Research / Pilot brief
Independent evidence for AI commerce decisions.
A focused research pilot to reconcile permissioned agent events with real merchant orders.
Led by Sean Bangalore. This initiative is in customer discovery. The audit workflow below is proposed; it is not a claim of completed customer work or a production analytics platform.
The thesis
Independent reconciliation may help merchants and agencies make better measurement and commission decisions, if authorized agent-side events provide useful evidence beyond existing reports. Independence alone is not enough: the evidence must change a real decision.
Who we want to learn from
- Merchant and agency measurement owners dealing with conflicting reports.
- Agent and commerce platforms able to provide authorized upstream event samples.
- Affiliate networks investigating a specific disputed transaction or commission allocation.
What we would compare
| Evidence | Potential use | Limit |
|---|---|---|
| Merchant orders and sessions | Verify transactions, refunds, and recorded journeys | Cannot reveal an outside agent’s unseen recommendations |
| Authorized agent-side events | Inspect recorded referrals or assistance | Coverage depends on participation and allowed uses |
| Existing measurement reports | Establish a baseline and inspect disagreements | Different attribution rules may explain apparent discrepancies |
A proposed pilot in three stages
- Access first. Identify one real, authorized upstream sample and the buyer decision it could inform. Confirm permitted fields and uses before exchanging records.
- Reconcile a limited sample. Compare the event chain with orders and existing reports. Review duplicates, refunds, and ordinary missing-referral explanations.
- Decide whether to expand. Only broaden the pilot if access works, the evidence adds information, and a buyer can act on it. Agree on any audit fee and scope in advance.
The proposed output
An auditable discrepancy ledger: the source behind each link, what can be observed, what is modeled, what remains unknown, and whether the difference is meaningful for the pre-agreed decision. Observed assistance, modeled attribution, and causal incremental lift stay separate.
A matched order does not demonstrate that an agent caused a purchase. Incremental lift requires a separate experiment with an appropriate design and sample size.
How this fits your current stack
The research starts with your existing merchant analytics, attribution tools, and surveys as the baseline. We are testing for useful additional evidence, not assuming every reporting gap requires another dashboard. Shopify and GA4 integrations are potential future work, not currently available product features.
Data and participation
No customer records are needed for the introductory call. Before any sample is shared, we would agree on permissions, purpose, minimum necessary fields, access, retention, and deletion. Do not send credentials or customer data through booking notes.
The 20-minute research call is free. A later paid audit would require a separately agreed scope and fee. There are no guaranteed findings or outcomes.
Start with one question
What was your last disputed attribution decision? Which records did you have, which were missing, and what would stronger evidence change?
Join the measurement pilot