Search a big store's app and some results are paid ads. The store records that you saw the sponsored listing and, on its own till data, that you then bought. It calls that a win and bills the brand. Here are 100 shoppers who saw one sponsored listing. 40 of them bought the product. The store grades the campaign on those 40.
Illustrative figures, invented for this demo40 of 100 shoppers who saw the ad then bought.
Reported return on ad spend: 4.0x. Verdict: strong campaign, spend more.
Before you believe it, make one guess. Of those 40 recorded sales, how many did the ad actually cause?
Your guess: 20 of 40 sales caused by the ad.
A matched group of 100 shoppers was never shown the ad. 34 of them bought anyway.
So only 40 minus 34 = 6 sales were caused by the ad. The other 34 credited sales were going to happen without it.
Real, incremental return on ad spend: 0.6x. The campaign lost money; the store's own grade hid it.
Figures are illustrative. The mechanism is not: comparing attribution against randomised experiments at large scale, Gordon, Moakler and Zettelmeyer (2023) found attribution systematically overstates the effect an experiment measures. The size of the gap for any named retail network has not been published, so the numbers here stand for the shape of the problem, not a measured amount.