I dispute what you are saying, as someone who has studied that Lewis & Rao paper in detail. Advertising academia is pretty much just as mixed up as advertising

I dispute what you are saying, as someone who has studied that Lewis & Rao paper in detail. Advertising academia is pretty much just as mixed up as advertising in practice. Here is another paper worth reading on it [0] (which cites Lewis & Rao), but provides fairly clear evidence of a distribution of outcomes of digital ad campaigns with a lot of probability mass on the side of positive lift. The place where issues come in via Lewis & Rao is on the side of poor statistical methodology, particularly in the calculation of necessary statistical power. But those discussions are all stuck in bad frequentist estimation techniques, and in industry there are practitioners using Bayesian methods, for example such as Gelman’s proposed “type sign” and “type magnitude” errors, with well-calibrated priors that treat the possibility of negative lift seriously. I strongly disagree with the idea that “ads don’t work” is an “open secret” like you say. Rather, “measuring ads is extremely complicated” is the known conclusion, and marketers are just coming around to the fact that you have to invest in extremely advanced statistical methods to even know if ads are working, let alone to optimize them for an audience, especially inclusive of things like audience privacy, opt-out policies, paychological well being and so on. [0]: https://marketing.wharton.upenn.edu/wp-content/uploads/2017/08/Johnson-Garrett-PAPER-VERSION-2.pdf https://marketing.wharton.upenn.edu/wp-content/uploads/2017/...