The two doors

WebMCP is a proposed web standard, running as a Chrome origin trial on Shopify storefronts until 16 November 2026. It is not part of the UCP specification.

Results

What we saw

• Both doors reached the store's checkout handoff with the right item in most runs: UCP in 107 of 120, WebMCP in 117 of 120.
• UCP was faster once the door was open (median 12.0 s against 20.5 s) and used fewer tool calls (2 against 4). The WebMCP door also has to load the store's page first: median 3.8 s, against 0.5 s to read a UCP profile.
• WebMCP used fewer model input tokens (median 28,800 against 51,400), so it cost less per task (median $0.023 against $0.028). Most of the UCP token cost is the size of the protocol's responses.
• Almost all the UCP failures came from two stores. At one, UCP checkout answered "Sold out" for items the store's own web page added to the cart without complaint. At the other, the UCP catalogue search did not return the exact item asked for, and the agent picked a similar one.
• Shopify's UCP checkout handed off with status "incomplete" and a checkout link, because no contact or address had been given. We count a store-issued checkout link as the handoff at both doors.

By store

How we ran it

• Stores: 12 US Shopify stores that serve both doors: a UCP profile (version 2026-08-25) and WebMCP tools including a checkout handoff, checked the day the study started.
• Tasks: two per store. One lets the agent choose ("buy a medium-size suitcase"); one names an exact product and option from the store's own catalogue.
• Agent: Gemini 3.6 Flash, the same prompt at both doors except where the doors differ (how the cart and checkout are reached), the same US buyer context, and the same scripted "Yes, go ahead" when the agent stops to confirm.
• Repeats: five per task and door. The two doors of a task run back to back, and which goes first alternates, so neither door always meets the store first.
• The stop: a run ends when the store hands the buyer a checkout link: a UCP continue_url, or the WebMCP page landing on the store's checkout. Tools that pay or place an order are refused before any call leaves, at both doors. No run pays; carts left at checkout are abandoned.
• The verdict: from the store's own evidence, never the agent's claim: the handoff link the store issued, and what the store's cart replies, catalogue entries or checkout page say the cart holds.
• The clock: starts once the door is open (UCP profile read, or page loaded). Opening is timed separately.

What this is not

• Not a verdict on either standard. It is one model, one platform and twelve stores, measured over one day.
• Not a completed purchase. Both doors stop where the store takes over; what happens after the handoff is outside this study.
• Not a grade of the stores. On Shopify both doors are the platform's, so differences mostly reflect the platform's two implementations and the page each store runs.