What is incrementality?
Incrementality is the portion of a business result that a marketing activity actually caused, as opposed to the portion that would have occurred without it. If a campaign is credited with 1,000 orders but 600 of those customers would have bought anyway, only 400 orders are incremental. Incrementality is measured by running a controlled comparison: one group is exposed to the marketing, a comparable group is deliberately held back, and the difference between the two is the lift.
That definition puts incrementality in a different category from attribution. Attribution is an accounting exercise. It takes the conversions that happened, looks at the touchpoints in each path, and divides credit among them according to a rule. Every model, from last-click through to multi-touch, distributes 100% of the credit no matter how much of the outcome the marketing genuinely caused. Incrementality tests the assumption underneath all of them.
Why the distinction costs real money
The gap between attributed and incremental results is largest exactly where spend is largest. Branded search, retargeting, and prospecting audiences that lean heavily on existing customers all look excellent in an attribution report, because the people in them were already close to buying. They were going to convert; the ad was simply the last thing they touched on the way. Reallocating budget toward those campaigns because the attributed ROAS is high is one of the most common and most expensive mistakes in performance marketing.
The reverse error is just as costly. Upper-funnel activity that genuinely creates demand often shows a weak attributed return, because the conversion happens weeks later through a different channel. Cutting it improves the report and damages the business. This is the whole subject of reported versus incremental ROAS.
How incrementality is measured
Every credible method rests on the same idea: create a group that does not see the marketing, then compare.
- Holdout tests. A randomly selected share of the addressable audience is excluded from a campaign for the duration of the test. The difference in conversion rate between the exposed and held-back groups is the lift. This is the cleanest design when the platform supports suppression at the user level.
- Geo experiments. Whole regions are used as test and control instead of individual users. Spend is turned up, down, or off in the test regions while the control regions run normally. Geo designs sidestep user-level tracking entirely, which is why they have become more common as identifiers have degraded.
- Ghost ads and public-service-announcement tests. The control group is served an unrelated ad, so both groups pass through the same auction and targeting machinery. This controls for the selection effect that plagues simple on-off comparisons.
- Time-based switchbacks. Spend is alternated on and off in short blocks. Weaker than a randomized design, but workable where geography and user-level suppression are both unavailable.
Most large ad platforms offer their own lift-study products. They are useful, and it is also worth remembering that the platform running the test is the party being graded by it. Keeping an independent read alongside the platform’s own is ordinary measurement hygiene, not distrust for its own sake.
Incrementality, attribution, and marketing mix modelling
These three are complements, not competitors, and mature measurement programs run all of them:
- Attribution is granular and fast. It tells you which campaign, ad, and keyword were involved, at a level of detail you can act on daily.
- Marketing mix modelling is aggregate and privacy-durable. It estimates channel contribution from spend and revenue history, including offline and brand activity that no tag can see.
- Incrementality tests are the calibration layer. They produce the causal ground truth that tells you how far to trust the other two, channel by channel.
A practical rhythm is to steer daily and weekly decisions with attribution, set the quarterly channel mix with a mix model, and run a small number of well-designed lift tests each year on the channels where the most money is at stake. For a worked walkthrough of designing one, see the guide to incrementality testing.
What makes an incrementality test trustworthy
Three things separate a test worth acting on from an expensive coincidence:
- Random assignment. Test and control groups must be comparable before the campaign starts. Comparing regions you already knew were different, or users the platform chose to serve, measures selection rather than lift.
- Enough signal to detect the effect. A test that cannot statistically distinguish a 10% lift from zero will report “no lift” for a campaign that is working fine. Sizing the test to the effect you care about is the step teams skip most often.
- A long enough window. The measurement period has to cover the real purchase cycle. A one-week read on a product people consider for a month measures the tail of demand, not the whole of it.
Incrementality is more work than reading an attribution dashboard, and it is the only measurement approach that answers the question a finance team actually asks: what did this spend change?