What are ghost ads?

Measuring an ad's effect needs a control group that would have seen it. Simply comparing exposed against unexposed users fails badly, because the ad system chose who to show it to — it targets people likely to convert, so exposed users differ from unexposed ones in exactly the way that matters. The comparison measures targeting rather than advertising.

The older fix was a PSA control: run the auction as normal, and for control users serve a charity advertisement instead. Exposure is then matched, since the same auction logic selected both groups. It also introduces two problems. You pay for the PSA impressions, which is money spent to learn nothing, and the PSA is itself a stimulus — an unfamiliar advertisement occupying the slot the real one would have taken, which is not the same as showing nothing.

Ghost ads solve both. The auction still runs and the system records that the control user would have won the impression, and then serves nothing — the slot goes to whatever would have been next. No spend, no substitute stimulus, and the same exposure-matched comparison. The measured group is the set of users who would have been shown the ad, which is exactly the population an incrementality question is about.

The practical payoff is cost. Because control impressions are never bought, a ghost-ads study can hold out a much larger share of the audience for the same budget, and a bigger holdout means a more precise estimate. A PSA design forces a trade between the size of the control group and the money wasted on it; ghost ads remove the trade entirely, which is why they became the standard once platforms supported them.

The limitation they share with any conversion lift study is the boundary. A ghost-ads test measures one platform's incrementality relative to everything else you are running, and cannot see demand shifting between channels. It also requires platform support, since the logging happens inside the ad server — this is not something an advertiser can implement independently, which is worth knowing before planning around it.

The same campaign is measured three ways on a £180,000 budget: a naive exposed-versus-unexposed comparison, a PSA-control design, and a ghost-ads design.

The naive figure is the one to dismiss outright: it is almost entirely the targeting system doing its job, since the people the algorithm chose to show ads to were far more likely to convert before any ad appeared. The two valid designs agree that the real effect is in single digits. The differences between them are instructive. Ghost ads afforded a 30% holdout against the PSA design's 10%, because control impressions cost nothing, and that tripled holdout is why the interval is less than half as wide. The point estimates also differ by more than noise alone would suggest, which is consistent with the PSA acting as a mild stimulus — the charity ad occupying the slot is not the same as an empty slot, and it plausibly depresses the control group's conversion slightly, inflating the measured lift. Ghost ads remove that ambiguity entirely.