How the AgentReady score works

What the score is — and is not

• ✓ A readiness measure: which facts and endpoints an agent needs are present, parseable and consistent on your store, today.
• ✓ Reproducible: two scans of an unchanged store return the same number. Every point is traceable to a named check on a fetched page.
• ✗ Not a conversion probability. A score of 82 does not mean 82% of agent purchases succeed — it is a readiness proxy, not a promise of revenue. What readiness correlates with is a question we answer with published data, not marketing.
• ✗ Not an industry standard. It is our methodology, published and versioned here. Where an open standard exists we align with it and say so — see our public conformance self-audit.

The four pillars and their weights

Score = 100 × Σ (pillar weight × completion ratio) / Σ weights. A pillar that is entirely not-applicable is renormalized away — "n/a" never penalizes.

Pillar	Weight	What it answers
🚪 Access	34	Can an agent get in at all? robots.txt and WAF posture for the shopping-agent class, crawlability, latency, and whether bots see the same page humans do.
🏷 Structured data	30	Can an agent understand the offer? Product/Offer JSON-LD with price, currency, availability and identifiers it can parse without guessing.
📦 Catalog	20	Can an agent navigate what you sell? Feeds, variants and the attributes a buying decision needs, checked against your vertical.
🛤 Transaction rails	16	Can an agent transact? The emerging agent-commerce rails — deliberately weight-limited (see below).
Total	100

"Do now" vs "Forward-looking" — read the score in two layers

The first three pillars (84 of 100 points) are Do now: mature, stable practices every store should implement this quarter. The transaction-rails pillar (16 points) is Forward-looking: emerging agent-commerce protocols, deliberately weight-limited because they are young and still moving. Almost every store scores near zero there today — that is expected, and it is where early movers pull ahead.

This split is deliberate honesty: it protects you from over-investing in a protocol that may change in six months, while still measuring who is ahead of the curve.

Forward-looking rail	Points	Effort
UCP agent endpoint (/api/ucp/mcp)	4	hard
A2A Agent Card (/.well-known/agent-card.json)	3	easy
ARD AI Catalog (/.well-known/ai-catalog.json)	3	easy
Machine entry points (feeds/APIs, not just HTML)	3	easy
Returns & shipping, machine-readable	2	medium
Add-to-cart / checkout markers	1	hard

Grades, and the one hard cap

A ≥ 85 · B ≥ 70 · C ≥ 55 · D ≥ 40 · E below. Since scoring v2, grade A means buyable — rails answering — not merely readable.

The cap: if a store blocks AI agents at the door (robots.txt blanket-blocking agent crawlers, or a WAF blocking them in practice), the score is capped at 25 no matter what else is right — an agent that can't get in can't buy.

Version history

• v2 — 2026-07-15 (current). Weighted matrix over pillar completion ratios; the transaction rails became a scored pillar. Reason: with read-layer checks alone, mature stores clustered at 90+ and grade A had stopped meaning anything. Reports made before v2 keep their original rendering and are excluded from v2 rankings.
• v1 — 2026-06. Three read-layer pillars, plus emerging-protocol checks as additive-only bonus points (max +10, nobody could lose a point).

How we verify — and where we stop

• We fetch public pages only, at a merchant's request, politely.
• Checkout simulation stops at the checkout boundary — we never place an order, never enter payment data.
• Our crawler identifies itself and cryptographically signs its requests (Web Bot Auth) — verify us at /bot.
• No model grades a store. Anywhere a model helps (fix suggestions, plain-language explanations), it is labeled advisory and never feeds the /100.