Latest · Ad platforms

Meta's Own Data Says Its Lift Numbers Are Fiction

September 27, 2026

Meta ran 663 experiments comparing its machine-learning lift estimates against actual randomized controlled tests. The machine-learning guesses put upper-funnel lift at a median 83%. The randomized tests, the ones that actually hold out a control group instead of modeling one, found 29%. That's not rounding error. That's the platform grading its own homework and inflating the score by 3x.

Here's why that should bother you specifically. Every time your Meta rep pulls up a lift study to justify more leasing budget, you're looking at the modeled number, not the randomized one, because the modeled number is the one Meta shows by default. You've been comparing Facebook's optimistic self-report against your ILS invoice and your actual leasing traffic, and wondering why the platform always looks like the hero and the spreadsheet never agrees.

The move isn't to pull spend off Meta. It's to stop taking its lift reporting as proof of anything. If you're shifting real budget based on a "lift study," ask what kind: modeled, or a genuine holdout test with a control group that never saw the ad. If your rep can't answer that clearly, assume modeled, and assume it's inflated.

Small operators can't run their own incrementality tests at scale. Fine. But you can still hold every platform's self-reported lift to the same standard: is this a number Meta produced about Meta, or a number someone else checked? One of those is marketing. The other is math.

Source: PPC Land, September 26, 2026