Speaker: Kourosh Gharachorloo and the Traffic Quality Team, Google Inc.
Title: Anecdotes from the Front Line
Abstract:
Since early 2002 when Google introduced cost-per-click (CPC) pricing
in its Adwords program, engineers at Google have been working on
improving our techniques for detecting and eliminating the impact of
click-fraud on our advertisers. The goal of this talk is to provide
some observations based on our experience, and to pose some
interesting questions and challenges for the budding research
community in this area.
By far the biggest challenge in click-fraud research is a lack of clear evaluation metric. This partly arises from the inherent nature of the problem, because determining whether an individual click is fraudulent is not feasbile in general. The lack of an evaluation metric makes it difficult to answer critical questions such as:
Related to metrics, we will also briefly touch upon the efficacy of overall click-fraud rates that have been publicized by several firms and often requested by journalists. Limitations of such rate estimates are numerous.
We will discuss basic click counting issues that can lead to discrepancies between events measured by an ad network and an advertiser. Some discrepancy is expected and inherent in the way traffic is redirected from a user's machine to an advertiser through an ad network (e.g., consider a redirect being dropped). We will briefly outline the various common reasons, including poor accounting of events, that can lead to major differences between what is measured by the two sides, and discuss extra information that can be provided by ad networks that can substantially simplify the task of reconciling events on the two sides.
Understanding and quantifying the impact of undetected fraud on advertisers requires an understanding of the auction dynamics, since we are not dealing with a static system and advertisers are continuously adjusting their campaigns/keywords/bids based on seasonality, competition, etc. One of our big concerns is advertisers that either do not monitor ROI or do not use the appropriate metrics for measuring ROI (we will discuss some common pitfalls). Such advertisers are likely not running optimized campaigns and can often spend their advertising dollars more efficiently. Furthermore, they are more prone to react to potentially incorrect information related to click-fraud; sometimes this leads to unnecessarily lowering spending on part of their campaign and hurting their overall sales. Such advertisers are also more prone to use click-fraud as the scapegoat reason for explaining decreases in ROI, often delaying them in identifying and reacting to real causes (e.g., better product prices at a competitor).
Finally, we will briefly mention some other issues that are tangentially related to click-fraud, including: impression spam, dealing with repeat fraudsters through detection of duplicate accounts and better online "id verification", cost-per-action and how it compares to cost-per-click.