A brand lift study is a randomized experiment that surveys two groups, people who were shown your ads and a control group who weren’t, and compares how many of each give the positive answer to a question like “which of these brands have you seen advertised recently?” The difference is your brand lift in ad recall, awareness or consideration. A conversion lift study uses the same random split but counts conversions instead of survey answers, so it estimates how many sales the ads actually caused. Both are offered by Meta and Google, both are genuine experiments, and both are far more credible than the attributed numbers in the same dashboards.
For most advertisers a platform lift study is the closest thing to ground truth they’ll get for free, and it’s quite normal for it to tell a different story from the attribution report in the same account. That’s exactly why lift studies are useful, but it’s worth understanding what’s going on underneath before you put a result in front of a CFO, starting with the intent-to-treat dilution that trips up most first readings.
How a conversion lift study works
Before the campaign runs, the platform randomly assigns everyone in your target audience to a test group or a control group (the holdout). People in the test group can be shown your ads as normal; people in the control group are prevented from seeing them. Because the assignment is random, the two groups are the same on average in everything that matters (age, interests, how close they were to buying), so any difference in their conversion rates after the campaign is the effect of the ads. Meta’s own description says incrementality is best measured by a randomized experiment, and that’s what this is.
The conversions are counted from whatever the platform can see: the pixel, a server-side API, app events, or offline conversions you upload. The result is usually reported as a relative lift (the test group’s conversion rate divided by the control’s, minus one), a number of incremental conversions, and a cost per incremental conversion. To turn that into incremental ROAS, multiply the incremental conversions by your value per conversion and divide by spend, as in the incrementality calculator. The design choices behind a good holdout (how big a control group, how long to run, what not to change mid-test) are covered in the guide to incrementality testing.
Meta Conversion Lift and Brand Lift
Meta offers both Conversion Lift and Brand Lift tests. Meta says advertisers can set up a test themselves through the self-serve option in Ads Manager if they meet its criteria. Meta doesn’t publish a single list of minimums on its Conversion Lift page, and the eligibility checks happen inside the Experiments tool, so the honest advice as of October 2026 is to open Experiments in Ads Manager and see what it lets you create rather than relying on thresholds quoted on agency blogs.
In practice, the things that decide whether a Meta lift study will be informative are the same ones that matter in any experiment: enough conversions in the test population for the effect you’re hoping to see, a conversion event that’s passed reliably (ideally server-side through the Conversions API, so it isn’t lost to browser tracking restrictions), and a campaign that you can leave alone for the test’s duration. A lift study on a campaign that gets 30 conversions a week will almost always come back inconclusive, which isn’t the platform’s fault.
Google Ads Conversion Lift and Brand Lift
Google offers four lift measurement types in Google Ads, and its help page comparing them is the best place to check current availability:
- Brand Lift: survey-based, measuring ad recall, awareness, consideration, favorability and purchase intent, for Video and Demand Gen campaigns on YouTube.
- Search Lift: measures how YouTube ads change searches for your brand or product, with budget minimums Google describes as similar to Brand Lift.
- Conversion Lift based on users: a user-level holdout measuring website visits, purchases or app downloads. Google lists Video, Discovery and Demand Gen campaigns as self-serve in Google Ads, with other campaign types requiring your account representative. Google notes it relies mostly on opt-in traffic, so there can be measurement gaps.
- Conversion Lift based on geography: a regional split that supports online and offline conversions and cross-channel measurement, for a wider range of campaign types, through your account representative. Google says it’s noisier, can’t be broken down by demographics, and tends to need higher budgets than the user-based version.
For Brand Lift, Google publishes its minimum budgets. Over a 10-day window, campaigns measuring the same product or brand must together meet $10,000 for one survey question, $20,000 for two and $60,000 for three in its “Country B” group, which includes the United States and the United Kingdom (cheaper and more expensive country groups exist). Surveys take up to 48 hours to be approved, and Google says a standard study completes within 3 to 14 days (9 to 14 for its Enhanced Lift option), often ending early once it reaches its target number of responses. Google also reports three versions of the result: absolute lift (the difference in positive-answer rates), relative lift (absolute lift divided by the control’s rate) and headroom lift (absolute lift divided by the share of control respondents who didn’t give the positive answer, which tells you how much of the available room you captured).
Intent to treat vs treatment on the treated
There’s a subtlety that catches most people the first time they read a lift study. Not everyone assigned to the test group actually sees an ad: some don’t use the app during the campaign, some are outbid in the auction, and some simply aren’t picked by the delivery system before the budget runs out. If the study compares everyone assigned to test with everyone assigned to control, that’s an intent-to-treat (ITT) comparison, and its lift is diluted by all the people in the test group who were never exposed. The effect on the people who were actually shown an ad, the treatment on the treated (TOT), is larger.
The standard adjustment is simple if you’re willing to assume that ads have no effect on people who never saw one: divide the ITT absolute lift by the share of the test group that was exposed. So if 30 percent of the test group saw an ad and the ITT lift is 0.048 percentage points, the lift among the exposed is 0.16 points. Converting that to a relative lift needs one more number, the baseline conversion rate of the people who were exposed, and here there’s a catch. The delivery algorithm chooses who to show ads to, and it tends to choose people who were more likely to convert anyway, so the exposed group’s baseline is often higher than the group average. That’s why the calculator below asks for a multiplier rather than assuming 1.0.
The dilution also has a cost in precision. A comparison of everyone assigned carries the noise of all the unexposed people, who contribute random conversions but no effect, so an ITT design needs a considerably larger sample than one that can isolate the would-be-exposed users in the control group. That’s the problem ghost ads were invented to solve.
Ghost ads, PSA tests and predicted exposure
The traditional way to know who in the control group would have been exposed was a PSA test, where control users are shown an unrelated public service announcement in your ad’s place. You then compare exposed test users with PSA-exposed control users. The trouble is that you pay for the PSA impressions, and an optimizing ad system may deliver the PSA to different people than it would have delivered your ad, which breaks the comparison. Johnson, Lewis and Nubbemeyer’s ghost ads method, published in the Journal of Marketing Research in 2017, instead has the ad system record the control users who would have seen your ad (the “ghost” impressions) without showing anything. The authors reported that this reduces the cost of experimentation and improves measurement precision relative to both PSA and intent-to-treat designs, and that a predicted version works with platforms that optimize delivery in real time.
When you read a lift study, the most useful question to ask is which comparison it reports. If it’s ITT, the lift will look small and the interval wide, and you can use the calculator above to translate. If it compares exposed users with would-be-exposed controls, the lift applies to the people the campaign reached, and you shouldn’t apply it to the whole audience. Either is fine as long as the arithmetic downstream (incremental conversions, iROAS, any calibration of your marketing mix model) uses the same definition.
How many responses a brand lift study needs
Brand lift studies have a sample size problem of their own, because they depend on survey responses rather than conversions. A survey answer is a yes or no, so detecting a two- or three-point change in ad recall on a 20 percent baseline needs thousands of responses split between exposed and control. Google’s own guidance is that a high-performing campaign can expect to detect lift at about 2,000 responses per metric, a campaign at the recommended budget minimum at about 4,100, and that if there’s no lift by 16,800 responses per metric you may be unable to detect one. Google also says its lift results use 80 percent two-sided confidence intervals, which is a lower bar than the 95 percent most analysts default to, and worth knowing before you compare a brand lift result with a conversion test.
The calculator assumes an equal split between exposed and control respondents and 80 percent power, so treat it as a rough guide to what’s detectable rather than a reconstruction of Google’s method. The useful thing it shows is how sharply the detectable lift falls as responses grow at first and how slowly after that: going from 2,000 to 4,100 responses buys a lot more sensitivity than going from 10,000 to 16,800. If your campaign can only generate a few thousand responses, ask one question rather than three, both because the minimum budget is lower and because every extra question needs its own responses.
Why platforms run lift studies, and their limits
Platforms offer free experiments because a credible lift result is the best argument for spending more with them, and because advertisers increasingly ask for one. That doesn’t make the results wrong. These are randomized experiments, and research using Facebook’s lift studies (Gordon, Zettelmeyer, Bhargava and Chapsky, in Marketing Science) treated them as the benchmark against which observational methods were judged, and found those methods often failed to recover the experimental effect. If the choice is between a platform lift study and platform attribution, the lift study wins comfortably; the guide to Facebook experiments vs observational methods goes into that paper in detail.
The limits are about scope and independence rather than validity. The platform designs the test, runs it, and grades its own homework, using a methodology you can’t fully inspect, so you’re trusting its choices about which users are eligible, how exposure is defined and which conversions count. It only measures conversions it can see, which means retail sales, phone orders and purchases on untracked devices are missing (a geo-based study or your own geo lift test gets around this). It only measures that platform, so it can’t tell you whether Meta is doing better than YouTube, and two platforms’ studies may count the same customer as incremental to each. And it measures the campaign it tested, in the weeks it ran, at the budget it had, which is a narrower question than “what is this channel worth to us?”
Using lift studies with your marketing mix model
The most valuable thing you can do with a lift study is often not to act on it directly but to use it to check and sharpen a broader model. A marketing mix model covers every channel but rests on assumptions; a lift study covers one campaign but rests on randomization. Feeding the second into the first, with care about whether they measure the same thing, is what calibration means, and it’s covered in the guide to calibrating MMM with experiments. Robyn’s demo script, for example, names identity-based tests like Facebook conversion lift as one of the usual experiment types to calibrate with, and recommends treating their result as the immediate effect rather than the total including carryover.
If you’re deciding whether a platform study is enough or you need an independent test, the guide to statistical power for marketing tests will tell you whether either can detect the effect you’re looking for, and the Vexpower MMM course works through how a lift result, a geo test and a model can be combined on one invented brand without counting the same evidence twice.
