Yahoo's Algorithm Aims To Improve Behavioral Targeting
A Yahoo patent filed with the U.S. Patent and Trademark Office explains how the company uses a "site sequence" algorithm it developed to deliver the most relevant ads to a variety of Web site visitors. The abstract begins by defining site sequence as the values representing a sequence of events that occur from a first site (A) to the second (B) for at least one person, along with an identified corresponding engagement level.
The patent, updated in January but originally filed in July 2008, studies visits from the first Web site to the second from a variety of people to determine an average engagement level. Some of the site visitors are frequent Internet users, while others are not. Taking an average allows Yahoo to use the technique it calls "site sequence."
The patent also describes ways to study visits from the second site (B) to the third site (C) to determine the engagement level of the person clicking from one Web site to the next. Monitoring the time it takes to get from one site to the next tells Yahoo whether or not a visitor to the first site was interested in the last.
The most interesting part of the patent application describes what Yahoo calls a widespread need for large ad networks to sell targeted advertisement to users. Key to realizing this vision is the development of technology enabling a system like Yahoo to capture and utilize a user's behavioral profile -- for example, associations between user preferences and user activities, such as the likelihood for visiting site B based on a visit to site A. Users who opt in and choose to reveal information about their behavior habits tend to generally be the more engaged group of users, as measured by time logged into the Internet.
It's not the first patent Yahoo has filed based on behavioral targeting. SEO by the Sea Bill Slawski has outlined many, including one that describes Yahoo search advertising and behavioral targeting. The post from November 2007 describes a behavioral targeting system that determines user profiles from online activity. The system includes several models that define parameters for determining a user profile score.