Can an AI agent
buy from your store?

Because every protocol above solves payment and identity — the last step. None of them finds the product, picks the size, or gets through checkout; today's shopping agents do that by driving the same web a human sees, since it's the only surface every store already has. Agent-native integrations arrive store by store, and the D2C long tail adopts last — so for stores like these, the browser journey is the agentic channel for years to come. And the walls this benchmark finds — OTP gates, forced logins, bot-blocks — are decisions about who may transact, not markup details. The stores failing an agent today are the ones that will be slowest onto the new rails.

It evolves with the surface — the rubric is versioned for exactly that. Rubric v1, this leaderboard, measures the browser journey: the 2026 reality. As agent-native surfaces appear, future versions shift with them: the static half becomes “does the store publish agent affordances — product feeds, ACP or UAP endpoints”, and the live half becomes “can an agent actually transact through them”. What the index measures never changes: the gap between agentic demand and merchant readiness.

Because readable is not operable. Across the 43 ranked stores, static machine-readability explains under 3% of what an agent actually achieves — stores with perfect structured data score zero on the journey, and stores nearly illiterate to a crawler walked all the way to payment. Every wall in the failure taxonomy is invisible in markup; it only appears when something tries to buy.

That is the failure mode this benchmark is built to avoid. When the agent itself fumbles, the run is voided as agent error and never counted against the store. A finding is published only when at least two of three independent runs reproduce it, and stores where nothing conclusive survived are excluded from the ranking entirely rather than shown with a number. The full policy is on the methodology page.