Add up the conversions your ad platforms claim and you will often find they exceed the orders in your accounting system. Nobody is lying. Every platform is counting the same sale, and each one credits itself for people who were already on their way to buying.
Advertisers running independent tests routinely find that platform-reported impact overstates true impact by 25 to 40 percent. As AI-driven buying — Performance Max, Advantage+ and their equivalents — takes over targeting decisions, the reporting has become less transparent at exactly the moment budgets got harder to justify. Incrementality testing is how you get an answer you can trust.
Attribution answers "which touchpoint preceded the sale?" That is not the business question. The business question is "would this sale have happened anyway?" Those diverge most in the places where spend concentrates: branded search, where many clickers would have found you regardless, and retargeting, where you are paying to reach people who already chose you.
Signal loss makes it worse. With tracking restrictions widespread and cookie-based measurement effectively dead, platforms increasingly model the conversions they report. Modelled numbers are not fabricated, but they are estimates built by a party with an interest in the outcome.
The logic is a controlled experiment: withhold advertising from a comparable group, run it for everyone else, and measure the difference in total business outcomes — not platform-reported ones.
Turn a campaign off in a set of matched regions while it runs normally elsewhere. Compare total orders or leads between the two. This is the most practical design for most businesses because it needs no user-level tracking at all.
Platform-native conversion lift tools hold back a randomised share of your audience and report the difference. Convenient and statistically clean, but you are asking the platform to grade its own work — useful, best corroborated.
The crudest and most under-used: switch a campaign off entirely for two to four weeks and watch total revenue. If nothing moves, you learned something worth far more than the test cost.
Start where the money and the doubt are largest. For most advertisers that is branded search or retargeting — the two line items with the best reported ROAS and the weakest incremental case. Split your regions into matched halves using the last twelve months of revenue, pause the campaign in one half, and leave everything else alone for a month.
Then act on the result, which is where most teams fail. If retargeting proves 60 percent incremental, its true cost per acquisition is roughly 1.7 times what the dashboard shows. Reprice it, reallocate the difference into prospecting, and re-test in six months.
Attribution tells you what the platform saw. Incrementality tells you what your business earned.
No single method is sufficient. The framework serious advertisers converged on in 2026 uses three layers: incrementality tests as causal ground truth, run a few times a year on the biggest questions; marketing mix modelling for portfolio-level budget allocation once you have enough spend history; and platform attribution as a daily steering signal only — useful for spotting a broken campaign, never for deciding annual budget.
Smaller advertisers can skip the modelling layer entirely. Two well-run holdout tests a year plus honest total-revenue tracking will beat any amount of dashboard-staring.
Budgets get defended with causation, not correlation. Pick your largest questionable line item, hold it out in matched regions for four weeks, measure total business outcomes, and reprice what you learn. It is the cheapest research you will run all year, and it is the only number in your reporting that nobody has an incentive to inflate.
We design and run incrementality tests, then rebuild media plans around what genuinely drives revenue rather than what the dashboard claims.
Talk to us