Funnel: from agent visit to order
The four observed stages — agent visits, served machine-readable, protocol requests, orders reported — and how order attribution is labeled, never guessed.
The Funnel (a Pro surface) shows what agent traffic actually turns into, for one store over a range you pick — 7, 30, or 90 days. Every number is an observed count for that range; nothing is modeled, sampled, or extrapolated, and the page says so at the top.
The four stages
| Stage | What it counts |
|---|---|
| 1 · Agent visits | Requests the gateway classified as AI agent or crawler in the range, with the agent/crawler split shown underneath. Human traffic is excluded. |
| 2 · Served machine-readable | Requests actually answered on the markdown path — machine-readable responses served, not pages merely eligible for one. |
| 3 · Protocol requests | Requests served on the protocol path (agent-protocol endpoints like UCP or MCP). Zero until protocol adapters are live for the store. |
| 4 · Orders reported | Orders that arrived via your Shopify webhook or the outcomes API, with total order value as reported (no currency conversion). If no orders have been reported yet, this card shows exactly how to wire an outcome source. |
Attribution honesty
The orders stage never blends its numbers into one “conversions” figure. Each order is labeled with its evidence class:
- Explicitly attributed — the order reporter sent evidence (a UTM, a referrer, an event id).
- Lookback-inferred — a product URL on the order saw agent traffic in the prior 7 days. This is inference, and the label says so.
- Unattributed — no evidence either way.
Below the stages, a daily bar chart shows agent visits per UTC day with orders alongside, and a platform table breaks out visits, markdown served, and “orders with evidence” per platform. An order whose evidence names several platforms appears under each — the column is evidence presence, not exclusive credit, and the caption under the table says exactly that.