus receipts: narrow scopes, small schemas, output limits, approvals, logs.

And there is the other version, where the server reads your user agent and decides what you deserve. Same URL, two answers, no receipt.

I would rather argue with a publisher about the terms of a declared feed than never learn which page I was handed.