Incrementality Testing: A Practical Guide

Incrementality testing measures how many of your sales your ads actually caused, by comparing people or regions that saw your ads with a similar group that didn't. The gap between the two is your incremental lift. It's the only common method that separates sales your ads created from sales that would have happened anyway.

That distinction matters because every ad platform is built to take credit. Meta, Google and every other channel report conversions from people who saw or clicked an ad, whether or not the ad changed anything. Incrementality testing is how you find out what your spend is really buying.

Key takeaways

• Attribution shows which ads people saw before they bought. Incrementality shows which purchases the ads caused. Expect your incremental ROAS to be lower than what the platforms report.
• There are four practical ways to test: platform lift studies, geo holdout tests, on/off tests and modeled estimates. The right one depends on your budget, conversion volume and question.
• Google says advertisers can now start an incrementality test with as little as $5,000 in experiment budget. Geo tests usually need more budget and more time.
• Decide what result would change your budget before the test starts. A test without a decision rule produces a number nobody acts on.
• Read the range, not just the headline number. A result can be useful even when it's wide, if the whole range points to the same decision.

What is incrementality testing, and how is it different from attribution?

Attribution assigns credit for a sale to the ads someone saw or clicked before buying. Incrementality testing asks a different question: would that sale have happened if the person had never seen the ad?

Attribution can't answer that, because it only looks at people who saw ads. As EMARKETER puts it, attribution assigns credit across touchpoints but cannot prove causation. A loyal customer who was going to reorder anyway still shows up as a Meta conversion if they scrolled past your ad first.

An incrementality test creates a fair comparison. One group can see your ads; a similar group can't. Everything else stays the same, so the difference in sales between the two groups is what the ads caused.

This is also why platform numbers never agree with each other. If you've wondered why Meta, GA4 and your store all report different numbers, our guide to why Meta Ads and GA4 never match explains the mechanics. Incrementality testing is the step after that: it tells you which number reflects reality.

Why incremental ROAS is usually lower

Incremental ROAS counts only the revenue your ads caused, so it's almost always lower than the ROAS a platform reports. EMARKETER notes that marketers used to last-touch figures need to adjust their expectations when they move to incremental measurement.

A lower number isn't bad news. It's the first number you can safely base a budget decision on.

Which incrementality test fits your budget?

There are four common approaches. They differ in cost, accuracy and what they can measure.

Method	How it works	Best for	Watch out for
Platform lift study (Meta Conversion Lift, Google Conversion Lift)	The platform randomly holds back a share of your audience from seeing ads and compares conversions	Testing one channel, with clean user-level randomization	The platform grades its own homework, and it can't see sales the holdout group made after seeing ads on other channels
Geo holdout test	You stop or cut ads in some regions and compare their sales with similar regions where ads keep running	Measuring total sales, including offline and other channels, independently of the platform	Needs enough regions and stable regional sales; noisier than user-level tests
On/off test	You pause a channel entirely for a period and compare sales with before and after	Small budgets, or a quick sense check on a channel you suspect isn't working	Seasonality, promotions and other changes can easily be mistaken for the ad effect
Modeled estimate (Meta's incremental attribution)	The platform uses a model to estimate which conversions its ads caused, without you running a test	An ongoing, free directional read inside Ads Manager	It's a model, not an experiment, and it only covers that platform

Platform lift studies

Google says advertisers can launch an incrementality test with an experiment budget as low as $5,000, using Bayesian methods that need less data than older approaches. According to Google's own comparison of lift types, user-based Conversion Lift is available directly in Google Ads for Video, Discovery and Demand Gen campaigns. For Search, Shopping and Performance Max, you need to contact your Google account representative.

Meta runs Conversion Lift through the Experiments section of Ads Manager. Meta randomly assigns people to a test group that sees your ads and a control group that doesn't, then compares conversions and revenue between them. Access varies by account: Meta's developer documentation says Conversion Lift access is currently limited and directs advertisers to their Meta representative.

There's no fee for either. You pay only for the ad spend during the test.

Geo holdout tests

Geo tests split regions rather than people. Because they compare total sales by region, they capture offline sales and the effect of other channels, which platform studies can't.

Two free, open-source tools help design and analyze them: Meta's GeoLift, an R package that uses synthetic control methods, and Google's Meridian GeoX. Google's GeoX guidance recommends at least 10 regions, uses a 28-day test as its standard example, and suggests extending to six or eight weeks when results are volatile.

Meta's incremental attribution column

Meta also offers an incremental attribution setting that estimates which conversions its ads caused. You can add incremental columns next to your standard results through the Compare Attribution Settings option in Ads Manager, without changing your campaigns. It's a useful ongoing signal, but it's a model, not a controlled experiment.

How do you run an incrementality test?

The mechanics differ by method, but every good test follows the same six steps.

Step 1: Decide what result would change your decision

Before anything else, write down your break-even point. If your gross margin is 60%, every $1 of ad spend needs to bring in at least $1.67 of incremental revenue (1 ÷ 0.60) just to pay for itself. That 1.67x is your break-even incremental ROAS.

Then agree on the decision upfront: "If incremental ROAS is above 2.0x, we increase budget. If it's below 1.67x, we cut it."

Step 2: Check your tracking

Your conversion data has to be reliable before a test can mean anything. For Meta, that means the Pixel and Conversions API are set up and deduplicated. For geo tests, you need daily sales data by region going back at least several months.

Step 3: Check you have enough volume

Small tests produce wide, inconclusive results. One practitioner guide suggests Meta Conversion Lift works best with around $50,000 a month in spend and 100–200 conversions a week on the campaigns being tested. Google's Bayesian approach lowers the bar, but it doesn't remove it.

A useful rule of thumb from Sellforte: the lift you expect should be at least twice the smallest lift the test can detect. If it isn't, run the test longer or skip it.

Step 4: Choose the holdout size

Bigger holdouts give clearer results but cost more in lost sales. For user-level tests, Sellforte suggests a 5–10% holdout can be enough when volume is high. For geo tests, holding out 30–50% of eligible regions is a common practical range.

Step 5: Run it without touching anything

Leave the tested campaigns alone: no budget shifts, no new creative, no bid strategy changes. Changes mid-test can invalidate the result. Plan for two to four weeks for platform lift studies, and four weeks or more for geo tests.

Google's GeoX guidance also recommends a one- to two-week cooldown after a geo test before returning to normal, because ads can keep influencing sales after they stop.

Step 6: Read the range, then decide

Every test gives a range, not a single number. If the whole range sits below your break-even, cut. If it sits above your target, scale. If it straddles both, the test was too small to decide, and that's a result too.

A worked example: when 2.0x ROAS is really 1.2x

This is an illustration built to show the math, not client data.

An ecommerce brand spends $60,000 a month on Meta. Ads Manager reports 1,200 purchases, a $50 cost per purchase. With a $100 average order, that's a reported ROAS of 2.0x.

The brand's gross margin is 60%, so its break-even incremental ROAS is 1.67x. On reported numbers, Meta looks comfortably profitable.

The brand runs a four-week geo holdout, pausing Meta in a matched set of regions. The analysis estimates that Meta caused about 720 of the 1,200 reported purchases: 60% of what the platform claimed.

• Incremental revenue: 720 × $100 = $72,000
• Incremental ROAS: $72,000 ÷ $60,000 = 1.2x
• Cost per incremental purchase: $60,000 ÷ 720 = $83

The test's 90% range for incremental ROAS is 0.8x to 1.6x. The entire range is below the 1.67x break-even, so the decision is clear even though the estimate isn't precise: at this spend level, Meta is losing money.

That doesn't necessarily mean "turn Meta off." It means the last dollars are the least efficient. The next step is to test a lower budget, or a change in targeting or creative, and measure again.

What are the most common incrementality testing mistakes?

Most failed tests fail for the same handful of reasons.

• Changing campaigns during the test. Any change to budget, creative or targeting muddies the comparison.
• Testing with too little volume. The result comes back "inconclusive," and the team concludes testing doesn't work.
• Stopping too early. Ads often keep influencing sales after they stop. In geo experiments run by the measurement firm Haus on TikTok, brands saw an additional 68% lift in their main KPI during the post-test window.
• Reporting one number without the range. A single "lift" figure invites overconfidence.
• Testing everything at once. Test one channel or one big change at a time, or you won't know what caused the result.
• Treating the result as permanent. Incrementality changes as you change spend, creative and audiences. A test result is true for the conditions you tested.

How often should you test?

This is our view, not a published standard: test each major channel at least twice a year, and after any big change in how the platform works. Google's AI Max auto-upgrade is a good example: if your Search campaigns changed in September, a lift test tells you whether the extra conversions are real.

Between tests, keep an eye on blended efficiency (total revenue divided by total ad spend) and Meta's incremental attribution column as a running check.

Frequently asked questions

What is incrementality testing?

Incrementality testing measures how many sales your ads actually caused. It compares a group that could see your ads with a similar group that couldn't, and the difference in sales between them is the incremental lift. It separates sales your ads created from sales that would have happened anyway.

How is incrementality different from attribution?

Attribution gives credit for a sale to the ads someone saw or clicked before buying, whether or not those ads changed their decision. Incrementality testing uses a control group to measure what would have happened without the ads. Attribution shows correlation; incrementality testing estimates causation.

How much budget do you need for an incrementality test?

It depends on the method and your conversion volume. Google says advertisers can start an incrementality test with an experiment budget as low as $5,000. Practitioner guidance for Meta Conversion Lift suggests around $50,000 a month in spend and 100 to 200 conversions a week, and geo tests usually need more budget than user-level tests.

How long should an incrementality test run?

Platform lift studies typically run two to four weeks. Geo tests usually need four weeks or more, and Google's Meridian GeoX guidance suggests six to eight weeks when results are volatile, followed by a one- to two-week cooldown. Avoid ending a test early or changing campaigns while it runs.

What is a good incremental ROAS?

A good incremental ROAS is one above your break-even point, which you can calculate as 1 divided by your gross margin. For example, with a 60% gross margin, break-even incremental ROAS is about 1.67x. Expect incremental ROAS to be lower than the ROAS your ad platforms report.

Is Meta's incremental attribution the same as an incrementality test?

No. Meta's incremental attribution setting uses a model to estimate which conversions its ads caused, without running a controlled experiment. It's a useful ongoing signal, but a Conversion Lift study or geo holdout test gives stronger evidence because it compares against a real control group.

The bottom line

Platform dashboards tell you who saw an ad before buying. Incrementality testing tells you who bought because of it. You don't need an enterprise budget to start: pick one channel, set a break-even point, choose the test your volume can support, and let the range decide your budget.

If you want help designing a test or reading the result, our measurement and attribution audit covers your tracking setup, your platform numbers and a test plan for your biggest channel.