Improving the performance of the overall system meant setting performance goals for each of those subsystems and incrementally ratcheting them up. The Web site team used tools such as JProbe, the Java performance analysis utility from Quest Software, to identify and eliminate performance bottlenecks in their software, and Mercury Interactive's LoadRunner to test for performance and scalability.

"First and foremost, you've got to have targets and goals, and your design has to be focused on achieving them," Vandevier points out. The operations research data helped him set more rigorous goals, he says, "and it's played a part in a lot of our product development ever since. We now include a response time element in every factor of our operations."

The performance targets were part of a broader effort to create a mathematical model of how customers choose between travel options, based on the utility of any given travel option, where utility is a factor of price, departure time, trip time and airline choice. Given an understanding of how important those things are to different types of customers—gained from Travelocity's study of actual purchase behavior—Travelocity could tweak the way results were displayed to ensure that the options most likely to entice a purchase would stand out. After each change, the researchers compared actual purchase data with the improvement predicted by their models and used that data to further refine the models.

A major finding was that, as of 2004, Travelocity was less likely than its major competitors to find the lowest airfare between two cities. Eliminating that gap became a major goal of the new version of the Web site Travelocity introduced at the beginning of 2005, Vandevier says, and required the Web site to analyze many more travel options and combinations of those options without sacrificing performance.

As a search problem, finding the lowest fare is more complicated than it might seem because of the multitude of possible flight combinations, including those that involve piecing together flight segments from multiple airlines for a single round trip. The researchers were able to attribute 60% of the cases where Travelocity failed to find the lowest fare to "unexplored fare combinations" where the Web site software had failed to cover all of the possibilities, says Faker Zouaoui, a researcher with Sabre's marketing department.

That shortcoming was costing Travelocity about a 10% gap in market share between itself and competitors among cost-conscious buyers, according to Zouaoui's analysis. Again, Travelocity used this knowledge to realign development priorities, this time toward a more comprehensive search of flight combinations. According to Vandevier, this was mostly a matter of refining to cover a broader spectrum of travel options and their combinations. Within three days of the implementation of a more thorough low-fare search, Travelocity was able to measure a 36% reduction in that market share gap, and over time it cut that gap by 55%, Zouaoui says.

Lance Jones, an analyst who studies Web site consumer satisfaction for Keynote Systems, says he can see the impact of Travelocity's efforts in his surveys of the air travel market. "Price satisfaction did improve substantially," he says.

Between 2004 and 2005, the number of study participants who reported they didn't feel they found a good price on Travelocity's site dropped from 18% to 6%, Jones says.

NEXT PAGE: Mining For Deals

>>> More Projects: Processes Articles          >>> More By David F. Carr