Kestrel & Vane runs a four-week holdout: 90% of its retargeting audience can see the ads and 10% cannot. Scaled to the same size, the exposed group buys 1,000 times and the holdout group would have bought 700. Meta credits the ads with all 1,000 at a 4.0 ROAS, but on the 300 incremental sales the return is closer to 1.2.

Attribution counts sales that touched an ad. Incrementality asks whether the ad caused them. Retargeting and brand-name audiences often look great in attribution and add little lift, while broad prospecting can look weaker and drive more new sales. Budgets set on attributed ROAS alone can fund the wrong campaigns.

Run holdout or geo tests on the campaigns that take the most budget, starting with retargeting. Use Meta's lift studies where available, or split regions and compare. Then use the gap between incremental and attributed results to discount platform numbers when you plan.