Measuring install lift: the growth marketer's incrementality guide

Every growth team can tell you their cost per install. Most can tell you their attributed ROAS. Far fewer can answer the question 'which of those installs would have happened anyway?' — and that is the question your CFO is actually asking.

A working definition of lift

Install lift is the difference between attributed installs in a treated cohort and modeled organic baseline in an untreated cohort, measured during an identical time window with identical seasonality.

If you can't articulate the untreated cohort and the time window, you do not have a lift number. You have a directional comfort blanket.

What's the difference between attribution and incrementality?

Attribution assigns credit for installs that happened. Incrementality asks how many of them would have happened without the spend. They answer different questions, and a campaign can score well on the first while being worthless on the second — retargeting an existing user who was going to install anyway attributes beautifully and adds nothing.

You need both. Attribution runs the week; incrementality decides the quarter.

How do you run a clean install lift test?

1. 1Pick one channel and one geography. Testing two variables at once produces a result you cannot attribute to either.
2. 2Define the holdout before launch — a geography, an audience segment, or a randomised share of the target audience the platform will genuinely exclude.
3. 3Size it so the expected effect is detectable. A 5% lift on 200 installs a week is not measurable; either widen the test or accept you are measuring nothing.
4. 4Run for at least one full purchase cycle, and never across a seasonal boundary. A test spanning Black Friday measures Black Friday.
5. 5Measure a downstream event, not the install. Installs are the cheapest thing to move and the least informative.
6. 6Compare treated against holdout on that event, then divide the incremental conversions by the spend to get a true cost per incremental action.

Three places your numbers are lying

• You're attributing organic search installs to brand campaigns. Brand search is the worst offender — set up a holdout, not a regression.
• You're double-counting between MMP and SKAN. Run them in parallel, but pick one for headline KPIs.
• You're conflating install with first event. An install that doesn't open the app twice is not a customer.

Why do lift results disagree with the ad platform's?

Because the platform is grading its own homework, and its holdout is not your holdout. Platform-reported lift is computed inside the platform, against a control group it selected, using its own conversion data. Every incentive points the same direction, and none of that requires bad faith to produce optimistic numbers.

The specific mechanism is usually the control group. Platforms tend to hold out users they judged unlikely to convert, which makes the treated group look better for reasons that have nothing to do with your ads. A geographic holdout you define yourself has none of that structure — which is why it will nearly always report lower lift, and why it is the number to take to a budget conversation.