Geo lift vs conversion lift: how to choose

Geo lift and conversion lift both measure how many conversions advertising caused, but they randomize different units. Geo lift compares regions, so it needs no user-level exposure data from the publisher. The trade-off is lower statistical power and less detail, because only a limited number of comparable regions exist. Conversion lift, on the other hand, randomizes individual users. This then gives larger samples and campaign-level results, at the cost of needing exposure data from the publisher.

A Decentriq data clean room makes the exposure data usable without the publisher handing it over. Exposure data and conversion data are matched without either party accessing the other's raw records, so conversion lift can still run.

Geo lift suits channels without user identifiers, such as linear TV and out-of-home, and questions that only need a channel-level answer. Conversion lift suits questions about specific campaigns or audiences, provided exposure data can be brought into the measurement.

What geo lift and conversion lift measure

Both methods are forms of incrementality testing: they compare an exposed group with a control group to estimate the conversions that would not have happened without the advertising. They differ in how the groups are formed: geo lift puts whole regions into the test or control group, and conversion lift puts individual users into them. For background on the broader concept, see Decentriq's guide to incrementality in marketing.

Geo lift assigns whole geographic regions, such as cities or states. Advertising runs in the test regions and is withheld or held steady in the control regions, and the difference in aggregate outcomes such as sales or sign-ups indicates the lift. Modern designs often build the control from a weighted mix of regions, known as a synthetic control, instead of pairing similar markets by hand.

Conversion lift assigns individual users. Eligible users are randomly assigned to a test group that can see the ads and a control group that cannot, and the difference in conversion rates between the two groups is the lift. Meta and Google offer versions of this test inside their own platforms. Because assignment is random at the user level, it is a true randomized controlled trial.

How geo lift works and where it fits

A geo lift test starts with a design phase. Analysts select the regions that will receive the advertising and build a control from the remaining regions using historical data. A power analysis then confirms the test can detect an effect of a commercially relevant size. The campaign then runs for a fixed period, usually several weeks, and the outcome in the test regions is compared with what the control predicts.

The main appeal is data minimalism. The method needs only aggregated, region-level outcomes, so it does not depend on user identifiers or on exposure logs from the publisher. That makes it compatible with strict privacy requirements, and it works for channels that leave no user-level trail, such as linear TV and out-of-home. It can also measure the combined effect of several channels at once, which a single-platform user-level test cannot do.

Open-source tooling, including Meta's GeoLift library, has lowered the setup effort, and geo lift results are widely used to calibrate marketing mix models. For large budget lines where a channel-level answer is enough, it is a sound and practical choice.

Where geo lift falls short

Geo lift trades detail and precision for independence from user-level data. Five limits matter most when choosing between the two methods.

Statistical power

A geo test has as many observations as it has regions, typically from a few dozen to a few hundred. A market divides into only so many regions that are comparable to each other and far enough apart to limit spillover. A user-level test can draw on thousands or millions of users. Fewer units mean noisier estimates and wider confidence intervals, so every geo test has a minimum detectable effect: a true lift below that threshold cannot be told apart from noise. For channels where the real lift is modest, that threshold can be hard to clear. Meta's GeoLift documentation recommends people-based experimentation where it is feasible, because it has higher statistical power than geo-based experimentation.

Spillover between regions

Region boundaries are porous. People commute and travel, and some media reaches audiences outside the intended targeting. When people in control regions see the advertising, the measured difference shrinks, so spillover generally leads to underestimated lift. Grouping adjacent regions into clusters reduces the problem, but it also reduces the number of units available.

Limited granularity

Geo lift answers whether a channel, or a combination of channels, produced incremental outcomes. It generally cannot compare creatives or audience segments, because splitting an already small number of regions further removes what little power the test has. Decisions about individual campaigns usually need a method with more units.

Opportunity cost and duration

Both methods withhold advertising from a control group, which means forgoing some sales. In a geo test the holdout is entire regions, so the sales at stake are larger than in a user-level test, where the control is usually a fraction of the audience. Each geo test also uses up a set of regions and a time window, which limits how many questions can be answered in a year. 

Data history and consistency

Credible designs depend on a stable history of region-level outcomes, measured the same way in every region. Seasonality and local promotions that differ by region can bias the comparison if the design does not account for them.

Each of these limits narrows the range of questions geo lift can answer, which is why it is often paired with a user-level method for finer-grained decisions.

Conversion lift and the exposure data problem

Conversion lift addresses most of the limits above. Randomizing at the user level gives far more units than a geo test, so smaller effects can be detected, and results can be broken out by campaign or audience segment. Reads also tend to arrive faster, since a study does not need dark markets or a long design phase.

Limits of platform-run conversion lift

The standard versions of this test run inside a single media platform, using that platform's exposure data and methodology. Results are comparable across platforms only with effort, because each platform defines and runs its study differently, and platforms can limit how many studies run at once or require cool-off periods between them. Because the platform also measures the effect of its own advertising, some marketers want an independent check.

Platform tests also depend on the conversions the platform can observe. Google notes that its user-based conversion lift relies mostly on opt-in traffic, so measurement gaps can affect results. Google also offers a geography-based version, which it describes as noisier than the user-based one, with no demographic breakdowns and a budget requirement that tends to be higher. That version shares the limits described above for geo lift.

The exposure data barrier

The underlying obstacle is where the data sits. The publisher holds the record of who was eligible to see the ads and who was held back as a control, and the advertiser holds the record of who converted. A user-level lift study needs both records matched, and in practice that means one side handing raw data to the other. Privacy requirements and competitive concerns often rule that out. Methodology is a second obstacle: each platform runs conversion lift inside its own walled garden, with its own definitions, so results are hard to compare across platforms. A single geo design applies the same method to every channel. Together, these are the main reasons teams fall back on geo lift even when they need user-level precision. 

Running conversion lift in a data clean room

A data clean room removes the need for a handover. In a Decentriq data clean room, the publisher provisions exposure data, the advertiser provisions conversion data, and the two are matched and compared inside a confidential computing environment. Neither party sees the other's raw records, and only the aggregate lift result is released. When the study concludes, the clean room is shut down and both parties' data is deleted. IAB Tech Lab publishes guidance and recommended practices for data clean rooms covering this and other topics.

The measurement is run by a neutral third party instead of the media platform, and it does not require the publisher to operate its own clean room infrastructure. Not every conversion lift study needs a clean room: where the advertiser and publisher can already collaborate on the data directly, a platform-run test may be simpler. The clean room matters when exposure and conversion data sit with different companies and neither can pass raw records to the other. The clean room matters when exposure and conversion data sit with different companies and neither can pass raw records to the other. It also gives a single study design that can run across publishers. This way, results are comparable as opposed to each platform’s own tool applying its own method. Decentriq's article on conversion lift measurement in clean rooms walks through the full process.

Geo lift and conversion lift compared

The table compares geo lift with two ways of running conversion lift. Each column is a method, and each row is a factor that usually decides the choice.

How to choose between geo lift and conversion lift

The right method depends on the channel, the question, and which data can be brought into the measurement.

Running both is the most robust option when the stakes justify it. A geo lift study can test whether a conversion lift program shows systematic bias, and a user-level study can fill in the detail that a geo test cannot resolve. When the two disagree, the usual cause is a design problem in one of them, such as low power or spillover, so the first step is to review both designs before trusting either result.

Frequently asked questions

Is geo lift less accurate than conversion lift?

Accuracy depends on the design of each test. Geo lift has fewer units and is exposed to spillover, so its estimates carry wider uncertainty and tend to understate lift. Conversion lift has far more units but relies on exposure data and, in platform-run versions, on the platform's own methodology.

When is geo lift the better choice?

Geo lift fits channels with no user-level trail, such as linear TV and out-of-home, and questions about the combined effect of a large budget. It is also the practical option when exposure data cannot be obtained at all.

Can geo lift and conversion lift be used together?

Yes. Many teams run geo lift for channel-level totals and conversion lift for campaign-level detail, then compare the results. Both types of result can also be used to calibrate a marketing mix model.

Is a data clean room required to run conversion lift?

No. Platform-run studies work without one. A clean room becomes useful in two situations: when exposure data and conversion data sit with different companies and neither can pass raw records to the other, and when consistency matters, since one study design can run across publishers while each platform’s own tool applies its own method. A Decentriq data clean room matches the datasets so that conversion lift can still run. 

How does a Decentriq data clean room protect the data in a lift study?

Data is encrypted and processed inside a hardware-based confidential computing environment, where it is never exposed, even to Decentriq. Neither the publisher nor the advertiser sees the other's raw records. Only the aggregate lift result is released, and all data is deleted when the study ends.

How do these tests relate to marketing mix modeling?

Both methods give experimental estimates of incremental return that can anchor a marketing mix model. Conversion lift results can inform campaign-level assumptions, while geo lift results can constrain the total for a channel or platform.