Google's Conversion Lift runs a real randomized controlled trial on your own campaigns, which is more than any attribution model can claim. It's the platform-native way to measure what your Google ads actually add.
Here's how to set one up, how to read the result, and how to avoid the tests that prove nothing.
What Conversion Lift actually does
It splits your audience into a test group that's eligible to see your ads and a control group that's held back, then compares conversion rates between them. The gap is the lift: conversions caused by the ads, not merely correlated with them.
Google runs it user-based, using ghost-ad-style withholding so the control is selected the same way as the test group. There are also geo-based lift designs for cases where a user-level split isn't feasible.
What you need before you start
- Enough conversion volume and spend, since small tests never reach significance. Lift studies have historically leaned toward larger advertisers and often involve your Google rep.
- A clearly defined conversion that fires reliably, because you're measuring its rate.
- A plan to leave the test alone for its full run, usually a couple of weeks or more.
Setting it up
- Confirm eligibility and the study type with your account or Google rep.
- Pick the campaigns and the conversion action you want to measure.
- Set the holdout size: big enough for significance, small enough that you're not withholding too much revenue.
- Launch and freeze the setup, so mid-test changes don't contaminate the groups.
- Run it for the full window without peeking and stopping early.
Reading the result
The output is incremental conversions and a lift percentage, with a confidence interval. The interval is the part that matters: a headline "plus 12 percent" whose range crosses zero is not a result, it's an underpowered test.
Compare the measured lift to what your attribution claimed for the same campaigns. When platform attribution says a channel is a hero and the lift test says it barely moves the needle, believe the test.
Where it breaks
- Too little volume, so the interval is enormous and the test decides nothing.
- Stopping early the moment the number looks good, which is how noise becomes a decision.
- Changing budgets or creative mid-test, contaminating the comparison.
- Measuring a conversion that fires inconsistently, so the rate is meaningless.
User-based or geo-based
Google offers two flavors, and the difference matters. The user-based study splits individual people into test and control and is the more precise option, but it needs scale and Google's involvement.
The geo-based version splits regions instead, which works when a user-level split isn't available and you can tolerate coarser results. If you can't get a user-based study, a geo design, including a DIY one, is the fallback.
How it fits your other measurement
Treat a lift study as the calibration layer, not a standalone report you run once. The number it produces is the ground truth you feed back into your MMM as a prior and use to sanity-check platform attribution.
Run it periodically, because a channel's true lift drifts as your audience saturates and creative fatigues. A single lift test from a year ago is a historical footnote, not a current fact.
And budget for the holdout honestly. Withholding ads from a slice of your audience costs you those conversions, which feels bad but is the price of knowing whether the rest of the spend even works.
Google's Conversion Lift is a real randomized trial, which is exactly why it beats any attribution number for the question of what your ads add. Give it enough volume, leave it alone, read the interval, and trust it over the dashboard when the two disagree.
Want a stronger data analyst role or a raise? Grab the FREE Product Analyst Playbook and get the exact roadmap to your next offer.
