AI Shopping Attribution: Measuring AI-Referred Commerce

The input to AI shopping attribution may be a referring domain, a campaign parameter, or another merchant-controlled identifier. The outputs are classified visits, sessions, checkout events, and purchases that meet the chosen attribution rules.

The report must declare the attribution window and treatment of unknown sources. Purchases without a supported signal stay unattributed rather than being assigned to AI by assumption.

A practical attribution pipeline has five explicit steps:

1. Classify the entry signal. Record the referrer, campaign parameters, landing URL, timestamp, and available platform identifier.
2. Persist session context. Carry the classified source through pageviews and checkout without replacing a stronger known source with an unknown one.
3. Collect conversion events. Record checkout and order events with the minimum identifiers needed for an approved join.
4. Apply the model. Declare the attribution window, eligible event sequence, cross-device limits, and treatment of direct returns.
5. Deduplicate. Prevent one order from being counted more than once across browser, server, retry, or platform events.

Reports should separate unattributed orders from zero-value outcomes. An order without a detectable signal is unknown for this model; it is not evidence that AI played no role.

Paz uses attribution as the final measurement stage after visibility monitoring, catalog diagnosis, reviewable optimization, and supported distribution. The reporting layer connects supported AI-origin signals with merchant events where the required data is available.

Teams should read attribution beside the underlying visibility evidence: configured query, engine, run time, appearance type, and observed position. A before-and-after change is still directional evidence unless the intervention, baseline, comparison group, and post-change window were designed to support a causal conclusion.