Research, guides and ecosystem updates
Original research on UCP adoption, the State of Agentic Commerce report, guides for each store platform, and the latest spec changes, written for developers and platform teams building for agentic commerce.
What shopping agents do when the store misbehaves
We built a UCP flower shop that fails in six specific ways and ran 150 sessions across Claude, GPT, Gemini and Grok, recording what the store received and what the agent told the customer. No model invented an order. Two followed instructions hidden in a product description in every session. One used a leaked override token five times out of six.
· 14 min read
Which AI Models Can Actually Complete a Checkout? The Benchmark Just Refreshed to the Frontier
UCP Playground's agentic-checkout leaderboard now spans 20 models across 8 providers — Claude Opus 5, Grok 4.5, Kimi K3, Qwen 3.7 Max and more, with the previous generation kept as lower-cost options. Two decisions make it a price-performance benchmark you can cite, not just a capability ranking.
· 3 min read
UCP Playground at 1,000+ Agent Sessions: What 16 Models and 97 Real Stores Reveal About AI Shopping
Two and a half months after we shipped UCP Playground with 114 agent sessions, the dataset has crossed the 1,000-session mark — across 16 frontier models, 97 real UCP-enabled stores, and roughly $96K of agent-driven cart value. Here are the patterns the larger sample is now strong enough to anchor.
· 16 min read
Agentic Commerce Optimization: What 4,491 Merchants Reveal About UCP Readiness
Everyone's publishing guides on how to get UCP ready. We measured who actually is. Data from 4,491 merchants on capability adoption, platform gaps, and what agents really do at checkout.
· 11 min read
We Tested 5 AI Models on Expiring Travel Inventory — Here's How They Failed
We built a mock travel server with expiring offers and tested Claude, GPT, Gemini, and Grok against it. No model checks TTLs before booking. Only one survived a 3-second expiry window. None flagged price changes unless asked. Here's the full data from 21 production sessions.
· 11 min read