Incrementality testing has been on every attribution conference agenda for five years. Most mid-market advertisers I work with have not done it. The reasons are budget, statistical literacy, and the operational complexity of running clean holdouts. The advertisers that have started doing it in 2025 and 2026 are producing genuinely useful results, but the path to those results is harder than the marketing literature implies.

What incrementality testing actually involves

Hold out a randomly selected portion of your audience from exposure to a specific channel or campaign. Measure the difference in conversion behaviour between the held-out audience and the exposed audience. The difference is the incremental effect.

The mechanics sound simple. The execution is full of traps.

Trap one: your channel cannot really hold out

Many of the channels that matter — paid search on branded queries, retargeting, certain types of social — do not support clean holdouts at the audience level. The held-out audience may still see the ads through other targeting paths. The exposed audience may not see the ads despite being in the exposed group.

The first conversation in any incrementality testing programme is which channels you can actually test cleanly. Branded paid search is hard. Display retargeting is moderately hard. Top-of-funnel display is the easiest.

Trap two: the test budget is bigger than you think

For a holdout to produce statistically significant results within a useful time window, the test budget needs to be large enough to produce a detectable lift over noise. For most mid-market advertisers, this means committing weeks or months of substantial spend to a structured test.

The math is unforgiving. If your baseline conversion rate is 2% and you are looking for a 10% lift with 95% confidence, you need substantial sample size in both arms. The smaller the lift you are trying to detect, the much larger the sample size requirement.

Trap three: confounding from other channels

Your incrementality test is happening while your other marketing continues to run. The other channels affect the same audience. Properly isolating the incremental effect of the test channel from the noise of the other channels requires either suppressing the other channels for the test population or modelling their effect carefully.

Suppressing the other channels is operationally painful. Modelling their effect is statistically demanding.

What is actually working

The mid-market advertisers running successful incrementality programmes have settled into a pattern. Test one channel at a time. Test cleanly with adequate budget and adequate duration. Accept that some channels cannot be tested rigorously and rely on directional methods for those. Combine the test results with MMM and other modelling to produce a coherent picture.

The combination matters. Incrementality testing alone does not answer the budget allocation question. It feeds into the broader measurement framework.

What this means for budget allocation

The advertisers that have done the testing are reallocating budget. Top-of-funnel display has consistently underperformed expectations. Paid search on branded queries has consistently been overcredited by attribution platforms. Connected TV has produced more incremental effect than the cost per impression suggested.

The reallocations are meaningful but not radical. Most channels work but the relative weights need adjustment. The advertisers that are running second and third rounds of testing are finding that their initial reallocations were correct but smaller than they could have been.

A note on the consultant question

Incrementality testing is the kind of work where outside expertise pays back. For organisations starting from zero, an AI consultancy or a measurement specialist with experience in the specific channels being tested is worth the engagement. The cost of the consultant is small relative to the cost of running tests that produce results you cannot trust.