Incrementality Testing

Incrementality testing asks how many conversions happened because of your advertising, as opposed to how many would have happened anyway. Google's Conversion Lift page calls incrementality experiments the way to measure the causal impact of ads, and describes lift as the increase in conversions caused by the presence of the ad. It is not a single product but a family of designs: user-based lift studies, geo experiments, and holdouts. All share one idea: compare a group that saw ads with a comparable group that did not, and treat the difference as the ads' effect. Google's lift studies are split automatically into a treatment group shown your ads and a control group that is not, so you do not need to build the groups yourself.

Standard attribution, such as last click or data-driven attribution, only distributes credit among touchpoints that happened before a conversion. It cannot show what would have occurred without those touchpoints. An incrementality test creates that missing comparison. Google's lift studies split the audience automatically into a treatment group that is shown ads and a control group that is not, and the difference in outcomes gives the lift. Randomization is what makes the result causal, because the groups differ only in ad exposure. The output is incremental conversions, incremental cost per action, and incremental return on ad spend, where values are available. Platform lift studies are not available in every account, so the simpler fallback is a holdout or regional on and off test.

The test interacts with the choice of channel and the scale of the question. Brand search is the classic target, since people searching your name often buy regardless. Retargeting is another. Prospecting campaigns usually show higher incrementality, because they reach people who did not know you. You can run tests inside a platform, through a lift study, or independently, by turning a channel off in randomized regions. The size and duration of the test matter: lifts are often small, and the test needs enough conversions in both groups to detect them. Results apply to the period, audience, and spend level tested. Check whether the channel is bidding on people who already know you, since those are the conversions most likely to happen anyway.

The most common mistake is running a test that is too small or too short, then reading an inconclusive result as no effect. A second mistake is changing something else during the test, such as a price promotion, which affects both groups unevenly if exposure and promotion overlap. A third is generalizing too far: a lift test on brand search in one quarter says little about a different season. A fourth is treating the platform's own lift study as a neutral audit; the platform runs it and reports it. Results are also estimates with uncertainty, so report a range, not a single number. Report results as a range with a plain statement of what you did, because a lift test estimate carries uncertainty.

Pick one channel where you suspect reported results are inflated. Define the conversion, the test duration, and the lowest lift that would justify the spend. Use a platform lift study if you have access, or a holdout or geo split if you do not. After the test, compare incremental cost per conversion with reported cost per conversion. Divide one by the other to get a correction factor, then apply it when setting targets for that channel. Repeat the test once or twice a year, because competitive conditions, creative, and audiences change. Repeat the test after major changes to creative, audience, or budget, since an old result may not hold.