Base·Note Data

Blog · Incrementality · Meta

How Meta's Conversion Lift works

How Meta's Conversion Lift test splits your audience, what you need to run one, what the results mean, and why 2026's attribution changes make it more useful.

Incrementality testing by platform: Overview · Meta · Google · TikTok · Amazon · Pinterest · Snapchat

Meta's Conversion Lift randomly splits your audience, hides your ads from one group, and counts how many extra conversions the other group produced. That difference is the lift your ads actually caused, not just the sales they were credited with.

The idea in one picture

Meta hides your ad from a random holdout and counts the extra conversions Eligible peoplerandom assignment Test groupads delivered as normal Control groupyour ad is not shown Liftextra conversions Third-party guides call a 10 to 20 percent holdout common. Meta designs and scores its own test.

The control group never sees your ad, so any gap in conversions is the ad's doing.

How a study runs

  1. You pick a campaign and a conversion source: Meta Pixel, Conversions API, app events or offline events.
  2. You choose a holdout share and a length. Third-party guides call a 10–20% holdout common.
  3. Meta randomly assigns eligible people to a test group or a control group, and they stay in that group for the whole test. PPC Land describes Meta's approach as auction-based: control people go through the same delivery pipeline, but your ad isn't shown to them.
  4. Both groups are tracked the same way for the whole window.
  5. Meta compares conversions in the two groups and reports the lift.

What you need

  • Access: if your account qualifies, you set it up in Experiments in Ads Manager. Meta's setup button can be disabled if you don't meet the spend or conversion requirements, and Meta's API guide still points advertisers to their Meta rep (Tofu, quoting Meta's help pages).
  • Budget and length: Meta's setup page recommends at least $5,000 and at least 28 days. Haus reports a hard minimum of seven days.
  • Conversions: guides citing Meta's eligibility rules ask for a campaign started in the past year with $5,000 in spend and 500 conversions (Triple Whale). Pick one with stable volume, not one still in the learning phase (AdBeacon).
  • Data: tests on standard web events require the Conversions API.
  • Audience: at least 10% of the audience in each test cell. Meta may reject a study that misses volume thresholds, and a test can raise costs by shrinking the audience (Haus).
  • Breakdowns need at least 100 conversions across test and control combined before results display.

What you get back

Meta reports lift %, incremental conversions, a confidence interval, incremental ROAS (iROAS) and cost per incremental conversion.

  • iROAS is incremental revenue divided by ad spend.
  • Incrementality factor is incremental conversions divided by platform-reported conversions. A factor of 0.6 means 60% of reported conversions were truly incremental. Multiply platform ROAS by it to get a calibrated ROAS (Haus).
  • A wide confidence interval usually means the holdout was too small. A neutral result is still useful: it may mean budget belongs elsewhere.

Why this matters more in 2026

Meta changed what its attributed numbers count three times this year, which makes a lift test the steadier benchmark:

  • January 12, 2026: the Ads Insights API stopped returning the 7-day and 28-day view windows (Lead Metrics).
  • March 2026: only link clicks count as clicks for website and in-store conversions. Likes, saves and shares moved to a separate engage-through category.
  • November 12, 2026: a new Pixel setting, "Measure conversions on customer devices," began rolling out switched on by default. It credits campaigns with purchases measured on shoppers' own devices. It isn't active in the EU or UK, and you can turn it off per dataset in Events Manager (PPC Land).

None of these changes how many people bought. If your Meta ROAS jumps in late November, a holdout test is how you find out whether sales did too.

Two related Meta tools

  • GeoLift is a free, open-source R package from Meta's Marketing Science team. It compares regions instead of people. Meta suggests it when people-based measurement isn't feasible, such as for healthcare advertisers, and notes people-based tests have higher statistical power (GeoLift docs).
  • Incremental Attribution is a model, not an experiment. It is trained on Conversion Lift data to estimate which conversions your ads caused. PPC Land dates its global rollout to June 2025, and reports Meta's own analysis of 307 studies found rules-based attribution undervalued Meta by a median 31% (PPC Land).

Mistakes to avoid

  • Ending the test early on interim results.
  • Changing creative, budget or promotions mid-test, or running through a seasonal peak.
  • Treating a neutral result as proof the campaign doesn't work.
  • Forgetting the scope: a Meta study measures Meta only, and Meta designs and scores its own test.

I design and read incrementality tests for brands, on the platforms and independently of them. See how a Measurement Audit works.

Much of Meta's help center isn't publicly readable, so parts of this post rely on third-party summaries that quote Meta's guidance.

Sources

More posts Services and pricing

Work with me

Want a second set of eyes on your measurement?

I help marketing and growth teams figure out what their spend is actually driving, with MMM, incrementality testing and attribution that finance trusts too.

Book a 30-min call