Why do the platforms disagree with the orders table?
Each ad platform attributes conversions it can match to its own touchpoints, under its own windows and modeling - and platforms cannot see each other. A customer who clicked two ads and a search result is a legitimate conversion to three systems. Meanwhile your analytics property applies one cross-channel model to the subset of journeys it observed after consent. Four instruments, four defensible answers, one orders table. The error is not the disagreement; it is expecting agreement.
What did the privacy changes actually break?
Device-level identity across apps and sites. When mobile platforms began requiring permission for cross-app tracking and browsers constrained third-party cookies, the platforms lost much of the deterministic joining that attribution reporting relied on - and replaced it with modeled conversions. The practical consequences: platform numbers became estimates with confidence rather than counts, view-through claims became softer still, and small segments became statistically decorative. ROAS did not stop being measurable; it stopped being measurable from a dashboard alone.
So what is ROAS tracking now?
A triangulation of three instrument classes, each trusted for what it alone can see:
| Instrument | Trust it for | Never trust it for |
|---|---|---|
| Platform reporting | Relative performance within the platform - which campaign, creative, audience | Cross-channel credit; totals that sum across platforms |
| Site analytics | Cross-channel shape of consented journeys; funnel behavior after the click | The full population - it sees only what consent and clients allow |
| The orders table + experiments | Ground truth on totals; incrementality when you can hold out geography or budget | Fine-grained credit assignment - it knows what happened, not why |
The decision rule that keeps teams honest: optimize creatives and audiences inside each platform with its own numbers; allocate budget across channels with analytics plus experiments; and report the business's health only from the orders table. Numbers cross those lanes only with a stated caveat.
Do attribution tools solve this?
Third-party attribution products - the category shoppers compare as alternatives to one another - re-model the same partial signals: pixels, platform APIs, sometimes surveys. Good ones make triangulation convenient and surface deltas quickly; none of them recover the identity signal the privacy changes removed. Evaluate them as instrumentation and workflow, not as oracles: ask what raw signals they ingest, how they handle consent states, and whether their outputs reconcile against your orders table - the same audit you would run on anything else.
What does a defensible setup look like?
Clean click ids captured and stored first-party; purchase events reconciled against orders with a known, explained gap; platform APIs fed with well-matched conversions so their optimization works with the best signal available; UTM discipline that survives redirects; and a written statement of which instrument answers which question. That last artifact costs an afternoon and prevents most attribution arguments from recurring, because the argument is usually two people reading different instruments and calling it a disagreement about reality.