One week last August, a store’s ad platform reported 79 purchases. Its own database recorded 103 orders. Neither number was wrong.

They answered two different questions everyone believed were the same one. And one of them cannot be corrected: there is no refund event, no negative value, no way to retract a purchase already reported. What you report is final the moment you send it.

Three of the six findings, all verifiable:

• A sender nobody at the store installed, copying every browser event hourly with no deduplication key, running since February, invisible in the interface they used.
• A later, better event cannot replace an earlier one. “Fire at checkout and correct it later” fails in both directions.
• The endpoint that would reveal double counting counts events before deduplication, so a store sending everything twice sees a healthy line going up.

One real store, arithmetic shown: 242 of 577 settled orders reversed, 41.9% by count and 42.2% by value. The shop counter out-earned the website 1.73 to one and the platform received an event for none of it.

It does not claim that changing what you report improves your advertising. No controlled measurement of that exists, by anyone, including me.

You get the 61-page PDF and the Revenue Signal Worksheet, Excel or Google Sheets: nine sheets turning your order table into your reversal rate, your three ledgers with each gap named, and a verdict on whether to act. Most come back DO NOT ACT YET.