How to read lift test results without fooling yourself

What a lift test result contains

A lift test result is an estimate of how much the ads raised the conversion rate, with a range showing how uncertain that estimate is.

Lift: (test rate minus control rate) divided by control rate. A 1.2% test rate against a 1.0% control rate is a 20% lift.

Confidence interval: the range of lifts consistent with the data at a chosen confidence level, usually 95%.

Incremental conversions: the rate difference times the number of people in the test group.

Everything else, from incremental CPI (the causal version of cost per install) to iROAS, is built from these.

How do you know if a lift result is real?

Look at the 95% interval. If all of it sits above zero, the result is unlikely to be noise; if it crosses zero, the test cannot distinguish the effect from chance. The p-value says the same thing in another form: below 0.05 matches an interval that excludes zero.

Same lift, different answers

Sample data: a made-up app to show the method, not a benchmark.

Two made-up tests both report a 15% to 20% lift. Only one is usable.

Test 1's headline looks fine, but its interval includes zero and even a small negative effect. Test 2, with five times the control conversions, pins the lift between roughly 10% and 30%. Paste either into the lift test reader to see the working.

Seven mistakes that turn noise into wins

1. Reporting the point estimate alone. "+20%" without its range invites budget decisions the data cannot support. Always show the interval.
2. Stopping when it looks good. Checking daily and stopping at the first significant day inflates false wins. Fix the length before starting; the holdout simulator shows an estimate that clears zero and then slips back.
3. Calling an inconclusive test a failure. An interval of −10% to +40% means the test was too small, not that the ads do nothing.
4. Mixing up percentage lift and percentage points. A 20% lift on a 1% rate is 0.2 percentage points, or 2 extra installs per 1,000 people reached.
5. Comparing test-group incremental installs with campaign-wide attributed installs. Scale to the same population first: incremental installs from the test group cover only the people in the test group.
6. Ignoring contamination. If the control group was reached by another campaign, a retargeting list or the same ads on another device, the measured lift shrinks. Note what else ran.
7. Reading many slices and reporting the best. Split a test into ten audiences and one will look significant by chance. Decide the cuts in advance.

From lift to incrementality

Most teams want a single share: of the installs this campaign was credited with, how many did it cause?

Sample data: a made-up app to show the method, not a benchmark.

Suppose the test group had 2,160 installs and 360 of them were incremental (a 20% lift on a 1% control rate across 180k people). If the attribution tool credited the campaign with 900 of those 2,160 installs, incrementality is 360 ÷ 900, or 40%. The other 540 credited installs would have happened anyway.

The denominator matters. Use attributed installs from the test group over the test period, not the campaign's lifetime totals.

What to report

A one-paragraph read-out that holds up in a budget meeting:

• What was tested: channel, audience, regions or holdout share, dates.
• The result: lift with its 95% interval, and incremental installs.
• The comparison: incrementality against attributed installs, and incremental CPI.
• Caveats: what else changed during the test, and whether the planned length was reached.
• The decision: what you will change, planned on the cautious end of the interval.

A guide to measuring true marketing ROI covers how to carry these numbers into a finance conversation. When platform and MMP numbers disagree before you even start, Attribution Debugger's guide on which number to trust helps settle the baseline.

Platform reports vs your own maths

Platform lift studies sometimes report at confidence levels other than 95%. Google's App Uplift experiments, for example, offer 80%, 85% and 95% and recommend 80% (Google Ads Help, read 26 Oct 2026). Lower confidence gives faster answers and more false positives, so note the level when you compare tests. The platform guide has the details.