We study the effects of internet display advertising using cookie-level data from a field experiment at a financial tools provider. The experiment randomized assignment of cookies to treatment (firm ads) and control conditions (charity ads) enabling us to handle different sources of selection bias, including targeting algorithms and browsing behavior. We analyze display ad effects for users at different stages of the company's purchase funnel (e.g., non-visitor, visitor, authenticated user, converted customer). We find that display advertising positively affects visitation to the firm's website for users in most stages of the purchase funnel, but not for those who previously visited the site without creating an account. Using a binary logit model, we calculate marginal effects and elasticities by funnel stage and analyze the potential value of reallocating display ad impressions across users at different stages. Expected visits increase almost 10 percent when display ad impressions are partially reallocated from non-visitors and visitors to authenticated users. We also show that results based on the controlled experiment data differ significantly from those computed using standard correlational approaches.

This paper introduces a near optimal bidding algorithm for use in real-time display advertising auctions. These auctions constitute a dominant distribution channel for internet display advertising and a potential funding model for addressable media. The proposed e_cient, implementable learning algorithm is proven to rapidly converge to the optimal strategy while achieving zero-regret and constituting a competitive equilib- rium. This is the _rst algorithmic solution to the online knapsack problem to o_er such theoretical guarantees without assuming a-priori knowledge of object values or costs. Further, it meets advertiser requirements by accommodating any valuation metric while satisfying budget constraints. Across a series of 100 simulated and ten real-world campaigns, the algorithm delivers 98% of the value achievable with perfect foresight and outperforms the best available alternative by 11%. Finally, we show how the algorithm can be augmented to simultaneously estimate impression values and learn the bidding policy. Across a series of simulations, we show that the total regret delivered under this dual objective is less than that from any competing algorithm required only to learn the bidding policy.

No abstract

No abstract

No abstract