The report is graded by the thing being tested
Platform-reported ROAS is the most confidently wrong number in marketing, and it isn’t fraud — it’s structural. Each platform can only observe the conversions it touched, so it reports those and stays silent about the counterfactual. Run the simple exercise of summing reported revenue across every channel and comparing it to the P&L. The gap is the double-counting baked into every optimization decision you’ve been making.
The damage isn’t academic. Budget flows toward whichever channel is best at observing conversions — typically branded search and retargeting, which mostly intercept demand that already exists — and away from the channels that created the demand. The dashboard improves. The business doesn’t. Measurement is the discipline that breaks that loop.
The three tools, and what each is actually for
Teams argue about attribution versus incrementality versus media mix modeling as if they had to pick one. They answer different questions and work best stacked.
Attribution is a fast, directional read for steering in-flight. It’s fine for “which ad is fatiguing this week” and dangerous as the basis for a budget decision, because it inherits every observation bias the platforms have.
Incrementality testing is causal truth about one channel over one window. A holdout, a geo test, a matched-market design, an audience suppression — you build the comparison the platform will never hand you and read the lift. It’s the ground truth everything else gets calibrated against.
Media mix modeling is the top-down view across the whole plan, including the channels that can’t be clicked. It’s the only way to allocate across CTV, video, audio, and out-of-home alongside the trackable channels without pretending the un-clickable ones don’t exist.
The structure that works: use tests to calibrate the model, use the model to allocate, and keep attribution in its lane as a daily steering instrument.
Reported ROAS answers “what did the platform observe?” Incrementality answers “what did the spend cause?” Those are different questions, and only one of them is safe to put in front of a CFO.
Measurement is the referee for the intent argument
The Intent lane makes the case that demand generation and demand capture are one system, and that capture channels quietly take credit for demand generation created. Measurement is what settles it. A holdout on branded search or retargeting routinely shows that a large share of that reported revenue would have arrived anyway — the channel was standing where demand was already going to land. That’s not an argument you can win with a dashboard; it’s an experiment. Incrementality is the only method that separates the spend that created demand from the spend that merely observed it, which is why the whole full-funnel case ultimately rests here.
What I do
Establish causality before allocating. Holdouts and geo tests are cheap relative to the budgets they govern. A test that costs a fraction of a quarter’s spend and reveals a channel is half as productive as reported pays for itself immediately.
Allocate on contribution, not revenue. Revenue-based ROAS targets quietly push spend toward high-revenue, low-margin products. Allocating on contribution margin changes which campaigns look like winners — sometimes dramatically.
Model the whole mix, including what can’t be clicked. CTV, video, audio, and out-of-home do real work that click-based systems can’t see. Modeling them alongside the trackable channels is the only way to stop underfunding the top of the funnel by default.
Keep every read executable. Measurement that doesn’t change a buying decision is a research project. Every read should terminate in an action: shift this budget, kill this line item, extend this test, raise this bid floor.
How I use AI here without handing over the keys
I build agents to do the labor — pulling and normalizing reporting, flagging anomalies, running forecast scenarios, preparing modeling inputs — each with guardrails and defined parameters. What they don’t do is decide. There’s always a human on top with the context to catch drift before a dashboard reports it. Measurement is where a wrong automated call compounds fastest, so it’s the last place I’d remove the person.
Further reading
- How to run a geo holdout test without fooling yourself
- Incremental ROAS: the only return number a CFO will actually trust
- Attribution, incrementality, MMM: stop asking which one is right
- Is branded search incremental? Run the holdout you’re afraid of
- Platform ROAS is grading its own homework
- The saturation point: finding where the next dollar stops working
- A real lift or just noise? The testing discipline most programs skip
Where this fits
- Intent → — how spend gets organized across levels of buyer intent, and the demand-gen vs demand-capture split this lane referees
- ← Back to Paid Media