The method
How the agent-readiness score works
"Agent-ready" has no industry definition, so here is ours, rule by rule. A site is agent-ready when an AI agent acting for a person can find it, understand it, use it for that person, and be trusted to. Four pillars, 25 points each, 100 in all, read from public documents only.
Grades
- A85+ · Agent-ready
- B70+ · Nearly there
- C50+ · Visible, not usable
- D30+ · Hard for agents
- F0+ · Closed to agents
1. Found · Can AI assistants reach the site?
| Rule | Points | Reads |
|---|---|---|
| The home page answers over HTTPS | 4 | the site |
| AI assistants acting for a person may visit | 15 | RFC 9309 |
| AI search engines may index the site | 6 | RFC 9309 |
2. Understood · Does it tell agents what it is and sells?
| Rule | Points | Reads |
|---|---|---|
| Structured data on the home page | 8 | schema.org |
| The business is described (Organization, LocalBusiness or Store) | 5 | schema.org |
| What is sold is described (Product, Offer or Service) | 6 | schema.org |
| llms.txt | 4 | llmstxt.org |
| A sitemap | 2 | sitemaps.org |
3. Usable · Can a customer's agent act here?
| Rule | Points | Reads |
|---|---|---|
| Publishes /.well-known/poppy.json | 5 | Personal Agent Protocol, Draft 0.2 §3 |
| Every rule an agent checks before using it passes | 12 | Personal Agent Protocol, Draft 0.2 §3 |
| Customers can sign in through their agent | 4 | Personal Agent Protocol, Draft 0.2 §4.4 |
| Something an agent can use: an API, an MCP server or a Company Agent | 4 | Personal Agent Protocol, Draft 0.2 §3.1 |
4. Trusted · Is that access accountable?
| Rule | Points | Reads |
|---|---|---|
| Agents prove who they are and act only with the customer's consent | 10 | Personal Agent Protocol, Draft 0.2 §4 |
| A customer can disconnect an agent | 5 | Personal Agent Protocol, Draft 0.2 §3.2 |
| Changes after the sale are handled as approved operations | 7 | Operations extension |
| HTTPS end to end | 3 | Personal Agent Protocol, Draft 0.2 §3 |
What the check reads, and what it never does
It fetches, over HTTPS and following redirects: /robots.txt, the home page, /llms.txt, /agents.md, /sitemap.xml (when robots.txt names none), /.well-known/poppy.json, the issuer's OAuth server metadata, and /.well-known/ucp. It identifies itself as Agent Adapt in its user agent.
It never signs in, never sends a token, never submits a form, never reaches a private address, and stores only the result, under the domain, for the report's link and badge.
Common questions
›Why four pillars of 25?
So that housekeeping any site can do (being found and understood) and the protocol work (being usable and trusted) weigh the same. A site with perfect SEO and no way for an agent to act scores at most 50.
›Why does blocking GPTBot cost nothing?
GPTBot, ClaudeBot, CCBot and the like fetch pages to train models. Whether to allow them is a separate decision. The check scores only the assistants that fetch a page because a person asked (ChatGPT-User, Claude-User, Perplexity-User) and the AI search bots that let answers cite you.
›Why partial credit for UCP?
A Shopify store's UCP profile lets a shopping agent browse and build a cart, which is real, so the Usable pillar gives 10 of 25 for it. UCP has no customer sign-in, revocation or operations for a person's agent, so the Trusted pillar can't score it.
›Can the score be gamed?
Only by doing the things it measures: the rules read the public documents an agent reads. A poppy.json that fails the issuer's checks earns less than one that passes, and the report says which check failed.
›Will the method change?
Yes, as the Personal Agent Protocol changes (it moved from Draft 0.1 to Draft 0.2 on 18 October 2026) and as agents change what they read. The version is in every stored report, and a change is announced here.