Forensic Reconciliation of Ad Platform and GA4 Numbers

Four systems measuring the same week will not produce the same number, and a finished investigation does not end with them agreeing. It ends with the variance decomposed: each cause named, quantified in your currency, and marked as either a defect to fix or a definition to accept. The usual result explains 90 to 97 percent of the gap; the remainder stays labelled unexplained, because a residual you can see is safer than one that has been folded into a plug figure. What is usually behind it: The same causes keep returning. Timezone drift — a GA4 property still set to a parent company's timezone while Yahoo! JAPAN Ads and LINE Ads report in JST, producing a one-day shift that reads like a six percent variance every Monday. Deduplication failure between the Meta pixel and the Conversions API, where event_id is absent or generated after the browser event fires. Modelled conversions in Google Ads under consent mode, read by everyone in the meeting as observed counts. View-through credited on one side and not the other. Refunds posted to the refund date in the ledger and to the order date in the platform. Currency converted at a month-end rate in one report and a transaction-date rate in the other. Two ad accounts linked to one property, both counting the same purchase. How the investigation runs: Read access first, then a single reconciliation period, normally 60 to 90 days — long enough that platform restatement has settled, short enough that a platform change has not intervened mid-window. I pull raw grain on each side: click and conversion exports per platform, the GA4 BigQuery export rather than the interface, and the order or contract table from your own systems. Joins run on order identifier where one exists and on timestamp windows where it does not. Each candidate cause is then tested by removing it and measuring how much of the gap moves, so nothing enters the report as a theory that merely sounds right. What you receive: A waterfall from platform-claimed revenue down to booked revenue, every step labelled with an amount. A written ruling on which figure belongs in the board pack and which figures are for channel operators only. A defect list kept separate from a definitions list, because the first is engineering work and the second is a decision your organisation makes once and then stops relitigating. Where a defect is fixable inside the engagement — a missing event_id, a wrong property timezone, a duplicated account link — I fix it and show the variance before and after. What I will not do: Make two platforms agree. If the success criterion is that Meta's reported revenue equals the ledger, decline this now; those figures measure different events and forcing equality means suppressing one of them. I also do not audit an agency's honesty. If the real question is whether your agency is inflating results, say so at the start — that is different work, quoted differently, and I will not conduct it under cover of a technical reconciliation. Fit: This gets bought by someone who has already lost an argument in a meeting because two decks disagreed. It earns its fee when the disputed spend is large enough that the answer moves a budget; below roughly 150 thousand dollars a quarter across the channels involved, the findings are real and nobody acts on them. Retail energy suppliers with a back-office confirmation lag, and multi-location service businesses with call-driven conversions, are where the gaps run widest and the fixes hold longest.