A New Serving of Sponsored Search
Is there a better way to sell search ads? David Pennock, a Yahoo! Research scientist in the microeconomics group, believes there is always room for improvement.
That’s why he’s investigating new sponsored search scenarios that could lead to more revenues for Yahoo! and a better experience for users and advertisers.
Pennock was recently joined by a handful of other Yahoo! researchers and engineers—including Pavel Berkhin, Chad Carson, Ashvin Kannan, Darshan Kantak, Sebastien Lahaie, Chris LuVogt, Jan Pedersen, and Tong Zhang—in a quest to make sponsored search even more positive—and profitable—for everyone.
The history of sponsored search has been one of incremental advancements. At first, Yahoo! and other search engines simply sold keywords to the highest bidder. The advertiser who paid the most for a particular keyword, such as “snowboard”, would see their ad appear at the top of the page.
The model was then modified so that relevance, or click-thru rates, would be factored into the equation. Today, rankings are based on the amount an advertiser pays for a keyword as well as the number of times users actually click on their ad. So an advertiser that has twice the click-thru of a competitor might actually be ranked higher, even though their bid for a particular keyword is less.
This is where the work done by Pennock and his team comes into the picture. Instead of basing advertising rank on the accepted model of highest bid multiplied by relevance, they decided to fiddle with click-thru rates and see what happens if relevance is given slightly less consideration.
Their model looks something like this: highest bid multiplied by relevance to the power of q, where q ranges between 0 and 1. So if q is zero, advertising rank is simply determined by the highest bid. But if q is say, .5, then the model still takes relevance into account, but squashes it down, giving it less weight than it normally would.
"We wanted to investigate a whole new family of keyword auctions by building a hybrid model based on what had come before," Pennock says. "The idea is to explore other mechanisms to help us maximize our own revenue as well as maximize the value we provide to advertisers and users."
As an added twist, the so-called squashing algorithm takes into account the concept of "quot;envy-free equilibrium," which was first put forward by another Yahoo! Research scientist, Michael Schwarz. This equilibrium assumes that advertisers will continually adjust their bids to compensate for any changes made to the auction process.
So what’s the most interesting discovery made by Pennock’s team? "It sounds counter intuitive, but our biggest takeaway is that ranking by revenue is not necessarily revenue optimal," says Pennock. That’s because of the subtleties of how advertisers change their bids when the rules of the auction change.
He adds that squashing can significantly improve revenue, at the expense of advertiser and user satisfaction. As a result, it is necessary to set acceptable thresholds for loss of relevance, and then optimize revenue based on those thresholds.
Pennock and Lahaie write about the research aspects of the project in a paper published at the 2007 ACM Conference on Electronic Commerce.
