Incrementality Testing for Google AdsHow to Actually Measure It

Incrementality testing measures the sales you'd lose if a campaign turned off, not the sales the platform attributes to it. The two numbers are always different, and for brand search and retargeting they can be wildly different, because platform attribution counts conversions that would have happened anyway. Only a controlled experiment isolates the real lift.

The standard method is a geo holdout: split matched regions into test and control groups, run ads in test regions only, and compare actual revenue lift against what the platform claims. This isn't a one-time exercise. Spend levels, seasonality, and competitive pressure change, so the incremental value of a campaign changes with them.

This post covers how to structure a geo holdout test, how long to run it, and which Google Ads campaign types are worth testing first because their reported-vs-real gap tends to be largest.

Why Platform-Reported ROAS Isn't Real ROAS

Google Ads attributes a conversion to the last (or a modeled) ad click before purchase, regardless of whether that click actually changed the buyer's decision. A customer who was already going to buy and searched your brand name gets counted the same as a customer who discovered you cold through a prospecting campaign. Both show up as "ROAS," but only one represents money the ad spend actually generated.

This gap is largest on brand search and retargeting/remarketing campaigns, where the audience already has high purchase intent independent of the ad. It's smallest on cold prospecting campaigns like non-brand search or PMax targeting new audiences, where the ad is more likely to be the actual reason for the sale.

Setting Up a Geo Holdout Test

Split your served geographies into two matched groups based on historical revenue similarity, not just population size. Pause the campaign you're testing in the control group's geos while running it as normal in the test group's geos. Compare actual revenue (not platform-attributed conversions) between the two groups over the test period, adjusted for any baseline difference recorded before the test started.

Four weeks is the practical minimum for most DTC accounts to get past week-to-week noise, and six to eight weeks is better for lower-volume accounts or highly seasonal products. Running a test for less than two weeks produces numbers you can't trust either way.

Which Campaigns to Test First

Test brand search and remarketing campaigns before non-brand or prospecting campaigns. These are the campaigns most likely to be overstated by platform attribution, and they're usually the ones with budget increases sitting on the table based on inflated ROAS numbers. If a geo holdout shows brand search has minimal incremental lift, that's the highest-leverage finding available, because it directly informs a rule that should already be standard practice: brand campaigns run on Maximize Clicks with a capped CPC, never on target ROAS, because there's no reliable incremental signal to optimize against in the first place.

Reading Results Without Fooling Yourself

A geo holdout that shows 40% incrementality on a campaign platform-reports as 400% ROAS doesn't mean the campaign is bad. It means the campaign's real ROAS is closer to 160%, and budget decisions should be made against that number, not the inflated one. Shopiator is a Google Ads agency for ecommerce brands, run personally by a 16-year PPC specialist with expertise in DTC, supplement/healthcare compliance, PMax architecture, and incrementality testing, and the recurring lesson across accounts is that clients who never ran an incrementality test are almost always over-crediting their brand and retargeting spend.

When Incrementality Testing Isn't Worth the Overhead

Below roughly $10-15k/month in Google Ads spend, a geo holdout often doesn't have enough volume per geo group to produce a statistically usable result, and the operational cost of running one outweighs the insight. At that spend level, a simpler heuristic works: assume brand and remarketing ROAS is overstated by 40-60% and budget accordingly, then revisit real testing once spend scales.