Agent shopping sessions, observed
UCP Playground runs real AI models against real UCP stores and records every tool call. Here is what its sessions show, from a snapshot taken on 29 September 2026.
UCP Playground is a separate product at ucpplayground.com.
26models run
8model vendors
531stores shopped
25.3%of sessions reached checkout
How far sessions got
Each session ends at the furthest point it reached. Reaching checkout includes the sessions that went on to pay.
The most-run models
Share of sessions shows which models people choose to run, not which shop best. Tokens and time are averages per session.
| Model | Share of sessions | Tokens | Time |
|---|---|---|---|
| Gemini 3.6 Flash | 14.1% | 75k | 22 s |
| Gemini 3 Flash | 13.2% | 77k | 23 s |
| Claude Sonnet 4.5 | 12% | 70k | 34 s |
| Claude Opus 4.6 | 7.9% | 61k | 30 s |
| Gemini 2.5 Flash | 6.8% | 45k | 14 s |
| GPT-5.2 | 6.7% | 52k | 33 s |
| Gemini 3.1 Pro | 5.3% | 62k | 42 s |
| GPT-4o | 5.1% | 34k | 16 s |
| Gemini 2.5 Pro | 5.1% | 47k | 37 s |
| DeepSeek v3.2 | 3.1% | 69k | 47 s |
| o4-mini | 2.6% | 70k | 40 s |
| Grok 4 | 2.6% | 36k | 45 s |
14 more models ran fewer sessions. Run any of them on any store in UCP Playground.
The tools agents called most
How long each answered on average, and how often it returned an error. Names are as the stores serve them: some are the UCP binding's, some a platform's own.
| Tool | Average answer | Errors |
|---|---|---|
search_catalog | 725 ms | 21.9% |
update_cart | 609 ms | 32.5% |
search_shop_catalog | 483 ms | 26.9% |
search_global_products | 347 ms | 0% |
get_product_details | 266 ms | 25.1% |
get_cart | 279 ms | 43% |
create_cart | 916 ms | 36% |
POST /catalog/search | 4,543 ms | 37.4% |
search_products | 1,229 ms | 14.3% |
get_product | 1,043 ms | 35.4% |
Watch an agent shop your storePick a model and a store; every tool call is recorded and replayable.
Open UCP Playground
What is UCP Playground?
UCP Playground is an agent shopping runtime at ucpplayground.com. It runs real AI models against real UCP-verified stores, logs every JSON-RPC tool call, and makes every session replayable. Think of it as Postman for agentic commerce — observability for agent behaviour, not a vendor benchmark.
How is UCP Playground different from UCP Checker?
UCP Checker evaluates your profile, surface signals, and lightweight probes — fast, safe, runs at scale. UCP Playground runs full end-to-end agent shopping simulations against your live infrastructure. Checker tells you "can agents find and parse you" — Playground tells you "can agents complete a transaction".
Is UCP Playground a UCP demo?
Yes — it functions as a live UCP demo. Every session is a real AI model exercising real UCP capabilities (search, cart, checkout) against a real store, in real time. Unlike a recorded demo video, every Playground session is reproducible: pick the same model and store and run it again. Several stores publish demo profiles specifically for this purpose (
demo-travel.ucp.dev, sandbox subdomains) so you can watch UCP in action without affecting production inventory.Which AI models are run most?
Gemini 3.6 Flash is the most-run model, with 14.1% of sessions, across 26 models from 8 vendors. Session share shows which models people choose to run, not which shop best.
How is the checkout completion rate calculated?
Sessions that reach a checkout state divided by total sessions. Currently 25.3% of sessions reach checkout. The remainder split between cart-created, search-only, and failed sessions — the funnel above shows the full breakdown.
How many stores has UCP Playground tested?
531 unique UCP-verified stores as of 2026-09-29. Sessions span Shopify, WooCommerce, Magento, custom Rails apps, and bespoke MCP servers. The same agent code runs against all of them — anything different in the outcome is the store, not the agent.
Is UCP Playground free to use?
Yes — running sessions is free at ucpplayground.com. Pick a model, pick a store, watch the agent shop.
How do I run a UCP test on my own store?
To run a UCP test against your own store, point Playground at your domain. If you have a valid
/.well-known/ucp profile and reachable MCP or REST endpoints, agent sessions will run against it directly. Stores that grade well on UCP Score typically also pass UCP tests in Playground, but a Playground test surfaces runtime issues that static checks cannot — schema drift, slow tool responses, partial cart state, missing variant data.