rubric

The Agent-Readiness Rubric

By Project Auxo · 2026-06-29 · v1

Why a rubric

The web is being visited by software acting on people's behalf, and there's no shared answer to a simple question: is this site ready for agents? Agent-readiness is measurable — from public signals a site already does or doesn't publish. This is the open methodology behind the agent-readiness report card; it's versioned, and we publish it so the score is auditable rather than a black box.

Discoverable · 50 pts · Can an agent find you and what you offer?

SignalWhat we checkPts
robots.txtPresent5
AI crawlers allowedGPTBot, ClaudeBot, PerplexityBot, CCBot, Google-Extended not blocked10
sitemap.xmlPresent or referenced in robots.txt5
llms.txtPresent — what is worth reading10
schema.org JSON-LDStructured data on the homepage8
capabilities.txtPresent + conformance grade12

Invocable · 30 pts · Can an agent actually act?

SignalWhat we checkPts
OpenAPI discoverableA spec at a common path agents can map18
MCP advertisedmcp.json / a live MCP server (optional)12

Identifiable & Trusted · 15 pts · Can an agent recognize your identity?

SignalWhat we checkPts
A2A agent card/.well-known/agent-card.json15

Provable · 5 pts · Can you prove what agents did?

SignalWhat we checkPts
Verifiable evidence artifactA cryptographically verifiable proof of a real governed action at /.well-known/chp-evidence.json — the grader recomputes it, so it is checked, not trusted5

Scoring

Each signal earns points within its category; the four categories sum to a 0–100 score and an A–F grade (A ≥ 85, B ≥ 70, C ≥ 55, D ≥ 40, F below). Most of the web scores C–F today — agent-readiness is new, and the point of the grade is the prioritized list of fixes, not the letter. The methodology is versioned; this is v1.

What is agent-readiness?

How prepared a website or API is for AI agents to discover what it can do, invoke it, recognize its identity, and be held accountable for consequential actions. It is the agentic-web analogue of mobile-readiness or accessibility — a measurable property of a site, not a product you buy.

How is the score calculated?

From public, deterministic signals at well-known locations — robots.txt AI-crawler rules, sitemap.xml, llms.txt, schema.org JSON-LD, capabilities.txt (graded for conformance), OpenAPI, MCP, and A2A agent cards. Each signal carries points within one of four categories; the categories sum to an overall 0–100 score and an A–F grade. No private data, no login required.

Why weight capabilities.txt as only one signal?

Because agent-readiness is a property of the whole stack, not any single file. A site can be discoverable via llms.txt and schema.org, invocable via OpenAPI, and still have no way to prove what an agent did. Weighting one file heavily would make the score self-serving and less useful. capabilities.txt is one Discoverable signal among several.

What is the "Provable" category, and how do you earn it?

It asks whether you can prove, later and to someone skeptical, what an agent actually did on your site — and whether it was authorized. Most sites still score zero because they have no standard evidence to point to. You earn it by publishing a VERIFIABLE evidence artifact at /.well-known/chp-evidence.json: an RFC 6962 Merkle store-head inclusion proof that a real governed action is committed under your append-only evidence head ({ scheme: "chp-store-head-v2", store_head, inclusion: { correlation_id, head_hash, leaf_index, tree_size, audit_path }, event_summary }). The grader recomputes the SHA-256 audit path — it must reproduce the published store_head — so the proof is checked, not trusted, and cannot be faked; no login required. This is the layer the Capability Host Protocol provides, and any host that emits governed, hash-chained evidence can produce one (e.g. the chp.adapters.audit.inclusion_proof capability). Maturity: v1 verifies inclusion under your published head; stronger tiers add a witnessed or externally-anchored signature over that head for full third-party authenticity.

How do emerging standards affect my score?

Stable, widely-supported signals (robots.txt, sitemap, llms.txt, schema.org, OpenAPI, capabilities.txt) carry the weight. Fast-moving ones (A2A agent cards, Web Bot Auth, MCP advertisement) count but are weighted lighter, so your score does not swing as those specifications evolve.

Get your score

Run the report cardMake your API agent-readyHow agents discover sites