What is Conversion Lift?
Also known as: Conversion lift study, Lift study
What is conversion lift?
Conversion lift is a randomized controlled experiment that measures the incremental conversions an ad campaign caused, by holding out a portion of the target audience and comparing their conversion rate against the exposed group. Meta launched the methodology in 2014 (Meta for Business) and most major platforms now copy the design.
The exposed group sees the ad. The holdout group is eligible but never served. The platform tracks both cohorts identically. The gap in conversion rate, multiplied by holdout size, gives the incremental conversions.
Reported conversions double-count buyers who would have purchased anyway. Conversion lift strips that out. It is the cleanest user-level read on whether an ad actually worked.
How conversion lift works
A conversion lift study runs in three parts. Random assignment, ghost-ad delivery, and statistical readout.
Test versus holdout group
The platform randomly splits the eligible audience. A typical split is 90 percent treatment, 10 percent holdout. Larger holdouts give more statistical power but cost more in foregone revenue. The split is locked at test launch and cannot be changed mid-flight.
Both groups are equally targetable. The only difference is whether the auction is allowed to serve them.
Ghost ads
Holdout users are flagged in the bidding system. The platform still calculates whether it would have shown them an ad, then logs that intent without serving. This "ghost ad" mechanism removes selection bias. A user who would never have qualified for the ad cannot dilute the control group.
Google calls this the Ghost Ads methodology, originally published by their research team in 2017.
Statistical significance
At readout, the platform calculates lift, a confidence interval, and a p-value. Lift is the percent change in conversion rate. The confidence interval is the plausible range. P-values below 0.05 are the standard threshold for declaring the lift real.
[UNIQUE INSIGHT] Most marketers fixate on the lift number. The confidence interval matters more. A 20 percent lift with a 95 percent CI of -5 to 45 percent is not a positive result. It is a noisy one.
Meta Conversion Lift vs Google Conversion Lift vs custom holdout
Each platform packages the same statistical idea with different constraints. The choice depends on channel mix and account size.
| Platform | How the holdout works | Min spend | Best for |
|---|---|---|---|
| Meta Conversion Lift | User-level split inside Ads Manager. Ghost bids on holdout. | Roughly $30k over the test window | Meta-only readouts, fastest setup |
| Google Conversion Lift | Account-level holdout across Search, YouTube, Display, Shopping | Higher, gated by Google rep | Cross-format Google measurement |
| Custom holdout (DSP or in-house) | DSP flags control users. Server-side conversion match. | Mid to enterprise volume | Multi-DSP, OTT, programmatic |
Meta Conversion Lift is the fastest path. Setup takes 10 minutes inside Experiments. Google Conversion Lift requires a rep to enable and runs at the account level, which means every campaign in the account is part of the test. Custom holdouts in The Trade Desk or DV360 give the most control but require manual conversion stitching.
Reading the report
Four numbers carry the readout. Lift percent, p-value, incremental conversions, and effective cost per incremental action.
Lift percent. The relative gap between exposed and holdout conversion rates. A lift of 18 percent means the ad drove 18 percent more conversions than would have happened without it.
P-value. Below 0.05 means the lift is statistically significant. Between 0.05 and 0.10 is borderline. Above 0.10 is inconclusive.
Incremental conversions. The absolute count of extra conversions attributable to the ad. This is the number that goes into the iROAS calculation.
eCPiA (effective cost per incremental action). Total spend divided by incremental conversions. eCPiA is the honest CPA. It strips out the conversions the ad would have gotten for free.
When to run a conversion lift study
Run a lift test when the budget question is bigger than the test cost. Three triggers usually justify it.
- Quarterly budget reviews. Before reallocating $100k plus across channels, prove which channels actually drive lift.
- New channel launches. A first test on TikTok or YouTube tells you whether the channel earns its seat.
- Brand or retargeting audits. These two slots flatter reported ROAS the most. They almost always need a lift check.
[PERSONAL EXPERIENCE] Teams running their first Meta Conversion Lift test usually underestimate the holdout cost. A 10 percent holdout on $300k of monthly spend means $30k of foregone exposure during the test window. That is the price of the truth.
Avoid running lift tests during peak seasonal weeks. Black Friday, holiday spikes, and product-launch windows distort baseline conversion rates and inflate noise.
Real-world example with numbers
[ORIGINAL DATA] A subscription meal-kit brand spent $240,000 a month on Meta. Reported multi-touch attribution ROAS sat at 3.8. The growth team wanted to defend the budget against finance.
They ran a 21-day Meta Conversion Lift study. Audience split was 85 percent treatment, 15 percent holdout. Conversion event was first-month subscription start.
Results:
| Metric | Treatment | Holdout |
|---|---|---|
| Users | 4.2M | 740k |
| Subscribe rate | 1.92 percent | 1.51 percent |
| Lift | 27 percent | - |
| 95 percent CI | 19 to 35 percent | - |
| P-value | 0.003 | - |
| Incremental subscriptions | 3,030 | - |
| eCPiA | $55 | - |
Reported ROAS was 3.8. Incremental ROAS came in at 1.7. Still profitable, but well below the dashboard number. The team kept Meta running, then tested a 15 percent budget shift to YouTube the following quarter.
Conversion lift in 2026
Conversion lift adoption has climbed every year since iOS 14. As deterministic attribution decays, lift studies have become the default trust signal. Three shifts now define the 2026 landscape.
Always-on lift. Platforms like INCRMNTAL and Haus now run continuous synthetic-control tests instead of fixed-window holdouts. The tradeoff is less statistical purity but no opportunity-cost penalty.
Lift inside MMM. Marketing mix modeling vendors increasingly accept conversion lift readouts as priors that calibrate their models. Google's open-source Meridian library was the catalyst.
Privacy-preserving lift. Meta's Private Lift Service uses multiparty computation to run lift studies without exposing user-level data. It became table stakes for regulated advertisers in 2025.
The headline does not change. A holdout is the only way to know what an ad caused. Everything else is correlation in a nice dashboard.
Related terms
Frequently asked questions
What is the difference between conversion lift and brand lift?
Conversion lift measures bottom-funnel actions like purchases, sign-ups, or app installs. Brand lift measures upper-funnel shifts in ad recall, awareness, and intent through survey panels. Both use a holdout. Conversion lift reads behavior. Brand lift reads attitudes.
How is conversion lift different from incrementality testing?
Conversion lift is one design inside the broader category of incrementality testing. It uses a user-level holdout inside a single platform. Geo holdouts, ghost ads, and synthetic controls are other incrementality designs that work across channels.
What sample size do you need for a conversion lift test?
Meta recommends a minimum of 100 incremental conversions for stable readouts (Meta Business Help). Most tests need 500 plus conversions across treatment and control to hit 80 percent power on a 5 to 10 percent lift. Below that, results are noisy.
Why does reported ROAS look higher than incremental ROAS?
Reported ROAS counts every conversion the ad touched. Incremental ROAS counts only conversions the ad caused. Buyers who would have converted anyway inflate reported ROAS. Brand search and retargeting are the worst offenders. The gap is often 30 to 80 percent.
Can you run conversion lift on a small budget?
Not really. Self-serve Meta Conversion Lift requires enough spend to clear a minimum daily incremental conversion rate. Below roughly $30,000 monthly spend, the holdout starves the test of statistical power. Smaller advertisers should run pre-post or geo tests instead.