Correlation versus causation in ad measurement
Every attribution model measures the same thing: did a customer interact with your ad before they purchased? If yes, the ad gets credit. This tells you nothing about whether the customer would have purchased anyway. A customer who searches for your brand name on Google, clicks a branded search ad, and converts is counted as a Google conversion. But that customer already knew your brand and was coming to buy. The Google ad was a taxi they took to a destination they were already walking toward.
Incrementality testing asks a different question: if this customer had never seen this ad, would they still have purchased? The answer requires a control group who did not see the ad. Platform attribution cannot provide this because platforms, by definition, show your ads to everyone in your target audience.
The geo holdout test
The simplest incrementality test divides your market into two groups by geography. Group A receives your normal ad campaigns. Group B receives no ads, or a holdout level of spend (typically below 10 percent of normal). You run this for two to four weeks and measure the conversion rate difference between the two groups.
The uplift, the difference in conversion rate between the treated and holdout groups, is your estimated incrementality. If Group A converts at 2.4 percent and Group B converts at 1.8 percent, your incremental lift is 0.6 percentage points, roughly 25 percent incremental. That means 75 percent of your Group A conversions would have happened anyway, with or without the ad.
Geo holdouts work best when your two groups are genuinely comparable. Match regions on population size, income distribution, historical purchase rates from your own data, and seasonality patterns. A holdout group that is systematically different from your treatment group will produce a meaningless result.
The user-based holdout test
Meta and Google both offer ghost ad experiments, where a holdout group is shown a public service announcement (PSA) or a placeholder creative instead of your actual ad. This is more precise than a geo holdout because the two groups are truly randomized at the user level, not at the geographic level.
The limitation is that you are dependent on the platform to run the holdout, which means the platform controls the methodology. Meta's Conversion Lift study is reliable but requires minimum spend thresholds and a minimum holdout size that makes it impractical for brands spending below 50,000 euros per month on Meta.
How to interpret your incrementality results
An incrementality rate of 30 percent means 30 percent of your attributed conversions are genuinely caused by your advertising. The remaining 70 percent would have converted through organic channels, direct traffic, or repeat purchase behavior regardless of whether you ran ads.
This does not mean your ads are worthless. It means you are over-crediting them. The relevant question is: what is the incremental revenue per euro of ad spend, not the total revenue per euro of ad spend? A channel with a reported ROAS of 5.0 and an incrementality rate of 30 percent has a true incremental ROAS of 1.5. That is a very different investment thesis.
Cadence and budget for incrementality testing
Run a geo holdout every quarter for your highest-spend channel. Budget 15 to 20 percent of that channel's spend toward the holdout region. The cost of the test is the revenue you forego in the holdout region during the test period, which is typically 2 to 4 percent of monthly revenue for a properly sized test.
The return on that investment is knowing whether your largest marketing expense is creating demand or merely claiming credit for demand that already existed. For most brands above 5 million euros in annual revenue, that knowledge is worth the test cost many times over.
---