This paper, written by Microsoft in 2007, is a little old, but it offers some interesting insight into an aspect of advertisement selling for search engines that we didn’t cover in class: variance in CTR among advertisements.

Search engines have an incentive to place well-clicking advertisements in prominent positions because doing so leads to more clicks which (depending on bid prices) can lead to more revenue. In order to sort by revenue expected, the search engines need to be able to predict the CTR value for an arbitrary advertisement. The naive approach is to simply measure CTR after the ad is displayed, and from there calculate its quality - CTR is both a function of the ads quality and its display location - but, because of the low CTR for most ads, it takes many clicks to get a reliable measure of quality.

The period of time in which the CTR is unknown represents a less-than-optimal allocation of advertisement slots, leading to lost revenue for the search engine.

The authors of the paper constructed a logistic regression model to study a variety of characteristics:

• The “landing page,” or the URL that the ad links to
• the bid terms the ad appears with
• the title of the ad
• the body text of the ad
• the URL displayed at the bottom of the ad
• The clicks received since being entered (The naive CTR estimation)
• view since being entered

They also measured a number of desiribility characteristics and trained the model to calibrate for the most common 10,000 words in the data set (taken from microsoft search advertisment data)

Taken all together, these reduced the error by 23%, with the common word accounting for about 3/4 of the improvement.

Further adjusting for the degree of specificity reduced error to 30% from the naive estimate.

This model has the potential to increase revenue dramatically for search provideers by allowing them to place ads more intelligently and by allowing them to advise advertisers on how to improve their advertisements, and is an interesting parallel to how the pricing should work independently of the ad quality.

The sophisticated model provides an advantage over the naive one until about 100 views, or about 200-300 searches.