Last-click credit is not a decision system
Last-click reporting can answer which touchpoint appeared immediately before an order. It cannot, by itself, explain why a shopper converted, whether a recommendation or offer changed the outcome, or what the merchant should do next.
Commerce teams need a decision record that connects the shopper context available at the time, the candidates considered, the action actually served, and the downstream outcome. Without that chain, attribution becomes a dashboard label rather than an operating tool.
The useful question is not only who gets credit. It is which governed decision should be repeated, changed, paused, or tested next.
Keep four kinds of evidence separate
A trustworthy revenue model does not collapse every number into lift. It distinguishes what was directly observed from what was associated, proven, or projected.
- Observed evidence: views, searches, carts, checkouts, purchases, returns, support events, and provider delivery records.
- Attributed outcomes: revenue or margin associated with an exposed recommendation, approved intervention, or provider action.
- Experiment-backed incrementality: the difference supported by a valid control or holdout design after exposure and guardrail checks.
- Modeled opportunity: a bounded estimate of what an unresolved problem may be worth, clearly labeled as a projection rather than realized revenue.
Build a tenant-scoped decision ledger
Every decision should have a durable identity, tenant, session or governed profile reference, placement, timestamp, policy version, candidate source, selected action, and execution result. The ledger should also preserve whether the shopper was eligible, whether consent permitted the processing purpose, and whether the experience was actually rendered.
Provider data can enrich the decision, but it does not replace the commerce source of truth. Shopify remains authoritative for product identity, variant, price, inventory, order, refund, and return facts. Klaviyo, support, loyalty, subscription, CRM, analytics, and recommendation providers contribute bounded context with provenance and freshness.
Measure exposure before outcome
Assignment is not exposure. A shopper belongs in an experiment result only after the intended experience was actually delivered under the configured eligibility rules. Logging a treatment assignment when the widget never rendered or the offer failed to publish biases the result before analysis begins.
The same discipline applies to recommendations and chat. Record the candidate source, final rank, impression, click, cart, checkout, purchase, and return events as separate facts. Keep delivery failures, stale inventory, and invalid product hydration visible instead of silently counting them as no-conversion outcomes.
Report economics, not only revenue
A conversion increase can still be a poor business result when discount cost, lower-margin product mix, refunds, or operational cost erase the gain. Decision-grade reporting therefore pairs conversion and revenue with average order value, gross margin where supplied, offer cost, return-adjusted outcomes, and platform cost.
When complete margin inputs are unavailable, the system should say so. A partial estimate is not improved by presenting it with more precision.
- Show influenced revenue separately from experiment-backed incremental revenue.
- Keep modeled opportunity separate from realized outcomes.
- Expose freshness, sample size, guardrail breaches, and delivery gaps alongside the headline metric.
- Retain the action, approval, execution, pause, and rollback history needed to reproduce the result.
Turn measurement into the next action
A useful report ends with an operating decision. Continue a proven treatment, collect more evidence, narrow an audience, repair a delivery path, pause a harmful action, or convert a repeated result into a merchant-approved rule.
That is the difference between attribution software and a continuous revenue improvement system: measurement closes the loop back into governed execution rather than ending in a slide deck.