Incrementality Testing: How to Measure What Your Ads Cause

On October 7, 2026, Google opened Conversion Lift to Search and Performance Max campaigns, giving advertisers “self-serve access to incrementality measurement that previously required help from a Google representative,” Search Engine Land reported. The price had already fallen. In November 2025 Google said a single experiment that once might have cost upwards of $100,000 could be run for $5,000. Meta, for its part, called incrementality testing, and Conversion Lift in particular, “the gold standard in measurement” in March 2026, PPC Land reported.

That matters because the number most advertisers manage to, the return on ad spend shown in Ads Manager or Google Ads, answers a different question. It counts the sales that followed an ad. Incrementality testing counts the sales the ad caused, the ones that would not have happened without it. This guide covers what incrementality is, the evidence that the two numbers differ, the three ways to test it, what each needs, and how to read a result.

What incrementality measures

Incrementality is the difference between what happened with your ads and what would have happened without them. You can’t observe both for the same customer, so every test builds a comparison: people who could see the ads against similar people who could not, or regions with spend against regions without.

Platform return on ad spend (ROAS) works differently. It divides the revenue the platform attributes to your ads by what you spent. A sale counts if the buyer clicked or viewed an ad within the attribution window, whether or not the ad changed anything. Someone who searches for your brand, clicks your ad and buys was probably going to buy anyway. The platform still counts the sale.

A worked example, with illustrative figures, shows why the gap matters. Ads Manager reports a 4.0 ROAS on $10,000 of spend, or $40,000 of attributed revenue. A lift test finds that the group able to see the ads bought $16,000 more than the holdout group, scaled to the same size. The incremental ROAS is 1.6. At a 40% margin you need a ROAS of 2.5 to break even on the spend, so a campaign that looks excellent in the dashboard is losing money.

The error does not always run one way. Haus, which sells incrementality testing, analyzed 640 Meta tests run by its customers and found in 2025 that Meta’s 7-day click attribution under-reported incremental direct-to-consumer sales by 15% on average. For YouTube, its customers’ geographic tests found 2.3 times the direct-to-consumer sales Google Ads reported, plus 75 cents of Amazon and retail sales for every dollar of website sales, the company said in October 2026. Attributed and caused are different numbers, and only a test shows which way yours are off.

What large experiments found

• eBay’s brand search ads. When eBay paused ads on searches for its own name, “almost all of the forgone click traffic and attributed sales were captured by natural search,” economists Thomas Blake, Chris Nosko and Steven Tadelis reported in Econometrica in 2015. Ads on other keywords influenced new and infrequent shoppers, but most of that spend reached frequent buyers whose purchases the ads did not change.
• Facebook’s first comparison. Across 15 advertising experiments at Facebook, “Generally, the observational methods overestimate ad effectiveness relative to the RCT, although in some cases, they significantly underestimate effectiveness,” Brett Gordon, Florian Zettelmeyer and colleagues found, in research published in Marketing Science in 2019. In one study, comparing exposed with unexposed users suggested a 316% lift; the randomized experiment measured 73%.
• 663 experiments later. A follow-up across 663 Facebook experiments found median lifts of 29%, 18% and 5% for upper-, middle- and lower-funnel outcomes. The better of two observational models put them at 83%, 58% and 24%, Gordon, Robert Moakler and Zettelmeyer reported. Their verdict on those models: “we are unable to reliably estimate an ad campaign’s causal effect.”

Two lessons follow. Purchases and other lower-funnel outcomes show the smallest true lifts, so tests on them need the most volume. And the sales most at risk of over-crediting sit closest to the purchase: brand search, as eBay found, and by the same logic retargeting.

Three ways to test, and what each needs

1. Platform lift studies

Google and Meta both split users at random into a group that can see your ads and a holdout that can’t, then compare conversions.

Google Conversion Lift covers Display, Search, Video, Demand Gen, App and Performance Max campaigns. The user-based version needs at least 1,000 observed conversions and a minimum campaign budget of $5,000, according to Google’s setup guide. Studies can run from seven days, Google recommends more than 14, and a study can be extended up to 56 days. You choose a holdback between 1% and 50%. Google reports an 80% credible interval and needs at least 150 conversions from the treatment group and 65 from the control. A geography-based version, which can include offline sales, is in beta for campaigns targeting a single country.

Meta Conversion Lift works the same way. Meta’s developer documentation says API access is “currently limited” and points advertisers to their Meta representative, while agencies such as Flighted describe a self-serve version in the Experiments tab of Ads Manager. Meta’s documentation sets no single minimum. Hunch, a creative performance platform, advises at least 50 to 100 conversions a week per test cell over two to four weeks, and the agency Flighted warns that accounts with fewer than 100 to 200 a week often struggle to reach significance. Meta also offers incremental attribution, a setting that estimates which conversions its ads caused, covered with Meta’s 2026 attribution changes in our piece on what to measure instead of last click on Meta. TikTok runs Conversion Lift as a managed service with spend minimums it doesn’t publish.

Platform studies randomize at the user level at little or no extra cost. Their limit is that each platform measures its own ads, and only the conversions it can see. Sales on Amazon, through other channels, or in stores whose data you don’t upload to the platform fall outside the test.

2. Geographic holdout tests

A geo test switches spend off, or up, in a set of regions and compares total sales there with similar regions where spend continued. Because it uses your own sales data, it counts every sale, online, in store or on marketplaces, and it works for channels no platform can randomize, such as TV, radio and billboards.

The best-practice guide for Meta’s open-source GeoLift recommends at least 25 pre-test periods across 20 or more regions, with daily data “strongly recommended,” and asks for history of at least four to five times the test length, ideally a year, and a test of at least 15 days on daily data or four to six weeks on weekly data, covering one purchase cycle. Google’s equivalent, GeoX, became generally available in its Meridian toolkit in September 2026. It requires daily data and at least 10 regions, recommends 50 to more than 100, and adds a cooldown of one to two weeks.

The constraint is geography. A business that sells mostly in one city, or in a small country dominated by its capital, may lack enough comparable regions. There, a platform lift study is usually the better choice.

3. Marketing mix models

A marketing mix model estimates each channel’s contribution from how sales move with spend over time, controlling for seasonality, price and promotions. It needs no user tracking and covers offline media, so it is the one method that sees every channel at once. Its weakness is that it finds correlations. Calibrated with experiment results, it gets much closer to cause.

Meta’s open-source Robyn says a model “will need a minimum of two years of historical weekly data,” with about ten observations per variable. Google’s Meridian, open to all advertisers since January 2025, can use reach and frequency data and lift-test results as inputs. Google’s documentation warns that two years of weekly national data is too little to estimate a many-channel model reliably, and that even three years may give only directional results unless you cut the number of channels. Regional data helps.

In short: if most spend sits on one or two platforms with steady conversions, start with their lift studies. If you sell across many regions, or offline and marketplace sales matter, run geo tests. If you use several channels, including offline media, and have two or more years of stable weekly data (more with many channels, unless you have regional data), build a mix model and calibrate it with tests.

How to read a result

Google defines the core numbers simply. Incremental conversions are treatment conversions minus control conversions. Relative lift is incremental conversions divided by control conversions. Incremental ROAS is the incremental conversion value divided by ad spend, and incremental cost per action is spend divided by incremental conversions.

Judge the result against your economics, not the dashboard. Break-even ROAS is one divided by your margin after product, shipping and other variable costs: 2.5 at a 40% margin, 4.0 at 25%. Above that line, the spend paid for itself on the first purchase. Below it, the case rests on repeat purchases, a calculation with its own traps, as our piece on the 3:1 LTV to CAC rule explains.

Read the range, not only the point estimate. Google grades certainty: at 90% or more there is “a very good chance” the ads caused the results, at 70% to 90% “a good chance,” 50% to 70% is directional, and below 50% it reports no lift and recommends a new test. An inconclusive result is not proof the ads did nothing. Extend the test, add volume or test a bigger change in spend. In a simulation of budget decisions, Haus found that acting on noisy results ended below doing nothing 38% of the time, against 18% with precise measurement. Size the budget move to the strength of the evidence.

Where to start

Begin with one number no platform can inflate: marketing efficiency ratio, total revenue divided by total marketing spend, as Shopify defines it. It won’t say which channel works, but every test result has to reconcile with it.

Then test where over-crediting is most likely: brand search, the pattern eBay found, and retargeting, which also reaches people who already know you. Video and connected TV sit at the other end, with sales that arrive later and elsewhere, and post-view attribution is not lift, as our look at Paramount’s CTV measurement argues. Run one test at a time, avoid big promotions during the test window, and decide before launch what result would move budget and by how much.

A few tests a year, feeding a mix model once you have the history, is enough for most mid-size advertisers. Tests also show what upper-funnel spend contributes, the question behind the end of performance-only budgeting. And they need clean conversion data; signal loss explains where tracking gaps come from.

Designing tests and acting on the results is the core of performance marketing. Before a test, an online presence analysis shows which conversions fire and which channels have no measurement at all. The same logic applies to your website, where A/B tests measure what a page change causes, the basis of conversion rate optimization.

Frequently Asked Questions

What is incrementality testing?

Incrementality testing measures the sales, leads or installs that advertising caused, rather than those that followed an ad. It compares a group exposed to ads with a similar group that was not, either users split at random by the ad platform or matched geographic regions, and credits only the difference to the ads. Results are usually expressed as relative lift, incremental conversions or incremental return on ad spend.

How is incrementality different from ROAS?

Platform ROAS divides the revenue an ad platform attributes to your ads by ad spend. It counts any sale that followed a click or view inside the attribution window, including sales that would have happened anyway, such as brand searchers who were already going to buy. Incremental ROAS divides only the extra revenue the ads caused, measured against a holdout group, by ad spend. The two can differ widely in either direction.

How much budget does a Google or Meta lift test need?

Google’s user-based Conversion Lift needs at least 1,000 observed conversions and a minimum campaign budget of $5,000, and runs seven to 56 days, with more than 14 recommended. Meta sets no single published minimum for Conversion Lift; practitioners advise at least 50 to 100 conversions a week per test cell, some say 100 to 200, over two to four weeks. Geo tests depend less on budget than on having enough comparable regions and daily sales data.

How long should an incrementality test run?

Long enough to cover at least one purchase cycle and to collect enough conversions for a clear result. Google allows Conversion Lift studies of seven to 56 days and recommends more than 14. Meta’s GeoLift guidance is at least 15 days on daily data, or four to six weeks on weekly data. Google’s GeoX documentation suggests extending a test from four weeks to six or eight to detect smaller effects, then allowing a cooldown of one to two weeks.

What is a good incremental ROAS?

One above your break-even ROAS, which is one divided by your margin after variable costs. At a 50% margin, an incremental ROAS above 2.0 means the ads paid for themselves on the first purchase; at a 25% margin you need more than 4.0. Below break-even, the spend can still be worth it if customers come back and buy again, but then the case rests on lifetime value, which you should measure rather than assume.