Bing users might not notice, but the Microsoft search engine tends to show fewer advertisements alongside search results than its rivals do. The reason might be that advertisers are drawn more to Google, but Microsoft says the difference also comes from an artificial-intelligence technique that aims to deliver only the most relevant ads.
The technique will soon get a wider test–Yahoo is now making the switch to Bing to power the search results on its site. And that is prompting Microsoft researchers to share more details about the algorithms Bing uses to try to please users and advertisers alike.
Although often annoying to users, the ads that appear next to search results play a crucial role in the online economy. They generate more than $25 billion in revenue each year.
But most of that money flows to Google, which handles about 63 percent of U.S. search requests. Yahoo processes about 19 percent, while Bing has a nearly 13 percent share, according to the latest data from comScore, a marketing research company. Microsoft intends to reach a wider audience of Web searchers through its partnership with Yahoo.
Microsoft also hopes that by using search-ad algorithms to make each ad more relevant, it can get more clicks–and more money from advertisers–out of each ad. This is key, because simply placing more ads on a page doesn’t necessarily increase revenue.
To try to predict how many clicks an ad will get, algorithms analyze historic trends to see what users have done when ads were shown alongside given search terms. In the past, this was handled by writing relatively simple rules into search software. For instance, if search users entered a query with the word “car,” they might see ads related to oil changes.
Bing incorporates “machine learning” into the equation. It uses a system called AdPredictor, which was developed by Joachin Quiñonero Candela, a Microsoft researcher in Cambridge, England. With the help of an artificial-intelligence tool called Bayesian modeling, AdPredictor takes continually updated data into consideration when calculating the probability of an event. In the case of search ads, AdPredictor analyzes how often an ad has been clicked on in a vast number of conditions, says Quiñonero Candela. For instance, ads related to cars might change depending on the time of day, the location of the person doing the search, and the placement of the ad on the page. (Quiñonero Candela says that for competitive reasons, he can’t reveal precisely what these conditions or parameters are.)
The result, says Quiñonero Candela, is that Bing can afford to show fewer ads, because it knows those ads are more likely to be clicked on. It doesn’t necessarily have to show the ads whose sponsors have paid the most for placement, he says. “When you try to improve the accuracy of your predictions, you want to try to annoy the user less and keep the advertisers happy at the same time,” he says.
Indeed, when Bing launched with the AdPredictor technology, its click-through rate was nearly three times higher than what Microsoft had seen with its predecessor search engine, Live Search, according to the marketing firm Didit. And according to a recent analysis by a firm called Adgooroo, Bing generates 3.85 ads per search keyword, compared to 5.72 at Google. Again, it’s not clear if this is purely the result of AdPredictor’s performance, or whether advertisers just prefer Google. (It’s also presumed that Google employs a relevance algorithm to decide which search ads to show, but the company wouldn’t comment.) With Yahoo now firing up Bing’s search engine, it will be telling if Yahoo’s relatively high rate of 6.85 ads per keyword comes down.
One important element of the Bing technology: It learns on the fly, online, as new data arrives. In contrast, “some algorithms need all the data before you can update the model,” says Qi Guo, a machine-learning expert at Emory University who has studied how people react to search ads. Although algorithms that can learn online are not new, Bing’s approach is a new application of the technology, he says.