Discrepancy Investigation Across Ad Platforms and GA4

Meta reports purchases, GA4 reports purchases, and finance reports revenue, and the spread between them runs 20 to 60 percent on an account nobody thinks is broken. In every investigation I have run, that spread decomposed into fewer than nine causes. The deliverable is a ledger where each cause carries a euro figure and a percentage, and the residual is stated rather than absorbed. Method: Raw data only. The GA4 BigQuery export at event level, each advertising platform's API at day and campaign level, and the order table from Shopify, the ERP, the PMS or the TMS. Nothing from a screenshot of a reporting interface, because interfaces apply filters that nobody remembers setting. Joins run on order id where the id survives the journey, and on timestamp and value with a stated tolerance where it does not, with the match rate reported rather than hidden. What it usually turns out to be: Attribution window and model mismatch. Seven-day click plus one-day view set against last non-direct is not a discrepancy, it is two different questions producing two correct answers. Ad account timezone against the fiscal day boundary in the ledger. A five-hour offset relocates an entire day of a month-end spike into the following period. Currency. Platform daily rate against settlement rate, and the date the conversion is booked rather than the date it is paid. Event id deduplication failure between pixel and conversion API, where I have measured duplicate rates from 5 to 30 percent on accounts that had passed a vendor health check. Consent-mode modelled conversions being counted alongside observed ones with no flag distinguishing them. Cross-domain tagging lost on a payment gateway hop, which converts paid sessions into direct at the exact moment the money arrives. Refunds and cancellations never written back to the analytics side. Bot and scraper traffic firing purchase events. And the one that hides longest: a QA suite in the deployment pipeline placing test orders that reach production tags every time the team ships. The target: 95 percent of the gap explained and assigned to a named cause, with the residual sized and left as residual. I do not make numbers match. Two systems measuring different things agreeing exactly is evidence of a filter, not of health, and I have found precisely that filter in an account whose owner believed it was reconciled. What you receive: A variance ledger, one row per cause with its contribution in euros and in percentage points. The queries that produced it, so your analyst re-runs the investigation next quarter without me. A fix list ordered by the size of the error rather than the ease of the fix, which is the ordering that gets argued with and the one worth defending. And a single page naming which number to use for which decision, because the answer is usually more than one number and the honest version says so out loud. Not included: Implementation of the fixes is separate at the entry tier and included above it. No handling of platform support cases on your behalf. No opinion on your media plan; this engagement establishes whether the instruments are telling the truth, and what you do afterwards is a different conversation. Who this is not for: A team that already knows what the answer should be. This work regularly finds that the channel someone spent a quarter defending is fine and the reporting was broken, and just as often the reverse. Both outcomes get written down with the same wording, and the person who commissioned the investigation does not get to review that section first.