Ad platforms take 100% credit for any purchase made within seven days of a click, whether the ad caused it or not. The only way to know if an ad actually moved the needle is a controlled experiment, where a random group sees nothing and you measure the gap. A new paper from Meta and Northwestern trained a model on 2,226 real experiments and showed that last-click attribution gets the call wrong about one in five times. Their method cuts that to one in eight.