What AX Audit actually checks,
line by line.
29 checks, weighted to 100 points. This is the literal weight table every real scan runs — not a sales pitch, a mirror of the report you actually get. For the handful of these that are named, external industry specs (AP2, UCP, x402, and others), see the standards map.
Discovery
Can an agent find out what this site offers before it does anything else?
.well-known/mcp.json or the AR-CAP-01 MCP server card — static manifests an agent reads before connecting.
llms.txt, agents.txt, agent.json, agent-card.json, AGENTS.md, agent-manifest.txt — the well-known family of agent-facing index files.
.well-known/api-catalog (AR-CAP-04) — a standard linkset describing the site’s own APIs.
A skills manifest at the site root, distinct from AGENTS.md.
.well-known/ard.json, with .well-known/ai-catalog.json as a legacy fallback.
Content & Documentation
Once an agent is here, can it actually read and use what it finds?
/api, /docs, /openapi.json respond; cursor-based pagination patterns are detected.
Accept: text/markdown returns real Markdown (not an HTML fallback), ideally with YAML frontmatter.
JSON-LD markup, Schema.org vocabulary, OpenGraph tags.
robots.txt, sitemap.xml, and RSS/Atom feeds.
How many tokens an agent burns just reading the homepage — and whether an llms-full.txt bonus exists.
Meta description and an H1 present in the initial HTML — server-rendered content an agent can read without executing JS.
Security & Auth
Can a server trust who’s asking, and can an agent trust the server back?
OAuth/OIDC endpoints, Bearer token support, Content-Signal licensing signals, and a Web Bot Auth key directory.
/.well-known/security.txt (and /security.txt) with a real Contact and Expires field.
A real 404 returns structured JSON, not an HTML error page masquerading as a 200.
Standard rate-limit headers an agent can read to self-throttle instead of guessing.
A negative signal — passes when there is NO CAPTCHA/bot-wall blocking agent signup.
A Markdown-negotiated request to a path that doesn’t exist returns a real HTTP 404, not a 200-coded "not found" page.
Agent Experience
Is there a surface built specifically for an agent, not just a human UI an agent has to reverse-engineer?
document.modelContext tool registration quality — the one check that needs a real browser, not just an HTTP request.
A ?mode=agent request returns a meaningfully different, agent-optimized response — not the same bytes as the baseline page.
A conversational-search endpoint at /ask, graded with partial credit for an added /mcp JSON-RPC binding.
/.well-known/provisioning — machine-readable terms and consent for what an agent is allowed to do here.
Claimable-resource / anonymous-deploy language (env claim, --allow-anonymous, claim links) an agent can act on directly.
Commerce & Payments
Can an agent actually transact here — browse, check out, and pay — not just read?
.well-known/x402 and/or a live HTTP 402 challenge for stablecoin agent payments.
.well-known/ucp — a manifest describing a catalog, cart and checkout an agent can drive.
OPTIONS /checkout_sessions + POST /agentic_commerce/delegate_payment — shipped into ChatGPT.
Cryptographically signed payment mandates, detected via the UCP discovery profile’s capability list.
A live 402 challenge carrying WWW-Authenticate: Payment — a multi-rail competitor to x402.
.well-known/poppy.json — an identity manifest establishing whom a personal agent acts for.
A plain-text convention describing how an agent should register and authenticate itself.