On Peculiarities of Positional Effects in Sponsored Search

Click logs provide a unique and highly valuable source of human judgments on ads' relevance. However, clicks are heavily biased by lots of factors. Two main factors that are widely acknowledged to be the most influential ones are neighboring ads and presentation order. The latter is referred to as positional effect. A popular practice to recover the ads quality cleaned from positional bias is to adopt click models based on examination or cascade hypothesis originally developed for organic search. In this paper we show the strong evidence that this practice is far from perfection when considering the top ads block on a search engine result page (SERP). We show that cascade hypothesis is the most questionable one because of important differences between organic and sponsored search results that may encourage users to analyze the whole ads-block before clicking. Additionally, we design a testing setup for an unbiased evaluation of click model prediction accuracy.