Why platform-reported ROAS over-credits your paid channels

Open any ad platform and it will tell you it is working. Meta reports a return on ad spend. Google reports one. TikTok reports one. Add them up and the picture looks excellent. The trouble is that the numbers describe the same revenue, claimed several times over, by systems that each have a reason to claim it.

This is not fraud. It is the predictable result of how attribution is built.

Each platform counts what it can see

An ad platform measures conversions through its own pixel, inside its own attribution window, crediting clicks and often views that touched a user before they bought. Every platform does this independently. None of them subtracts the influence of the others. So when a customer sees a Meta ad on Monday, searches your brand on Google on Wednesday, and buys on Friday, both platforms can record the sale as theirs.

Sum the platform numbers and you are double counting by design. The more channels you run, the wider the gap between what the platforms claim and what actually happened.

Reported is not the same as incremental

The deeper problem is not arithmetic. It is causation. Platform ROAS answers the question “how many buyers had we touched?” The question that matters for a budget decision is different: “how many of those buyers would have bought anyway?”

Brand search is the clearest example. A customer who already intends to buy will often search your name and click whatever appears first. If that is a paid ad, the platform books a conversion at an excellent ROAS. Remove the ad and most of those sales still happen through the organic result below it. The reported return was real. The incremental return was close to zero.

Retargeting has the same shape. It is cheap, it looks efficient, and a large share of the people it reaches were already on their way back.

Why a single vendor does not fix it

The usual response is to add an attribution tool. Triple Whale, Northbeam, and the rest exist precisely because platform numbers cannot be trusted on their own. They help. But each one is a model, and every model encodes assumptions: a particular attribution logic, a way of handling view-through, a default window. Those assumptions are choices, and different vendors make different ones.

So a single vendor moves you from one biased number to another biased number. It is usually a better number. It is still one point of view, presented as if it were the answer.

Triangulation: agree, and disagree, on purpose

The way out is not to find the one true measurement. It is to run several independent ones and read them against each other.

In practice that means three things working together. A media mix model, or ideally a few of them built on different methods, estimates the contribution of each channel from spend and outcome data without relying on any pixel. Geo experiments and holdouts test specific claims directly, by turning spend down in some regions and watching what happens. And the vendor tools stay in the mix as a third lens.

When these methods agree, you can act with confidence. When they disagree, the disagreement is the finding. It tells you exactly where the reported number is hiding something, and that is usually where the reallocatable budget is.

What to do this week

You do not need a full measurement rebuild to start seeing the gap.

Stop summing platform-reported ROAS across channels. Track a single blended efficiency number against total revenue, and watch that instead. Then pick your largest “obviously great” line, often brand search or retargeting, and run one honest holdout.¹ Turn it down somewhere measurable and see how much revenue actually moves. The result is frequently uncomfortable, and it is the most valuable thing you will learn all quarter.

Measurement is not about producing a prettier dashboard. It is about knowing which of your good numbers are real, so the next dollar goes where it actually compounds.

Footnotes

1. A holdout suppresses a channel for a randomly chosen share of your audience or geography, then compares the two groups. It is the closest thing to a controlled experiment most brands can run without specialist tooling. ↩