How do you measure an ad's effect on someone who never clicked it?

How do you measure an ad's effect on someone who never clicked it?

Most attribution credits a conversion to whoever got the last click. But the people an ad truly moves often never click at all — they see it, then convert later through another path.

The method. Facebook researchers (Gordon, Zettelmeyer, Bhargava, Chapsky, 2019, Marketing Science) compared randomized controlled trials against observational attribution across 15 advertising studies covering roughly 500 million user-experiment observations. Their tool was "ghost ads": the control group is logged as eligible to see an ad and recorded at the moment they would have, but shown nothing — giving a clean counterfactual for lift.

What they found. Observational methods (last-click, even multi-touch and propensity matching) mis-estimated the true randomized lift badly. In several studies, observational approaches overstated effect by 100% or more; in some, they got the sign wrong entirely. Matching on observables closed only part of the gap because the people who see ads differ systematically from those who don't — selection, not persuasion.

The nuance. This is not a click-versus-view debate. It's that any attribution built on observed behavior inherits the bias of who was targeted. No reweighting fully removes it.

Bottom line for practitioners: if a decision is big enough to matter, an experiment beats a model. Multi-touch attribution (spreading credit across touchpoints) is useful for diagnosis and pacing, but treat its ROAS numbers as hypotheses to be tested by holdouts, not as causal truth. The 100%-plus overstatement was the rule across studies, not the exception.