Attribution answers 'which touchpoint preceded this conversion?' Incrementality answers a fundamentally different and more useful question: 'would this conversion have happened without the ad?'
These diverge most sharply in remarketing and brand search — the two places where reported ROAS is usually highest. Someone searching your brand name was already going to find you; the ad often just intercepts a click you'd have received for free.
Measuring it requires an experiment rather than a dashboard. You need a group that doesn't see your ads and a comparison of outcomes against a group that does.
Why Incrementality matters
Incrementality frequently reverses the conclusions of attribution reporting. Campaigns showing the strongest attributed ROAS are often the least incremental, meaning budget flows toward the activity doing the least actual work.
Running a geo holdout
Split comparable markets into test and control. Run normal spend in test markets and pause entirely in control. Measure the difference in total revenue, not attributed revenue. If revenue in control markets barely moves, the channel was largely harvesting demand that existed anyway. Run for at least four weeks to clear noise and delayed conversions.
Common mistakes
Running tests too short
Purchase cycles and delayed conversions mean a two-week test measures mostly noise. Four to six weeks minimum, longer for considered purchases.
Choosing non-comparable markets
Test and control regions need similar seasonality, demographics, and baseline performance. Mismatched markets produce confident, wrong answers.
Testing during unusual periods
A holdout across Black Friday or a major product launch measures the event, not the channel.
Never testing brand search
It's the most commonly over-credited channel and the easiest to test. Most brands avoid it because they suspect what they'll find.
Where we work on this