Measuring Meta’s halo: what a conversion lift study told us about our budget

A three-week holdout found that a third of the subscriptions our other paid channels were claiming would not have happened without Meta ads.

Last-click attribution has one great virtue: it never returns an ambiguous answer. Somebody clicked, somebody subscribed, the channel that served the click gets the credit. It is tidy, it reconciles, and it is wrong in a specific and expensive direction — it pays the channel that closed and tells you nothing about the one that made closing possible.

We had a suspicion about which channel that was. Suspicions are not a budget case, so we went looking for a number.

Designing the test around the doubt

The instrument was a Meta Conversion Lift study with channel lift methodology. The design is the important part: a randomised group is held out from Meta ads entirely, and the difference in what that group does — including on other paid channels — is the causal estimate. Nobody is asked to model an attribution window. The counterfactual is measured.

We ran it over three weeks, from 13 April to 4 May 2025, and fed subscription events in server-side through the Conversions API and our mobile measurement partner, on last-click, so the study was reading the same conservative data our reporting already used. If the effect existed, we wanted it to show up against our own least generous view of Meta.

A third of the credit was in the wrong column

of other channels’ subscriptions were incrementally Meta’s

33%

lift in subscriptions from Meta

45%

range of Meta’s halo across channels

12–62%

On average, a third of the subscriptions our other paid channels were reporting would not have happened without Meta advertising. The spread underneath that average matters more than the average: depending on the channel, Meta’s contribution ranged from twelve percent to sixty-two percent. A single blended correction factor would have been wrong almost everywhere.

What we changed

The immediate consequence was that Meta had been underfunded, and by a margin large enough that the previous allocation was not defensible once the number existed. The more durable consequence was methodological. A platform’s own reported return is a lower bound on its contribution, not an estimate of it, and the gap between those two things is not a constant you can carry in a spreadsheet.

So incrementality testing stopped being an occasional exercise and became part of how we plan. Holdouts are uncomfortable — they cost real revenue while they run, and the whole point is that you must not peek. They are also the only way to find out what your advertising actually did, as opposed to what your attribution was willing to say about it.