Measure the purchase-rate difference against a holdout.
Conversion lift starts with the difference between the treatment and control purchase rates. Multiply that difference by the treatment population to estimate incremental purchasers. A confidence interval shows the sampling uncertainty around the difference. For advertising incrementality, use a valid randomised holdout with comparable outcome measurement in both groups.
Enter the number of people assigned to each group and the unique purchasers observed in each group over the same window. Include assigned treatment people who never saw an ad. Comparing only people who were reached with a no-ad control changes the population and can introduce selection bias.
Use people and purchasers, not clicks and attributed orders
This is a two-group purchaser-rate calculation. Each person contributes one observation and can be counted as a purchaser once. Multiple orders from one person do not become multiple purchasers. A platform's attributed purchases divided by landing page views is a different metric. Use our attribution gap diagnostic to reconcile reporting sources before analysing an experiment.
For a non-randomised comparison, select that study design. The calculator shows the observed rate difference and its sampling interval, while leaving incremental CPA and incremental ROAS unavailable. Statistical precision cannot remove differences in audience selection, seasonality or measurement.
The lift and return equations
Treatment rate = treatment purchasers ÷ treatment assigned people. Calculate the control rate from its own assigned population. Absolute lift = treatment rate − control rate. Multiply by 100 to express the difference in percentage points.
Relative lift = absolute lift ÷ control rate. Relative lift is unavailable when the control rate is zero. Absolute lift remains defined.
Incremental purchasers = absolute lift × treatment population. This scales the comparison to the people actually assigned to treatment, even when treatment and control groups are different sizes.
Incremental CPA = additional treatment spend ÷ positive incremental purchasers. Additional spend is what treatment spent above the control policy on the treatment population. Incremental ROAS = incremental purchasers × assumed revenue per purchaser ÷ additional spend. Revenue per additional purchaser is an explicit valuation assumption. It can be first-order AOV when each purchaser is valued at one order. It does not independently measure a revenue effect.
Read the 95% interval
The calculator constructs a Wilson score interval for each purchase rate, using z = 1.959963984540054. For the rate difference d = pT − pC, the Newcombe lower bound is d − √[(pT − LT)² + (UC − pC)²]. The upper bound is d + √[(UT − pT)² + (pC − LC)²], where L and U are the individual Wilson bounds.
An interval entirely above zero supports a positive rate difference under the study assumptions. An interval spanning zero is compatible with no difference as well as effects on either side. An interval entirely below zero supports a negative difference. This is a fixed-analysis interval; repeated peeking, selecting the best of many tests, clustered assignment and missing outcomes require additional methods.
A synthetic holdout example
Assign 100,000 people to treatment and 50,000 to control. Observe 2,500 treatment purchasers and 1,000 control purchasers. Purchase rates are 2.5% and 2.0%, an absolute difference of 0.5 percentage points and relative lift of 25%.
The control rate predicts 2,000 purchasers across a population the size of treatment. The treatment group has 500 more, giving 500 estimated incremental purchasers. The 95% difference interval is approximately 0.342 to 0.655 percentage points, or 342 to 655 incremental purchasers when scaled to treatment.
With $20,000 of additional treatment ad spend, point-estimate incremental CPA is $40. Valuing each additional purchaser at $100 gives $50,000 of modelled incremental revenue and 2.5× incremental ROAS. These return figures inherit uncertainty in the purchaser effect and the value assumption.
Before changing budgets, check the assignment, holdout integrity, observation window and outcome coverage. Read our Meta attribution guide for the difference between platform reporting windows and an experiment, or use the creative testing planner for a production and testing schedule.
Method references checked 16 November 2026: NIST's Wilson interval equations, UCL's Newcombe-Wilson difference method, and Meta's Conversion Lift methodology, which uses randomised eligible groups and Bayesian inference in its own reporting.
Conversion lift questions
What is the difference between absolute and relative lift?
A rise from 2% to 2.5% is a 0.5 percentage point absolute increase. Dividing that increase by the original 2% gives 25% relative lift. Keep the units attached when sharing a result.
Can I use unequal treatment and control groups?
Yes. The calculator compares rates using each group's own size, then scales the difference to the treatment population. Enter everyone assigned, not only reached people or purchasers.
What happens when the control group has no purchases?
The absolute rate difference and its Wilson-based interval remain available. Relative lift has a zero denominator, so it is unavailable. A small or zero-event control group can leave considerable uncertainty even when the point estimate looks large.
Why is incremental CPA unavailable for negative lift?
There are no positive estimated incremental purchasers to divide spend across. A negative cost per acquisition would be misleading. Negative incremental revenue and ROAS remain visible when spend and purchaser value are supplied.
Will this reproduce Meta's Conversion Lift report?
Meta uses Bayesian inference in its Conversion Lift reporting. This calculator uses a frequentist Newcombe-Wilson interval on independent purchaser counts. Platform studies can also apply measurement and modelling adjustments. Use the official study output for that study's reported estimates and settings.
Can an observational comparison prove that ads caused sales?
No. Groups selected by ad exposure, clicks, geography or different time periods can differ before advertising. Choose the non-randomised mode to describe the rate difference, then design a suitable randomised study before treating it as incremental advertising impact.