Predicting Purchase Behavior of Website Audiences

This paper proposes a methodological framework that extends the advantages of behavioral targeting while preserving the privacy of the individual. Instead of profiling individual users according to their general interests, we profile website audiences according to their online purchase behavior. This presents a trade-off between looser, aggregate audience profiling and deeper understanding of actual purchase behavior, the holy grail of online advertising. Our framework is based on the analysis of raw clickstream data of Web users who explicitly agreed to participate in an online audience panel. We experiment with data collected by an online analytics company, SimilarWeb, which consists of 3,463,796 records of online purchases and 1.1 billion records of Website visits. We train a multilabeled classification model on the clickstream of panel members with distinctive online purchase profiles to predict the purchase potential of the entire panel. We aggregate the individual purchase behavior profiles (both ground-truth and predicted) into purchase behavior profiles of Web domain audiences and test the resulting methodology on 3,408 Web domains, with very promising results. If privacy-related regulation tightens up in the near future, the proposed panel-based, purchase-focused ad targeting mechanism might be the panacea for online advertisers.