What agent readiness actually measures

Four questions, one score

The methodology lists every check under one of these four headings. The public score counts the core checks only, so a shop and a blog are measured on the same questions.

Fetch

An assistant reaches a site as an unattended HTTP client with no cookies, no JavaScript budget worth the name and no patience for a challenge page. If GPTBot, PerplexityBot or ClaudeBot is blocked in robots.txt, or the edge answers them with a bot wall, nothing downstream matters: the site is simply absent from the answer.

This is also the most common single reason a well-built site scores badly. A bot rule written years ago to keep scrapers out now keeps the buyers' assistant out too.

Read

Fetching returns HTML; reading requires meaning. A model resolves a page far more reliably when the facts are also present as structured data: what the entity is, who publishes it, what it costs, when it changed. Client-side rendering that leaves the server response empty is the equivalent of handing over a blank sheet.

The rubric rewards the machine-readable copy of a claim, not its visual presentation. A price rendered by JavaScript into a styled div is invisible; the same price in Product markup is quotable.

Trust

Assistants hedge. Given two comparable sources they name the one they can attribute: an organisation with an identity, content with a named author, a policy that states what automated clients may do. Attribution is not decoration; it is the reason one of the two gets cited.

Act

The last question is whether an agent can do something beyond reading: follow a feed, call an endpoint, complete a checkout. This is where the score stops being an SEO metric. A site that can only be read is a source; a site that can be acted on is a participant.