Conversion lift is a platform-native randomized experiment: users (not geos) are the unit of randomization, and the holdout is enforced by the ad platform’s delivery system. Meta, Google, and TikTok all ship lift products under that name. The conceptual formula is lift = (test conversion rate − control conversion rate) / control conversion rate, and the dollar form is incremental conversions = (test rate − control rate) × test audience size.
The point of running one is the counterfactual that attribution cannot honestly answer: would this customer have bought anyway? Attribution credits the last click; lift credits only the conversions that wouldn’t have happened without ad exposure. The two numbers usually disagree, and incremental ROAS measured via a holdout is often substantially lower than the ROAS the platform’s attribution reports, because much of the attributed volume — existing buyers, brand demand, retargeting overlap with already-converting users — was going to land regardless.
The practical cost is real. Studies require committing to a holdout, which means accepting lost short-term revenue to learn the answer, and they typically run two to four weeks. They also need a large enough audience for statistical power on a DTC brand; exact thresholds depend on control-group conversion volume and vary by platform, so check the platform’s lift-study requirements before committing. Where a platform won’t or can’t run a user-level holdout, geo-lift is the platform-agnostic, market-randomized alternative. Brand lift applies the same randomized design to upper-funnel survey outcomes.