Skip to main content

How Agent Readiness is scored (method v6)

Agent Readiness is scored out of 100: AI Trust counts 40 and AI Readability counts 60 (Machine-readable 36, Schema 24). Agent-integrated is an add-on outside the score.

Levels

  1. Not reachable: Measured, and Reachable is not met.
  2. Reachable: The homepage is readable in the raw HTML, robots.txt lets in at least half the AI agents we test, and pages can be found (sitemap, Link headers or homepage links).
  3. Readable: Reachable, and Machine-readable at 60% or more.
  4. Trusted: Readable, and: valid certificate, served over HTTPS, HTTP redirects to HTTPS, no downgrade to HTTP, HSTS, DNSSEC not broken, DMARC at quarantine or reject, SPF valid.
  5. Agent ready: Trusted, and: score 70 or more, AI Readability at 60% or more, AI Trust at 70% or more.

How the score and levels work

  • Score = round(40 x AI Trust share + 60 x AI Readability share), between 0 and 100. AI Readability is Machine-readable (36) plus Schema (24). A share is the points earned over the points we could check.
  • A check we could not run for a reason of our own leaves the share's denominator and costs nothing.
  • A score is provisional when we could check less than 90% of either vector, or could not read robots.txt. A provisional site's level can hold or fall, but not rise.
  • A connection problem (certificate, HTTPS, redirect to HTTPS, downgrade to HTTP, DNSSEC broken) holds the score at 69 until it is fixed. Nothing else is a cap.
  • Levels are cumulative: each needs the one before. A numeric line (70 points, 60% and 70% shares) must be cleared by 2 points to move the level up and missed by 2 to move it down. Exactly on a line counts as reached. A pass or fail condition has no buffer.
  • A level change shows after 2 runs agree. We run a confirm check within a day of a change.
  • One badge, AGENT TRUST, for every site at Reachable or above, showing its level name. Not reachable and not measured sites have no badge.
  • Every check is graded on what we observe from outside, the same way for every company, against published standards. A score never rewards using our product.

AI Trust: can AI trust this domain? (40 points)

CheckPointsFixed byHow it is tested
Valid certificate 2.2 IT RSA 2048 bits or more, or ECDSA, and no SHA-1 signature. A confirmed failure holds the score at 69.
Served over HTTPS 2.2 IT The site is served over HTTPS. A confirmed failure holds the score at 69.
HTTP redirects to HTTPS 2.2 IT http:// redirects to https://. A confirmed failure holds the score at 69.
No downgrade to HTTP 2.2 IT No redirect goes from HTTPS back to HTTP. A confirmed failure holds the score at 69.
HSTS 2.2 IT A Strict-Transport-Security header with a max-age of at least one year. Six to twelve months is a warning (0.75); shorter or missing costs 1.5 and never caps.
DNSSEC 2.2 IT DNSSEC is signed and validates. Unsigned costs 1; broken caps. A confirmed failure holds the score at 69.
Mail provider 2.2 IT Mail servers are published, or a null MX (RFC 7505) says there is none. Neither costs 1.5 and never caps.
DMARC 4.2 IT A DMARC record with p=reject or p=quarantine at pct=100. pct under 100 costs 1 and does not cap; p=none costs 1.5 and caps; no record costs 3 and caps. Those are the whole cost: the subdomain and reporting checks do not apply then.
SPF 1.1 IT A valid SPF record. No record costs 1.5 and caps; ?all or no all costs 1.5 and does not cap; +all, two records or over 10 lookups cost 1.5 and cap. SPF never costs more than 1.5.
SPF within 10 lookups 1.1 IT SPF stays within 10 DNS lookups; over 10 costs 1.5 and caps. Does not apply when the SPF record is missing or already fails.
DKIM 2.2 IT A DKIM key under a selector we can find; caps only when confirmed broken.
Certificate not near expiry 1.4 IT The certificate is not near expiry.
CAA 0.7 IT A CAA record.
CAA names an issuer 0.7 IT The CAA record names an issuer. Never caps.
Old TLS versions off 1.4 IT TLS 1.0 and 1.1 are turned off.
No insecure content 1.4 Developer No insecure content on the homepage, and no form that posts to http://.
Domain not near expiry 1.4 IT The domain registration is not near expiry.
Bare domain and www both work 1.4 IT The bare domain and www both work.
Content-Security-Policy 1.4 Developer A Content-Security-Policy header.
Frame protection 1.4 Developer Frame protection (X-Frame-Options or frame-ancestors).
Basic security headers 2.6 Developer X-Content-Type-Options: nosniff.

AI Readability, Machine-readable (36 points)

CheckPointsFixed byHow it is tested
Agents can open your homepage 4.8 Developer Two looks from our own servers: our signed checker, and a plain browser. Full points when our checker reads the homepage; half when only the plain browser does; none when neither does. Reading means no block or challenge and readable main text in the raw HTML.
AI crawlers allowed 4.8 Developer robots.txt (RFC 9309) lets in the agents that act for a user and the answer engines, on the homepage and the sampled paths. Training crawlers are information only. An unreadable robots.txt fails.
Sitemap 4.8 Developer A sitemap we can fetch and parse.
Pages agents can read 6 Developer On each fetched page: no block, and readable main text in the raw HTML. Partly when only the plain browser read it.
Title, description and excerpt 3.6 Content and product On each fetched page: a title, a meta description and an og:description, all non-empty. Partly when one or two are.
Headings 2.4 Content On each fetched page: at least one H1, plus an H2 when the main text is over 300 words.
Clean copy 2.4 Content and product On each fetched page: no unbalanced quotation marks, garbled characters or unfilled template text in the title, description, headings or main text. Code blocks, inch marks and « » are ignored.
llms.txt 0.9 Developer An llms.txt that is served and parses. Highest rung counts.
llms-full.txt or a Markdown link 1.8 Developer llms-full.txt, or Markdown (Accept: text/markdown, or a fetchable rel=alternate Markdown link) on the homepage and at least half of the sampled pages. The homepage alone is partly. Markdown content negotiation is shown here as evidence, not scored apart. Highest rung counts.
Markdown version 3.6 Developer Highest rung counts.
Link preview when shared 1.2 Content og:title, og:description and og:image on the homepage.
Canonical link 1.2 Developer A canonical URL on the homepage.

AI Readability, Schema (24 points)

CheckPointsFixed byHow it is tested
Organization on your homepage 6 Content An Organization node in the homepage JSON-LD.
Organization complete for its type 7.2 Content A description, three or more official profiles (sameAs; a profile counts unless it is confirmed broken), and three of: contact, address, legal name, founding date, related organizations.
Page markup valid and complete for its type 6 Content On each fetched page: JSON-LD in the fetched HTML that parses and is valid schema.org. Markup added only by script does not count.
Markup agrees with the page and the company 4.8 Content and product On each fetched page: the publisher or author organization carries the homepage Organization's name, URL and logo, and links to it.

Reported, not scored (0 points)

Shown with their result. Rows marked bonus are not scored yet and are never a cap.

CheckPointsFixed byHow it is tested
Privacy and terms pages 0 Content The homepage links a privacy page and a terms page. Reported only
Secure cookies 0 Developer Cookies the homepage sets carry the Secure attribute. Reported only
No known-vulnerable scripts 0 Developer JavaScript libraries matched by retire.js; caps when a script loads from a known-compromised host. Server and CMS versions are information only. Bonus
No subdomain takeover 0 IT One GET of each known CNAME'd subdomain; a finding is a takeover-prone service showing its own unclaimed-site page. Bonus
STARTTLS on mail servers 0 IT The mail servers offer STARTTLS. Extra credit, and not measured until a probe host can reach port 25; then Fixes offers a prompt to check it yourself. Reported only
MTA-STS 0 IT An MTA-STS policy. Reported only
DMARC reporting 0 IT DMARC reports go to an address, and an external one has authorized it. Does not apply when DMARC is p=none or missing. Reported only
BIMI 0 IT Optional: a BIMI record. Never a failure. Reported only
Universal Analytics retired 0 Developer No leftover Universal Analytics tag. Reported only
Verified bot identity 0 Developer Our checker's identity can be verified by the site (Web Bot Auth). Reported only
Tags 0 Content Reported only
AI use preferences 0 Content A line in robots.txt that says how your content may be used for search, AI answers and AI training, parsed from a `Content-Signal:` directive. Reported only
Says what it offers and to whom 0 Content and product On each fetched page: a category and a specific audience ("Everyone" or "All" does not count). On the homepage, knowsAbout and areaServed count instead. Reported only

Agent-integrated (add-on, not in the score) (outside the score)

Whether an agent can act with you. It never changes the score or the level.

CheckPointsFixed byHow it is tested
Agents can find your AI tools 0 Developer A discovery file (agents.json, A2A card, MCP server card, skills index, resource catalog, API catalog, ARD manifest, or an llms.txt link) points at an interface that passes its proof call.
DNS-AID 0 Developer Extra credit. Your DNS publishes an SVCB agent discovery record under _agents. Never a failure.
WebMCP 0 Developer A read-only WebMCP tool on the homepage or the contact, demo or pricing page, called by our browser, answers with content.
MCP server 0 Developer initialize, then resources/list and resources/read. Otherwise one tool, only if it is marked read-only, its name reads as a read, and it needs no arguments. We never send arguments that could write.
A2A 0 Developer An agent card names an agent, and asked for a made-up task the agent answers that it has no such task, or with a sign-in challenge an agent can follow. We never send a message.
API agents can use 0 Developer A documented GET, HEAD or OPTIONS with no body returns real content, or a 401/403 with a challenge an agent can follow. Redirects stay on your domains.
Agent checkout 0 Developer ACP or UCP discovery on your domain, and its capabilities read. We never create a cart or send a payment token.
Contact path for agents 0 Developer An agent can get from the homepage to the demo, contact or pricing page, and its form or booking link has labeled fields and no visible challenge. Standard embeds and invisible CAPTCHA pass. Never submitted.
OAuth and OpenID discovery 0 Developer Required when we see a real sign-in: a 401 or 403 sign-in challenge on one of our proof calls, OAuth protected-resource metadata, or OAuth or OpenID metadata we found. Then: sign-in server metadata (RFC 8414 or OpenID discovery) at your site's root or through your protected resource metadata, with an issuer that matches where it is published. A WebMCP tool, an open MCP server or API, or an interface we found but could not try is no sign-in: then it does not apply.
auth.md agent sign-up 0 Developer Extra credit. Your sign-in server's metadata carries an auth.md agent_auth block with the fields auth.md requires. Never a failure.
Personal Agent Protocol 0 Developer Listed with the extra credit, worth 0 points until the Personal Agent Protocol has a published specification. /.well-known/poppy.json with a protocol version that names your domain. Never a failure.
x402 payments 0 Developer Extra credit where we detect agent checkout or a payment interface (the three payment checks earn +1 between them): your site asks for payment the x402 way. Otherwise not applicable, and not shown.
Machine Payments Protocol 0 Developer Extra credit where we detect agent checkout or a payment interface (the three payment checks earn +1 between them): your OpenAPI description says which requests take payment, the Machine Payments Protocol way. Otherwise not applicable, and not shown.
Agent Payments Protocol (AP2) 0 Developer Extra credit where we detect agent checkout or a payment interface (the three payment checks earn +1 between them): your A2A agent card declares the AP2 extension. Otherwise not applicable, and not shown.

The previous method, v5