How to Run a Meta Conversion Lift Study Step by Step
Set up a Meta Conversion Lift test in Experiments, see how Meta splits test and holdout groups, and read incremental conversions, confidence and lift percent correctly.
On this page
- Who can run a Conversion Lift test?
- How does Meta split people into test and control?
- A setup checklist for the Experiments tool
- What do incremental conversions and the confidence figure mean?
- Intent to treat or treatment on the treated: which one is Meta reporting?
- A worked readout against Ads Manager (hypothetical)
- Is the holdout itself biased?
- Why the dashboard numbers look off
- Sources
To run a Meta Conversion Lift study step by step, confirm the account qualifies, build the test in Meta's Experiments tool, let Meta randomize people into test and holdout groups, leave the setup alone until the end date, then read incremental conversions and their interval beside what Ads Manager attributed over the same dates. Ranking guides blur three details that change the budget call: who qualifies, what confidence measures, and which audience the lift percent uses. Our sources are Meta's documentation and a peer-reviewed study of Facebook lift experiments, read on 6 November 2026. If the concept is new, start with what incrementality is.
Who can run a Conversion Lift test?
The Meta Business Help Center lists one requirement for Conversion Lift in Experiments: "At least one high-quality conversions data source, such as Conversions API". The "$120,000 USD in ad spend in the last 90 days" minimum in the same table sits in the Brand Lift column, and Meta's best practices for Experiments says some Conversion Lift and Brand Lift tests "may display spend minimums or recommended spend amounts when you start creating your test." Creating a test costs nothing extra, though the About Experiments page warns that lift tests can affect CPM, since the control group never sees your ads. The lift study guide on Meta for Developers opens with "Conversion Lift Measurement is currently limited" and sends advertisers to a Meta representative.
If Experiments shows a spend minimum your account cannot reach, regions usually give a cleaner answer than people, as covered in how to run a quarterly geo-holdout test.
How does Meta split people into test and control?
The Meta Conversion Lift setup randomizes your audience into a test group that can see the ads and a holdout (control) group that cannot. The developer guide says Facebook "randomizes the audience for your ads and assigns Accounts Center accounts to either the test or control group." Its control percentage "Defines a holdout percentage of the Accounts Center accounts who will not see ads," and for one test group, treatment plus control must equal 100. For single-cell holdout tests in Experiments, Meta says the control group "will be 10 percent of the size of your test group," scaled up to the test group's size before totals are reported.
Brett Gordon and Florian Zettelmeyer of Kellogg, with Neha Bhargava and Dan Chapsky of Facebook, explain why the control stays comparable while delivery optimizes. Their paper (author copy dated April 12, 2018, hosted by the Kellogg School of Management, published in Marketing Science in 2019) describes how, when the focal ad would win an auction for a control user, the second-place ad is served instead, and how bid upweighting "is applied equally to users in the test and control group." We found no current Meta help page that spells out those auction mechanics.
A setup checklist for the Experiments tool
Screen labels move between releases, so this list tracks decisions.
- Eligibility. A high-quality conversions source such as Conversions API, and any spend minimum Experiments displays when you create the test.
- Scope. We start at account level, which covers all Meta advertising, and test single campaigns later.
- Events. Pick them now. The API guide says "A lift study requires at least one objective. You cannot modify objectives after the study starts running."
- Power. Meta's confidence page suggests an estimated power of 80 percent or higher.
- Dates. The start time must be in the future. Once the study runs, the start time and treatment percentage lock and test-group campaigns cannot be removed; the end date can still be extended.
- Post-test window. The API's observation end time adds a "post test conversion window" in which Meta keeps matching conversions to people in their original groups.
- Freeze list. Our own rule: write down what stays fixed (budgets, bid strategy, exclusions, creative rotation), because every mid-test change blurs what the result describes.
- Comparison baseline. Note each campaign's attribution setting so you can pull matching Ads Manager results later.
What do incremental conversions and the confidence figure mean?
In Meta's Conversion Lift methodology, incremental conversions are the test group's conversions minus the control group's, adjusted for group size, and the confidence figure is the probability that this number is above zero. The lift metrics glossary in the Meta Business Help Center defines incremental conversions as "additional conversions that would not have occurred without the ads being tested, and must come from Meta accounts reached by those ads." The export carries lower and upper bounds "of the 90% credible interval." Confidence is "The probability that ads in this test cell generate incremental conversions greater than zero." Cost per incremental conversion is spend divided by incremental conversions.
Meta's page on confidence in tests, also in the Meta Business Help Center, sets the bar: "For lift tests, a 90 percent or higher confidence percentage represents a statistically reliable result." The 65 percent bar on the same page applies to A/B test winners only. Confidence says the effect is probably above zero and says nothing about its size, so the interval should carry the budget decision. This next point is our own arithmetic: if that interval is equal-tailed (the glossary does not say), any confidence below 95% puts its lower bound at or below zero.
Intent to treat or treatment on the treated: which one is Meta reporting?
Intent to treat (ITT) in a lift study compares everyone assigned to test with everyone assigned to control, while treatment on the treated (TOT, written ATT in the research) estimates the effect on the people the ads reached. The Kellogg-hosted paper derives ATT as the ITT effect divided by the share of the test group exposed, which "undilutes" it. In its Study 4, test and control converted at 0.045% and 0.033%, an ITT lift of 38%. With 37% of the test group exposed, the ATT lift was 73%, with a 95% bootstrapped interval of 49% to 103%.
Meta defines conversion lift percent as "The percentage increase in conversions among Meta accounts exposed to your ads in this test, compared to the estimated number of conversions you would have received without these ads." That is a treatment-on-the-treated figure. The incremental count stays the same under both views; only the denominator moves. Before you set a Meta lift percent beside a geo test or a marketing mix model, compute the ITT-style version: incremental conversions divided by scaled control conversions.
A worked readout against Ads Manager (hypothetical)
Every number below is invented to show the arithmetic. The field names come from Meta's glossary.
| Metric (export field) | Hypothetical value |
|---|---|
| Test group accounts (population_test) | 10,000,000 |
| Control group accounts (population_control) | 1,000,000 |
| Accounts reached (population_reached) | 4,000,000 |
| Spend (amount_spent) | $150,000 |
| Test group purchases (conversions_test) | 30,000 |
| Control purchases scaled to test size (conversions_control_scaled) | 25,000 |
| Purchases from reached accounts (conversions_exposed) | 18,000 |
| Incremental purchases (conversions_incremental) | 5,000 |
| 90% credible interval | 3,400 to 6,600 |
| Confidence (conversions_confidence) | 99% |
| Ads Manager purchases, 7-day click and 1-day view, same campaigns and dates | 12,000 |
How we read it:
- Cost per incremental purchase: $150,000 ÷ 5,000 = $30.00, or $22.73 to $44.12 at the interval ends. Ads Manager shows $150,000 ÷ 12,000 = $12.50 per purchase.
- Lift among reached accounts (Meta's lift percent): baseline 18,000 − 5,000 = 13,000, so 5,000 ÷ 13,000 = 38.5%.
- ITT-style lift across the test group: 5,000 ÷ 25,000 = 20.0%.
- Incremental share: 5,000 ÷ 18,000 = 27.8% of purchases by reached people were caused by the ads.
- Incremental purchases per Ads Manager purchase: 5,000 ÷ 12,000 = 0.42, or 0.28 to 0.55 across the interval.
Ads Manager counts purchases that followed an ad inside the attribution window, which Meta lists as 1 or 7 days after a click and 1 day after a view. The lift test counts purchases the ads caused, and the 0.42 ratio converts one into the other; the multiplier method carries it into weekly reporting. We would budget on the $44.12 end until a second test narrows the range. Meta's scaled control is also "calibrated to ensure greater accuracy," so a real export may differ from our plain 10:1 scaling.
Is the holdout itself biased?
Holdout bias in the Meta Conversion Lift methodology mostly comes from comparing the wrong groups. Inside the test group, heavier users and the people delivery favors are more likely to be exposed, so exposed people tend to convert more often even without the ads. The Kellogg-hosted paper says that comparison will "overstate the effectiveness of advertising"; in Study 4 it gave 316% against the experiment's 73%. In half of its 15 US experiments, the estimated percentage increase in purchases was off by a factor of three across all observational methods. The randomized holdout avoids that trap, with two limits: a result "may not generalize to similar future campaigns if market conditions change," and conversions after the measurement window make the estimate conservative.
Why the dashboard numbers look off
Most Meta Conversion Lift dashboard bias and reporting issues trace back to documented behavior. Meta's results guide says test totals are "not the same as your ad campaign results" because they include every action by people in either group for the whole test. Results can appear after 100 conversion events, but Meta recommends waiting until the test has completed. A campaign appears in the breakdown only "if all ad sets in the campaign have the same result type," and API breakdowns need at least 100 conversions across test and control combined. Ads Manager's incremental attribution model predicts whether a conversion was caused by an ad, and Meta warns that comparing across attribution models "will lead to inaccurate conclusions." We treat the randomized test as the check on that model.
One lift study is one reading at one spend level in one season. Its lasting value is the ratio it sets for everyday reporting, so we pair each test with a review of how the account assigns credit, using what marketing attribution is as shared vocabulary and Attribution Doctor as a further resource. Designing and reading these tests is part of our data analytics work.
Sources
- Meta Business Help Center, About Experiments
- Meta Business Help Center, Best practices to get started with Experiments
- Meta for Developers, Lift Study
- Meta Business Help Center, View and understand holdout test results across Meta technologies
- Meta Business Help Center, Facebook Lift metrics glossary
- Meta Business Help Center, About confidence in your tests and experiments
- Meta Business Help Center, Conversion lift percent
- Meta Business Help Center, About attribution models and attribution settings
- Gordon, Zettelmeyer, Bhargava and Chapsky, A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook