Incrementality testing measures the additional business outcome caused by a marketing activity, as distinct from the outcome merely associated with it. The method is experimental: a randomly selected group is withheld from exposure, and the difference in outcomes between the exposed and withheld groups is the incremental effect. Everything else, including platform-reported conversions and attribution models, describes correlation between exposure and outcome rather than causation.
The problem it solves is that attribution systematically over-credits activity aimed at people who were already going to convert. Branded search advertising is the standard illustration: it reports strong performance because people searching for a brand name have high intent, but a portion of those clicks would have arrived through the organic result at no cost. The same logic applies to retargeting, to prospecting audiences built from existing customers, and to any channel that finds people close to purchase. Reported return on ad spend can be excellent while incremental return is close to zero.
Several designs are used depending on what can be randomized. Geographic holdouts split markets into matched regions, running activity in some and not others, which suits channels that cannot be targeted at the individual level. Audience-level holdouts withhold a random percentage of an addressable audience, which suits email, retargeting, and platforms supporting suppression lists. Platform conversion lift studies randomize at the auction level, though they depend on the platform to run and report their own examination. Switchback designs alternate exposure over time within one market, which suits activities that cannot be split by audience.
Design details determine whether the result is trustworthy. Contamination, where the holdout group is reached anyway through another channel or through spillover between regions, biases the measurement toward zero. Insufficient scale produces intervals so wide that the test cannot distinguish a substantial effect from none, and incrementality tests generally need more volume than teams expect because the effects being measured are smaller than the correlational figures suggest. The measurement window must be long enough to capture delayed conversion, and short-window tests systematically understate effects on considered purchases.
The organizational difficulty is greater than the methodological one. Incrementality tests frequently show that channels with excellent reported performance contribute far less than believed, which threatens budgets, agency relationships, and the internal reputation of whoever built the case for the spend. Committing to the test before seeing the result, agreeing in advance what decision each outcome will trigger, and running the analysis with people who have no stake in the answer are what prevent the finding from being negotiated away afterwards.
Test design should also account for the fact that incrementality varies by audience and context rather than being a single property of a channel. Advertising that is largely redundant for existing customers who would return anyway may be genuinely incremental for prospects who have never encountered the brand, and a single blended result conceals both. Where volume permits, running the holdout within defined segments produces far more actionable findings than a channel-level average, because the resulting decision is usually not whether to stop a channel but which audiences within it to keep paying for. This is also why results from one period should be re-established periodically, since incrementality shifts as brand awareness, competitive activity, and audience saturation change.
Because the results directly determine where budget should sit, this work belongs at the intersection of measurement and planning. The experimental design and analysis are normally handled by data analytics, the execution by marketing services, and the reallocation decisions by growth management, since the value of the test is realized only when spend actually moves in response to what it found.