Incrementality
Incrementality is the revenue marketing caused, above what would have occurred without it. Attribution assigns credit to touchpoints that preceded a sale. Incrementality measures whether the sale depended on the marketing at all. A campaign can receive attribution credit for many sales that would have happened regardless; only the sales that would not have happened count as incremental.
How it actually works
Incrementality is measured with a controlled experiment: a comparable group is held out from a campaign, and the exposed group is compared against the holdout. The difference between the two is the incremental lift. Platforms report their own version, iROAS, but the platform defines the baseline and assigns its own credit, so a platform-reported iROAS is an estimate built on the seller's assumptions.
A reliable incrementality read requires a control group the advertiser designs and owns. With one, the question of whether a campaign added revenue has a measured answer. Without one, the answer rests on the platform's attribution model.
The design work happens before the test runs. The minimum detectable effect, the smallest lift the test is sized to distinguish from noise, is computed in advance from traffic volume, holdout size, and test length; without it, the most expensive error in experimentation goes through unchecked, which is reading "no statistically significant lift" as "no lift" and cutting a channel the test was never large enough to measure. The experiment also has a price that can be computed the same way: the holdout share multiplied by the expected incremental revenue over the test window is revenue deliberately forgone to buy the answer, so a 10% holdout on a channel believed to drive $2 million over the window costs about $200,000 of expected revenue, a number that is weighed against the budget the answer will govern. In geographic holdouts, contamination sets a floor on what the design can see: exposed users who travel or shop across the holdout boundary, and media that bleeds across it, deliver treatment into the control group, which biases the measured lift toward zero and makes a real effect look absent.
When incrementality is not measured, teams allocate budget by attribution credit, which concentrates spend on demand that already existed. Campaigns creating new customers report lower attributed efficiency and get cut, and growth slows while the dashboards report that the spend is working.
In practice
At a large insurance brand, the incumbent agency, which sold and managed the ads, reported a lower acquisition cost in its first three-month test, and a $7 million renewal depended on that result. Brought in as a second set of eyes, we ran a controlled experiment measuring net new acquisitions and contribution margin. The reported gain came from a seasonal rise in demand that neither party had been monitoring, and in every treatment where the ads ran, acquisition costs rose. The renewal did not happen. The attributed result and the incremental result pointed in opposite directions, and only the experiment could tell them apart. Subtraction is the simplest form of the test. The owner of several chiropractic clinics was spending nearly 15% of revenue each month on advertising heading into the pandemic. He shut off all his ads, and revenue stayed the same: customers were finding him organically through Google My Business and local SEO. When he was ready to grow, a rebuilt program got better results for a tenth of what he had been spending, and he opened two more clinics ([platform bias and attribution](/insights/why-your-vendors-win#platform-bias-and-attribution)). At a CPG brand, the same subtractive test shut down organic Instagram, Facebook, and YouTube, and revenue did not move; roughly $50,000 a month in Facebook spend the agency insisted was protecting revenue was turned off, and revenue did not move ([the psychology of recovering margins](/insights/what-turnarounds-teach#the-psychology-of-recovering-margins)).
Where we come in
We design the controlled experiments and holdouts that isolate real lift, and we build the test structure and measurement outside the platforms, so teams allocate budget by measured cause. Our principals have used subtraction-based frameworks in client work since 2016: turn something off, or hold it out, and measure whether the number moves. It is the same approach behind our work with Potbelly, Digital Realty, and Equinix.
Start a Revenue Health Pre-Assessment →See it in action
Related terms
- Control group / holdout
- A comparable set deliberately not shown a campaign, used to measure what would have happened without it. It is the basis of any causal claim about marketing.
- Lift testing
- Measuring the difference in behavior between the exposed group and the control. That difference is the incremental result.
- Marketing mix modeling (MMM)
- A statistical estimate of each channel's contribution using historical data, without tracking pixels. It works at the portfolio level, where per-user attribution breaks down.
- Minimum detectable effect (MDE)
- The smallest lift a test is sized to distinguish from noise, computed before the test from volume, holdout size, and duration. A result below the MDE is unmeasured rather than zero.
- iROAS
- Incremental ROAS. When the platform calculates it, the platform also sets the baseline, so it is an estimate under the seller's assumptions.