Testing your store
A working UCP profile and endpoint are the start. Testing shows whether agents can complete a purchase on your store, not just connect to it.
Recommended workflow
Work through these steps in order, from discovery to checkout:
- Check discovery — open your store in Inspect (or the public Inspector) and confirm your UCP profile at
/.well-known/ucpis found and read correctly. Check that every service endpoint and payment handler appears as you expect. - Call each tool by hand — in Inspect, the Steps view walks your UCP flow in order and the Tools view lists every tool with a form built from its schema. Start with
search_catalog(orsearch_shop_catalogon older Shopify stores), then product lookup, then the cart and checkout tools. Check that each response is valid JSON with the shape you expect. - Fix schema issues — look for the usual problems: required fields missing from your tool definitions, ID formats that differ between search results and lookups, prices in dollars where cents are expected, and missing variant IDs that agents need for the cart.
- Start a run — on Runs, pick your store, type a realistic task such as "Find a blue medium t-shirt and add it to the cart" (or start from one of the Pick, Buy and Rely examples) and send it. The run's page shows the chat as the buyer sees it.
- Read the run — on the run's page, the Trace tab lists every tool call, the arguments the agent chose and what your store returned. The Timeline tab shows the turns in order. Look for where the agent hesitated, retried or gave up.
- Iterate — fix what you find, retry the affected tools in Inspect, then start another run to confirm the fix. Repeat until runs reach checkout reliably.
- Check conformance — review your schema quality grade, and read Identity enforcement to check who is calling your store and whether your identity posture behaves the way you intend.
Using Inspect
Inspect shows what an agent sees when it connects to your store: every tool's input schema, description and parameter rules, and the exact request and response for each call you send. If your schema says a parameter is optional but your endpoint fails without it, you find out here before agents do.
Reading how a run ended
Each run ends at the furthest point it reached: Searched only, Built a cart, Reached checkout, Tried to place the order, Order placed or Failed. When it stops short, the run says whose fault it was: the store, the agent, ours, or none. If runs keep stopping at the same point, look at your schema or implementation for that step:
- Runs stop after search — your search results probably lack the IDs or URLs needed to look up a product. The agent cannot go further without them.
- Runs fail at the cart — check that
create_cartsays what format variant IDs and quantities take. IDs that differ between search results and cart inputs are the most common cause. - Runs stall at checkout — checkout tools often need buyer details, a shipping address or a payment token. Make sure your tool descriptions say what is required and in what format.
Testing your website (WebMCP)
If your storefront registers WebMCP tools on its own page (through document.modelContext), you can test those too. WebMCP is separate from UCP: it is your page's own tools, not a UCP transport.
- In Inspect — choose WebMCP under Calls go over to list the tools your page registers. The public Inspector's WebMCP view adds findings on each tool and a Page tools vs UCP tools table that matches them to your UCP tools by job.
- In a run — signed in, set How the agent reaches the store to Website. The agent runs one shopping task through your page's tools in a browser on our side, shown in The agent's browser pane. It stops at your checkout and never pays.
This shows whether an agent can get from your page to your checkout using only the tools you register, and what a shopper would have seen on the way. The run keeps screenshots of the page it opened, of each step that changed what a shopper sees, and of the checkout it reached. Run the same task over UCP to see where the two routes differ. See WebMCP: the store page for limits and details.
Test across models
Models read tool schemas differently. Run the same task with two or three models, using the Model pill. If only one model fails at a step, the problem is likely that model. If every model fails at the same point, the problem is in your store's schema or implementation.
Compare with a test store
Our test stores answer every tool end to end with test payments only: flowers.ucpsandbox.com (Retail), stays.ucpsandbox.com (Stays) and food.ucpsandbox.com (Food). Scenario stores such as flowers-decline.ucpsandbox.com show a specific case, like a declined payment. Running the same task on your store and on a test store shows which differences come from your store.
Use Share on a run's page to send it to your engineering team. The shared run keeps the full tool call history with request and response payloads, so developers can diagnose a problem without reproducing it.
For how stores are scored for UCP compliance, see UCP Checker, which grades stores and publishes results on its leaderboard.
Something wrong on this page?