A conversion is defined by an advertiser as a specific action taken by the search user at the advertiser's landing page, e.g. filling out a form or buying an item. Advertisers use Sponsored Search to drive traffic to their site at a conversion rate and cost per action (CPA) per conversion that provides value to them. Improvement for advertiser value is an increase in conversions and/or a decrease in cost per action. Higher conversion rate is also a proxy for increase in search user satisfaction. This project involves measuring and predicting conversion rates and using it to influence ranking, pricing and placement of ads to improve the value of Sponsored Search to search users and advertisers.
There are two broad areas of the project: measuring and predicting conversions and using this data in different applications.
Measurement and prediction of conversion rates involves a number of modeling challenges due to the opt-in nature of conversion-tracking by advertisers, ranging from the very definition of what is a conversion to self-selection bias and incomplete/scarce data for certain scenarios. We use machine learning techniques to predict conversion rates in a robust manner using a number of features - for example, a score for the match between the query and the ad, and the type of match of the ad to the query (is it an "exact match" - that is, the ad was bidded on the search query, or an "advanced match" in which the ad was bidded on a related query). Other useful features could capture user intent (how much intent is commercial), quality of match between query and advertiser landing page, or market type.
Once we can measure and predict conversion rates accurately, we can apply that data in a number of areas. It can be used for discounting clicks on channels with lower traffic quality. It is used to monitor the health of Yahoo's Sponsored Search marketplace - for example, trend of conversion rates and CPA with time for the entire market or in particular market segments of interest, or in evaluating tests of alternate configurations of the SS system. Another application is to use the predictions to discount the bids of ads that are advanced matched for use in the ranking, pricing and placement modules in order to increase the conversion rate to the advertiser, decrease their CPA and improve the search-user experience. Going a step further, this would allow us to provide advertisers with a service in which, given their bid per click and bid per conversion (either estimated via conversion tracker or given by advertiser), we optimize their CPA and rank their ads accordingly.