The first model doesn’t make sense to me. You treated everyone in the control group who made a purchase as a 0? In that case, of course, the model is very biased to predict only people receiving promotions would respond which would greatly inflate the metrics. But maybe I’m misunderstanding what you actually did.