Attribution tells you which tracked touches came before a sale. It cannot tell you whether the sale would have happened anyway. An incrementality test can. You withhold your marketing from a randomly chosen group of people or places, compare their outcomes with the group that saw it, and the difference is what your marketing caused.
Most lift tests that come back inconclusive were set up to fail, usually because nobody sized them. The design you pick and the sample you plan for decide whether you get an answer at all.
The sale that would have happened anyway
Picture a retargeting campaign that shows ads to everyone who visited your product page in the last seven days. Plenty of those visitors were going to buy regardless. When they do, the retargeting ad was on their path, so every attribution model hands it some credit, and the platform dashboard may show an excellent return on ad spend.
A lift test asks something sharper. If those ads had never run, how many of those people would still have bought? If the answer is “almost all of them”, the true return is a fraction of what the dashboard says. Branded search ads, retargeting and loyalty emails are the usual suspects, because they reach people who are already close to buying.
What you want out of the test is incremental conversions, and from those, incremental return on ad spend (iROAS): incremental revenue divided by spend.
Four ways to run one
User-level holdouts (platform conversion lift)
The platform randomly splits your target audience into a test group that can see your ads and a control group that cannot, then compares conversion rates. Google and Meta both offer this.
As of October 2026, Google’s Conversion Lift setup page sets these requirements for user-based studies:
- at least 1,000 observed conversions, not counting modelled conversions
- a minimum campaign budget of $5,000
- Display, Search, Video, Demand Gen, App or Performance Max campaigns (iOS app campaigns and travel ads are excluded)
- a holdback between 1% and 50%
- a run of at least 7 days, with 14 or more recommended.
Before launch, Google shows a “study power” estimate and recommends aiming for 90% or higher; results between 50% and 90% count as directional. The same page notes Conversion Lift “isn’t available for all Google Ads accounts”, and Google’s comparison page says some campaign types still need an account representative. In November 2025, trade press reported that Google had lowered its experiment budget threshold to $5,000 from figures approaching $100,000. That put these tests within reach of many smaller advertisers.
Meta runs conversion lift through the Experiments tool in Ads Manager. Eligibility and settings vary by account, so go by what your own account shows, not by third-party guides.
Randomising at the person level makes the comparison clean. The catch is that the platform marks its own homework, and it only measures its own ads.
PSA and ghost ads
Early ad experiments showed the control group a placebo ad, often a public service announcement (PSA), so that both groups went through the same delivery. You paid for those placebo impressions, though, and modern delivery algorithms show the PSA to different people from the ones who would have seen your real ad.
The ghost ads method, published in the Journal of Marketing Research in 2017, fixed this. It logs which control-group users would have been shown your ad, without showing them anything or paying for it, and compares them with the people who did see it. Cleaner and cheaper. Nobody builds ghost ads in-house, but platform lift tools rest on ideas like it, so ask your vendor which method theirs uses.
Geo holdouts
Switch a channel off (or on, or up) in some regions, leave it alone in others, and compare sales. You measure in your own sales data, so a geo test works for any channel, offline sales included, and doesn’t depend on the platform’s reporting. Google’s comparison page describes it as the more transparent, cross-channel option. The price is noisier results that cannot be split by demographics.
Matched markets and synthetic controls
A simple geo holdout picks treatment and control regions that look alike. A synthetic control builds a weighted blend of control regions whose past sales closely track the treatment regions, then uses that blend to predict what would have happened without the change.
There are two open-source tools for this. Meta’s GeoLift uses the augmented synthetic control method; its walkthrough works from 90 days of daily pre-test data across 40 cities and recommends that a test covers at least one full purchase cycle. Google’s Meridian GeoX supports holdback, go-dark and heavy-up designs and multi-cell tests that share a control group, and it can feed results into a Meridian marketing mix model.
| Your situation | Design to start with |
|---|---|
| One platform, enough conversions, account is eligible | Platform conversion lift |
| You need to measure offline or total sales | Geo holdout |
| Few regions, uneven region sizes | Synthetic control (GeoLift or GeoX) |
| Testing several spend levels at once | Multi-cell geo test |
| Very low conversion volume | Longer geo test on your biggest channel, or wait |
Sizing the test before you spend anything
The maths is the same as for any experiment. Our lesson on A/B testing statistics and the A/B test sample size tactic go through it step by step.
Here is a hypothetical user-level case. The target audience converts at 1.0% without ads. You want to detect a 10% relative lift (1.0% to 1.1%) at the usual 95% confidence and 80% power.
| Split | Control group | Exposed group |
|---|---|---|
| 50/50 | about 163,000 | about 163,000 |
| 90/10 (10% holdout) | about 87,000 | about 785,000 |
The small holdout loses fewer sales while the test runs, and needs a far bigger total audience to get there. Put your own numbers into an A/B test calculator before committing budget.
Geo tests follow the same logic, except the sample is regions and time. Fewer or noisier regions mean a longer test or a bigger expected effect. GeoLift and GeoX both include power analysis that simulates your test on past data; use it.
A few rules of thumb hold across designs. Halving the lift you want to detect roughly quadruples the sample you need. The test should cover at least one purchase cycle, plus a cool-down if people tend to convert days after seeing an ad. And decide up front the smallest lift that would change your budget; if the test can’t detect that, don’t run it.
Reading the result
A lift readout usually gives you the lift (absolute or relative), a confidence or credible interval, and the incremental conversions.
Look at the interval before the headline. “12% lift, interval −3% to 27%” means you cannot rule out zero effect. “12% lift, interval 6% to 18%” is something you can act on.
Then convert it to money:
- Incremental conversions = conversions in test group − expected conversions without ads (from the control group, scaled to the test group size).
- iROAS = incremental revenue ÷ ad spend during the test.
- Incrementality ratio = incremental conversions ÷ conversions the platform reported for the same period.
The ratio is the one you’ll use week to week. A platform that reported 1,000 conversions against 400 incremental ones has a ratio of 0.4, and you can apply that as a correction factor to the channel’s dashboard numbers until the next test.
Where lift tests go wrong
- No hypothesis. “Let’s see what happens” produces results nobody acts on. Write the decision down first, as covered in hypothesis design.
- Too little power. The test runs for a week because that is what the budget allowed, and a noisy result gets called a win or a loss.
- Changes mid-test. New creative, a price change or a sale in one region breaks the comparison.
- A contaminated control group. Other campaigns, emails and promotions reaching the control group are fine if they reach both groups equally. A regional promotion in the control regions only is not.
- Spillover. People in a “dark” region still see national TV, travel into treatment regions, or order online to another address. Choose regions that are well separated.
- The wrong channel. Test the channel you spend most on and trust least, not the one that is easiest to test.
- Treating one result as permanent. Creative, competition and prices move. Re-test important channels once or twice a year at least.
- Taking the platform’s test as independent. It is still the platform measuring itself. Where you can, cross-check with a geo test in your own sales data.
Picking your first test
Rank your channels by spend and by how confident you are that they are incremental. Your first test is the channel at the top of the spend list and the bottom of the confidence list. Before anything else, write one sentence: “If lift is below X%, we will cut spend by Y.”
If your Google or Meta account shows a conversion lift option and you clear its stated minimums, start there. If it doesn’t, pull at least 90 days of daily sales by region and check whether you have enough regions for a geo test. Either way, size it with a calculator first. When the test can’t detect the lift that would change your budget, change the design. Don’t run it and hope.
Questions people ask
What is the difference between an A/B test and an incrementality test?
An A/B test compares two versions of something, such as two ads or two landing pages, and tells you which performs better. An incrementality test compares showing your marketing against not showing it, and tells you how many conversions it caused that would not have happened otherwise.
How long should an incrementality test run?
Long enough to cover at least one full purchase cycle and reach the sample size you calculated in advance. Google's Conversion Lift help page allows studies as short as 7 days but recommends at least 14, and longer consideration cycles need longer tests.
What is a good holdout size?
It is a trade-off. A larger control group gives a clearer answer faster but costs more lost sales. Many platform tools allow a range, for example Google's Conversion Lift accepts 1% to 50%. Size it from your power calculation, not by habit.
Can a small business run incrementality tests?
Yes, if you can switch a channel off in some regions or for a randomised group and still measure sales in both. A simple geo holdout on your largest channel is often the most useful test a small team can run.
Sources
- Google Ads Help: Set up Conversion Lift based on users
- Google Ads Help: Comparing lift types
- PPC Land: Google lowers incrementality testing threshold to $5,000 for advertisers
- Kellogg School of Management: Ghost Ads, Improving the Economics of Measuring Online Ad Effectiveness
- Meta Open Source: GeoLift walkthrough
- Google for Developers: Introduction to Meridian GeoX
- Haus: Understanding Meta incrementality testing
