This is a test store that anyone's AI agent can shop. These figures come from the store's own request log for the last 30 days. They are shares and timings only: no agent is named unless it proved who it is, and no search text is shown.
Of every session that wasn't a browser. An agent can name itself with a UCP profile or a bot user agent; it proves it by signing its requests with keys it publishes.
Browsers are counted on their own in the grid: a browser rarely says what it is. How we decide each level
Share of all requests, by door and by how much the sender proved. Darker means more. Discovery files are counted inside the door they were read from. Crawler addresses are checked from 31 Oct 2026; crawlers before that count as declared.
| Door | Verifiedsignature checked out | IP-checkedcrawler from its operator's ranges | Declared, unprovennamed itself, no valid signature | Browserperson or agent: owner console | Scriptcurl, python, no user agent | All |
|---|---|---|---|---|---|---|
| Storefront pages | 9% | 10% | 17%<1% signature failed <1% claims to be a crawler | 33% | 3% | 73% |
| UCP over REST | 4% | <1% | 3%<1% signature failed <1% claims to be a crawler | 4% | 5% | 18% |
| UCP over MCP | 6% | — | 2%<1% signature failed | <1% | <1% | 9% |
| All doors | 20% | 11% | 23% | 38% | 9% | 100% |
Share of reads of the store's discovery files
/robots.txt | 43% |
|---|---|
/.well-known/ucp the UCP profile | 31% |
/llms.txt | 13% |
/mcp tool list, read before a call | 7% |
/ucp/openapi.json | 5% |
| A2A agent card | 1% |
What the store's UCP profile offers, and the share of agent sessions that used each
| Transport | In the UCP profile | Used by | Paid |
|---|---|---|---|
| UCP over REST | Yes | 21% | 1% |
| UCP over MCP | Yes | 4% | 9% |
| WebMCP in the page | In the pages | — | — |
The rest of sessions used storefront pages only (76%). A transport can be offered and still hardly used; that is a fact about agents today, not a fault in the store.
The latest paid journey from a verified agent. Requests from the same agent are joined into one journey when they share a signer or a UCP profile, across every door.
ucpplayground.com Verified · UCP request signature
UCP over MCP
initialize · tools/list · notifications/initialized
33 ms
search_properties results found
574 ms
create_booking_session ready to pay
58 ms
complete_booking_session · paid test order
94 ms
Share of sessions. An agent is named only when it signed its requests with keys it publishes and the signature checked out, and it visited at least 3 times. Everyone else is grouped.
| Browsers | 44% |
|---|---|
| Search crawlers and bots | 36% |
| Scripts and tools | 10% |
ucpchecker.com | 6% |
ucpplayground.com | 1% |
| Declared a profile, not verified | 1% |
river.ai | 1% |
| Other verified agents | 0% |
Of the sessions from agents (not browsers or crawlers)
The store answers every agent, signed or not. Checking signatures began on 3 Oct 2026.
Share of sessions reaching each step
| Sessions | 100% |
|---|---|
| Searched dates | 4% |
| Viewed a room | 19% |
| Started a booking | 1% |
| Ready to pay | 0% |
| Booked | 0% |
Every session ends in exactly one of these
| Never searched | 78% |
|---|---|
| Viewed, did not add | 18% |
| Searched, picked nothing | 3% |
| Checkout started, not finished | 0% |
| Paid | 0% |
| Ready to pay, then left | 0% |
| Search found nothing | 0% |
Share of sessions by door, and the share of each that paid
| Storefront pages | 76% 0% paid |
|---|---|
| UCP over REST | 21% 1% paid |
| UCP over MCP | 3% 4% paid |
Which result agents go on to open or add
Too few picks to show a split yet.
By operation, for operations called often enough to show. 94.2% of all calls succeeded. A call fails when the store answers with an HTTP error or a protocol error.
| Operation | Succeeded | Typical | Slowest 5% |
|---|---|---|---|
GET /.well-known/ucp | 100% | 21 ms | 467 ms |
tools/list | 100% | 189 ms | 614 ms |
GET /ucp/openapi.json | 100% | 15 ms | 245 ms |
initialize | 100% | 191 ms | 258 ms |
search_properties | 100% | 339 ms | 574 ms |
notifications/initialized | 100% | 209 ms | 251 ms |
GET /ucp/search | 0% | 51 ms | 226 ms |
POST /ucp/booking-sessions | 75% | 77 ms | 112 ms |
Share of the error answers the store gave, by code
invalid_profile_urlThe agent left out its profile, or it could not be read42%missingA required checkout field was not sent22%eligibility_invalidThe store answered with this error code14%not_foundThe agent asked for an id the store does not have11%payment_failedThe test payment was declined6%invalid_datesThe store answered with this error code6%Things a store can't see on a normal analytics page
A session is one agent's requests with no gap longer than 30 minutes, joined across doors by a signer, a UCP profile, or a hash of network and user agent whose salt changes every day. Sessions under a scenario host (for example stays-slow) are included. How this page decides who arrived. Run your own version of this on a private copy of the store: how the sandbox works.