The 1,050 attributed versus 600 actual is the exact failure mode, and the Christmas tree framing is the right one: brands try to solve at the wrong level. The part I would add is that the star metric only works if the definition of "new customer" is the same in every platform and in your own store. A big chunk of the double-counting is not the model, it is that each platform's default window counts a returning buyer as new. Fix the definition and the blended number stops contradicting itself. Reading new-customer revenue and spend by channel against one agreed definition, without exporting anything, is what nomu does off a Meta or Google Ads account.
The 1,050 attributed conversions against 600 real orders is the whole attribution problem in one line: every platform grades its own homework and gives itself an A. The trap is reaching for a single impartial model to settle it, because the models that aggregate every signal are still black boxes and still wrong in ways nobody can inspect. What reconciles is triangulation — platform numbers for direction, a holdout or geo test for incrementality, and the backend for truth. Any one of those alone will lie to you confidently.
The line about every platform grading its own homework is the whole thing, and the 1,050 attributed conversions against 600 real orders is the number I'd put in front of the client, not the platform reports. One addition from the media buying side: the gap is usually biggest on Meta, because view-through and click-through both get counted, while Google is closer to last-click reality, so comparing the two platforms' reported CPA without adjusting for that makes Meta look better than it is and Google worse. Keeping the two platforms as separate reported columns and reconciling against the store's order count is what I'd suggest before any geo-lift work, and it's what I ask Analytics Copilot for when I need the per-campaign numbers without opening four dashboards.
The 1,050 vs 600 example is the clearest way I have seen to explain why platform dashboards cannot be summed. For brands below the spend where geo-lift tests get clean reads, where do you draw the line? Is there a monthly budget threshold where you would tell a team to stay on blended MER and holdouts until a test can actually reach significance?
The 1,050 attributed versus 600 actual is the exact failure mode, and the Christmas tree framing is the right one: brands try to solve at the wrong level. The part I would add is that the star metric only works if the definition of "new customer" is the same in every platform and in your own store. A big chunk of the double-counting is not the model, it is that each platform's default window counts a returning buyer as new. Fix the definition and the blended number stops contradicting itself. Reading new-customer revenue and spend by channel against one agreed definition, without exporting anything, is what nomu does off a Meta or Google Ads account.
The 1,050 attributed conversions against 600 real orders is the whole attribution problem in one line: every platform grades its own homework and gives itself an A. The trap is reaching for a single impartial model to settle it, because the models that aggregate every signal are still black boxes and still wrong in ways nobody can inspect. What reconciles is triangulation — platform numbers for direction, a holdout or geo test for incrementality, and the backend for truth. Any one of those alone will lie to you confidently.
The line about every platform grading its own homework is the whole thing, and the 1,050 attributed conversions against 600 real orders is the number I'd put in front of the client, not the platform reports. One addition from the media buying side: the gap is usually biggest on Meta, because view-through and click-through both get counted, while Google is closer to last-click reality, so comparing the two platforms' reported CPA without adjusting for that makes Meta look better than it is and Google worse. Keeping the two platforms as separate reported columns and reconciling against the store's order count is what I'd suggest before any geo-lift work, and it's what I ask Analytics Copilot for when I need the per-campaign numbers without opening four dashboards.
The 1,050 vs 600 example is the clearest way I have seen to explain why platform dashboards cannot be summed. For brands below the spend where geo-lift tests get clean reads, where do you draw the line? Is there a monthly budget threshold where you would tell a team to stay on blended MER and holdouts until a test can actually reach significance?