Advertiser Learning in Direct Advertising Markets

Abstract

Direct buy advertisers procure advertising inventory at fixed rates from publishers and ad networks. Such advertisers face the complex task of choosing ads amongst myriad new publisher sites. We offer evidence that advertisers do not excel at making these choices. Instead, they try many sites before settling on a favored set, consistent with advertiser learning. We subsequently model advertiser demand for publisher inventory wherein advertisers learn about advertising efficacy across publishers' sites. Results suggest that advertisers spend considerable resources advertising on sites they eventually abandon—in part because their prior beliefs about advertising efficacy on those sites are too optimistic. The median advertiser's expected CTR at a new site is 0.23%, five times higher than the true median CTR of 0.045%.

We consider how pooling advertiser information remediates this problem. Specifically, we show that ads with similar visual elements garner similar CTRS, enabling advertisers to better predict ad performance at new sites. Counterfactual analyses indicate that gains from pooling advertiser information are substantial: over six months, we estimate a median advertiser welfare gain of $2,756 (a 15.5% increase) and a median publisher revenue gain of $9,618 (a 63.9% increase).

Keywords: Display advertising, Learning models, Bayesian estimation

JEL Classification: C11, C51, D61, D83, L14, L82, L86, M31, M37