What Is Incrementality Testing?

What It Actually Measures

Attribution tells you which touchpoint a customer saw last before they bought. Incrementality testing asks a different question: would that sale have happened anyway? Those are not the same question, and the gap between them is where most wasted ad spend hides.

Last-click attribution can't answer the causal question because it has no way to see the customer who would have converted with zero ads. It just counts who touched what. Incrementality testing fixes that by building a control group — a slice of the audience that sees no ad, or a placebo ad — and comparing its conversion rate to the group that was exposed. The difference is the real, incremental contribution of the media.

This matters most for channels where the platform grades its own homework: Google, Meta, and any walled garden reporting self-attributed conversions. It's also the only honest way to size the contribution of channels that generate impressions without clicks, like display and video, where last-click credit is close to meaningless.

The Formula and a Worked Example

The standard calculation: (Test group conversion rate − Control group conversion rate) ÷ Test group conversion rate = Incrementality %.

Say a test group of 20,000 people sees the campaign and converts at 3.0%. A matched control group of 20,000 sees nothing and converts at 1.8%. (3.0% − 1.8%) ÷ 3.0% = 40% incrementality. That means 60% of the conversions in the test group would have happened anyway — the campaign only caused four in ten of them.

Here's why the platform number and the incremental number can diverge sharply. Across 30 of our advertisers in July 2026, the book attributed 57,137 purchases at a blended $5.49 cost per acquisition, with 20.1% of those from shoppers new to the brand. That figure is real and it's useful — but it's an attributed number, not a tested one. It tells you what the platform counted, not what the platform caused. Reconciling that number against a holdout, inside Amazon Marketing Cloud where DSP and sponsored ads stop double-counting each other, is the only way to know how much of it was incremental.

Test Types, Side by Side

Not every channel can run the same test. The table below covers the four structures you'll actually encounter.

When the Number Comes Back Bad

A low or negative incrementality result isn't a failed test. It's the test working. Three things to check before you act on it:

• Holdout too small. If your control group is under roughly 10% of total reach, noise can swamp the signal and a real lift can look flat.
• Seasonality contamination. If demand for your category swings independently of ad spend — a launch, a competitor promo, a holiday — a test run across that window can blame or credit the ads for something else entirely.
• Sample ratio mismatch. If the test and control groups aren't actually comparable in size, geography, or prior purchase behavior, the comparison is void before you've even read the result.

We've made the small-holdout mistake ourselves, early on, sizing a control group to protect media efficiency rather than statistical power — the test came back inconclusive and had to be rerun for another cycle. The fix isn't to distrust every bad number. It's to check the design before you trust any number, good or bad.

Google, Meta, and the Specialist Vendors

Google runs its own conversion lift studies inside Google Ads, and Meta runs conversion lift studies inside Ads Manager — both are native, both are free to run, and both are graded by the platform whose spend is being tested. That's a real limitation, not a fatal one: they're a reasonable first check, especially for single-channel questions.

Independent vendors go further. Measured runs continuous, always-on experimentation across hundreds of data sources and uses the results to calibrate a media mix model — a genuinely strong fit for brands running many channels at once who want one reconciled view. Haus specializes in geo-experiments and incrementality-as-a-service, and suits teams that want a dedicated testing partner without building the infrastructure themselves. Both are legitimate answers to this question and outperform doing nothing.

What none of them do natively is see inside Amazon's own walled garden the way a clean room built for it can. Sponsored ads and DSP report separately by default, which means a brand running both can double-count the same shopper twice before any incrementality test even starts.

Where Amazon DSP Fits

Dr. DSP is Amazon DSP — Demand-Side Platform, Amazon's programmatic display, video, and audio buying — run as a managed product, not the Delivery Service Partner courier franchise. It's built on the same holdout-and-control logic described above, reconciled inside Amazon Marketing Cloud so DSP and sponsored ads stop claiming the same sale twice. Every change we make carries the evidence behind it, a measurement plan, and a rollback trigger before it goes live. Full Circle has managed more than $500M in revenue across 100+ brands; there's no published price, a demo and the first 30 days are free, and pricing is set on the call against actual budget and scope. Whether or not that's the right fit for you, the underlying test — holdout versus exposed, causal versus correlated — is the same one worth demanding from anyone measuring your media.