Loading
How we score a website for AI agents
Six areas, 24 checks, six recorded calls and a level from 0 to 5. This is what the Index reads and what it leaves out.
Method, October 1, 2026, 4 minute read. By Encore AI.
A customer who brings an AI agent to a website wants it to do what the customer would do: find information, check an account or an order, change a booking, reach a person. The Index asks how far an agent can get through each way into the organization.
Six areas
- Discoverability, 15 points. Can an agent find and read your pages?
- Protocol access, 20 points. Can an agent connect to you through MCP or an API?
- Identity and delegation, 20 points. Can a customer give an agent limited access without sharing a password?
- Web experience, 15 points. Does your site work when an agent uses it?
- Voice and IVR, 20 points. What happens when an agent calls your phone line?
- Policy and governance, 10 points. Do your terms allow agents that act for customers?
| points | |
|---|---|
| Discoverability | 15 |
| Protocol access | 20 |
| Identity and delegation | 20 |
| Web experience | 15 |
| Voice and IVR | 20 |
| Policy and governance | 10 |
The score is the share of available points an organization earns in the areas we could measure, scaled to 100. Each of the five website areas passes through one published calibration curve before it counts, so that a typical large organization scores near 50. The Method page gives the curve. A check we could not run, for example because the site blocked the scanner, is left out and is never counted as a failure.
A few checks are bonuses. A bonus check is worth nothing on its own, adds points when an organization has it and never takes points away. The first is the Poppy file for personal agents, a draft standard from Sierra and Meta. An organization that publishes a valid file earns up to three bonus points in Protocol access.
Levels
Each report also places the organization on a track of six levels. A level is earned only when every condition for it is met, so a high score cannot hide a missing capability.
- Level 0, Closed. Declared agents are blocked from the public website.
- Level 1, Tolerated. Pages load for an agent. Nothing is built for one.
- Level 2, Legible. An agent can read and cite your content.
- Level 3, Connectable. An agent can connect through an MCP server or a public API program.
- Level 4, Trusted. A customer can authorize an agent with limited, revocable access.
- Level 5, Native. Agents are a supported channel on web, API and voice.
How the scan behaves
- It identifies itself with a user agent that names Encore AI and links to gainencore.ai.
- It reads public pages and published technical files only.
- It loads the sign in page and may click a visible sign in control to see what opens. It never types into a field and never tries to authenticate.
- It compares what a declared agent receives with what a standard browser receives.
- When a host does not answer, the report says so.
The six calls
The Voice and IVR area is measured with 6 recorded calls to the number the organization publishes on its own website. The caller says it is an AI agent and that the call is recorded. It gives no personal information and stops at the first authentication step. Each report shows the recording, a timed transcript and the facts taken from the call.
Five calls are placed in business hours, on weekdays in the middle of the day. One call is placed at night. It reports that someone else is using the account and checks whether the line answers, states its hours and reaches a security or fraud path. Fraud does not wait for the office to open, so the index tests both.
The area counts toward the score once at least half of the calls have finished. Until then the score covers the other five areas.
Limits
The scan sees public pages. It cannot see private partner interfaces, contracts or anything behind a sign in. A missing signal means it was not found at the addresses checked. Findings are dated, and an organization that publishes a new route will score higher on its next scan.