How it is measured

Every check is automatic and reads only the public website: the homepage, robots.txt, the sitemap and the well-known files agents look for, about 35 requests a site, with no crawl.

Sites are read as a standard browser and as a plainly named checker, never under another company’s agent’s name and never through a proxy. robots.txt is read as a signal, not obeyed.

Measure, the sixth domain, is not scored: what agents do with an organization cannot be seen from outside.

Reachable is the check most often failed. Many large sites show a bot wall (a CAPTCHA or “checking your browser”) to any request that is not a full browser. Most agents fetching a page for a person are not full browsers, so the Index counts the wall as closed.

The sixteen checks

Each is pass, partial, fail or does not apply. The weight is its share within its domain.

DomainCheckHow it is measuredWeight
DiscoverReachableThe homepage answers a standard browser and the checker, with no challenge wall.3
DiscoverCrawl rules for AI agentsrobots.txt names AI agents and lets the ones acting for a person in.3
DiscoverSitemapA sitemap is found, with last-changed dates.2
DiscoverContent signalsrobots.txt says yes or no to search, AI input and AI training.1
DescribeReadable without scriptsThe homepage’s HTML carries its words before any script runs.3
DescribeStructured dataschema.org JSON-LD names the organization, its website and its logo.3
Describellms.txtAn llms.txt file exists, is well formed and links to pages.2
DescribeMarkdown for agentsThe homepage answers a request for markdown, or a Link header points agents to their files.1
DescribeEveryday pages linkedContact, pricing, support, legal and about pages are plain links from the homepage.1
ConverseAgent endpointA working MCP server, an A2A agent card or a Personal Agent Protocol file.3
ConverseInstructions for agentsAn agent skills index, or an agents.md or auth.md page.1
TransactActions a machine can listAn API catalog or a complete OpenAPI description.2
TransactAgent sign-inOAuth metadata tells an agent how to sign in on a person’s behalf.1
TransactAgent checkoutAn agent can check out through UCP, ACP or x402; does not apply to a site that sells nothing online.2
TrustSecurity contactA valid security.txt says whom to tell about a security problem.1
TrustSigned-agent supportA Web Bot Auth key directory lets the site’s own agents prove who they are.1

Scoring

A check
Pass earns full credit, partial half, fail none. A check that does not apply (no checkout on a site that sells nothing) leaves the total.
A domain
The weighted share of its checks’ credit, from 0 to 100.
Overall
The weighted mean of the domain scores: Discover 25, Describe 30, Converse 20, Transact 15, Trust 10.
Reach first
A site that fails Reachable scores 0 overall and in every domain: an agent that cannot get in sees none of the rest.
Ranks
By overall score, then level. Equal scores and levels share a rank.

The Maturity Model

The level is the highest rung whose conditions hold, each needing every rung below it.

LevelFrom outside, it means
0 · ClosedNot reachable, a challenge wall, crawl rules that shut agents out, or no words before scripts run.
1 · VisibleReachable and readable, with nothing shutting agents out.
2 · DescribedStructured data passes, and llms.txt or markdown for agents passes.
3 · ConversationalAn agent endpoint passes.
4 · CallableActions a machine can list, or agent checkout, passes.
5 · TrustedNot awarded from outside: needs a maturity review.

Editions and the sample

Monthly editions
An organization’s score in an edition comes from its latest check that month: the latest, never the best, whoever asked for it.
Reviewed, then published
In the first week of the next month the edition is frozen, reviewed for checker failures and outliers, and published with its method version.
Versioned method
A change to a check, a weight or a level is a new method version, announced with a date and applied to past editions too, so editions stay comparable.
The sample
Well-known organizations in the Institute’s industry networks, picked for size, recognition and a spread of countries. Nobody is in the Index because they asked or paid. Public sector is not measured yet.

Corrections

An organization that thinks its score is wrong writes to us with the subject “Index correction”. We answer within ten working days.

Our measurement was wrong
The site is checked again and the edition’s score replaced, marked Corrected with the date and the reason.
The site has changed since
Nothing is corrected; the next edition shows it. A recheck can be asked for at any time.
The method is wrong for everyone
The working group takes it up, and the fix comes as a new method version applied to every edition.

Ask for a correction. A correction never removes an organization from the Index.

Fairness: the Maturity and the Index working group

Keep the Maturity Model and the Index’s method fair, open and repeatable.

  • Scope: the checks, weights, levels and sample rules; disputed corrections; a yearly review of the method against how agents actually behave.
  • Membership: certified practitioners or members’ nominees, at most a third from any one employer and at most one from the data provider. Members declare employers and clients and step out of decisions about them.
  • It proposes; the Institute’s staff decide and publish the reasons.
  • It never looks at one organization’s score before a method change that affects it alone, never accepts payment tied to the Index, and publishes its decisions.

Forming. Until it meets, the Institute’s staff maintain the method and log every change.

Independence

The method, the sample, the editions and corrections are the Institute’s. Nobody pays to change a score, and being a member or a customer of anyone changes nothing about how an organization is measured.

The checks are run by a data provider, Common Square, whose readiness engine reads each site and records its raw signals. The Institute scores those signals by its own method; the provider’s own score is never used.

Disclosed: Common Square sells tools that help organizations raise their readiness, and the Institute was started by Common Square’s founder. The open method, the latest-not-best rule and the working group’s limits exist because of it. commonsquare.com is not in the sample.