Incrementality Testing (Ghost Ads, PSA & Geo Lift)

By the end of this article, you should be able to: explain why exposed-vs-unexposed comparisons over-credit ads, how ghost-ad and PSA holdouts fix it, when to fall back to geo lift, and how to read an incremental ROAS honestly.

Step 1: Meet the problem

An advertiser runs a campaign and sees that people who saw the ad convert at 5%, while people who didn't see it convert at 2%. The dashboard says the ad "drove" a 3-point lift. Should they pour in more budget?

Not so fast. The people who saw the ad aren't a random slice — the ad platform targeted them precisely because they looked high-intent (they searched for the product, visited the site, match a lookalike). So they would have converted at a higher rate even with no ad. The 5%-vs-2% gap mixes the ad's real effect with the fact that the exposed group was already more likely to buy. That's selection bias, and it makes almost every naive ad measurement over-credit the ad.

The question incrementality answers is causal: how many conversions did the ad actually cause that would not have happened otherwise? To get it, you need the missing counterfactual — what the exposed people would have done without the ad.

Step 2: The counterfactual you can't see

For a person who saw the ad and converted, you can't observe the parallel world where they didn't see it. So you approximate it with a control group that is as similar as possible to the treated group — same targeting, same intent — but who didn't get the ad.

The trap is that a naive control ("people who didn't see the ad") is not similar: they didn't see it often because the system judged them lower-intent, or they weren't reachable. Comparing to them bakes the selection right back in. You need a control that was eligible and targeted identically, and only randomly withheld from the real ad.