AI

Near Impressions for Observational Causal Ad Impact

Abstract

Advertisers often estimate the performance of their online advertising by either running randomized experiments, or applying models to observational data. While randomized experiments are the gold standard of measurement, their cost and complexity often lead advertisers to rely instead on observational methods, such as attribution models. A previous paper demonstrated the limitations of attribution models, as well as information issues that limit their performance. This paper introduces "near impressions", an additional source of observational data that can be used to estimate causal ad impact without experiments. We use both simulated and real experiments to demonstrate that near impressions greatly improve our ability to accurately measure the true value generated by ads.