IBM’s Industry Solutions group building in Hawthorne, NY, where Mohan works, houses a big customer demonstration area showcasing new concepts and technologies, many of them aimed at e-commerce. A telematics demo shows how directions, traffic alerts, and promotions can be sent over the Internet to cars, based on their Global Positioning System location. And all around are systems for payment and authentication—from a cryptographic chip for wireless transactions to conversational biometrics, which perform voiceprint analysis while asking questions only authorized users should be able to answer.
The e-commerce technologies pursued here, and at other labs worldwide, cover an ever expanding range of areas. However, underlying virtually all the personalization and customization efforts are Web analytics and data mining. “We are paying close attention to the tendency of shoppers to visit our online store and then, within a week or so, come in and buy at our retail stores,” says REI’s Broughton. “We want to get better at providing our online customers whatever they need—product information, store locations, articles about the activities they’d like to enjoy—so that they shop our retail stores as well.”
Even when customers don’t hand retailers detailed information about themselves, a lot can be gleaned from what they do online. For nstance, geolocation and data-mining company Digital Envoy of Norcross, GA, tracks two billion Internet addresses a day, culling demographic data that advertisers and retailers love. Working like a search engine that studies the Web’s infrastructure rather than its content, the firm’s system can track an Internet transaction backward from a website to the network node at which it originated in order to answer two questions: what city is the user in, and how fast is her Internet connection? From that information, the system can make a good guess about what business the user is in. Digital Envoy can also identify a person’s local area code, time zone, and zip code—and determine what language she speaks.
The hunger for such information is growing fast, in large part because of e-commerce, notes Digital Envoy cofounder Sanjay Parekh. For one thing, the information improves fraud detection. If a buyer claims to be in Florida, but his Internet address shows he’s in Wisconsin, that’s a tip-off that something is amiss. Even more to the point, Parekh notes, if a retailer knows that a customer is in New York, not Palm Springs, it might display a different style of clothes on its home page—or dispatch a coupon good at a store in Manhattan. But the biggest force driving data mining is the push to provide better context for the paid keyword ads linked to search terms. Online marketing and advertising service provider DoubleClick, for instance, uses Digital Envoy’s technology to help companies create ads that are location specific. The goal, says Parekh: “Drive to a sale much quicker. That’s what everything is about.”
IBM researchers tackle many of the same issues, though differently. One of the company’s projects involves recording a customer’s mouse clicks and tracking what was viewed—a red blouse, say. From that data, it’s possible to get a good idea of a visitor’s feelings about price, color, and size preferences—even his or her gender. If the shopper makes a second visit, the retailer might offer a discount on an item already examined, or something similar. Every purchase gives the store more information about its customer.
Such efforts, of course, raise privacy concerns. In 2001, Big Blue founded the IBM Privacy Management Council, a coalition of privacy and security leaders from health care, finance, retail, and government that seeks to find ways—through technology, standards setting, and business practices—to get ahead of looming privacy issues. A big push is related to database management—so that when you enter personal information like your name, income, and tastes in lingerie into a company database, it is stored in such a way that none of the company’s employees can put all the pieces together and trace them to you. Instead, customers are profiled in the aggregate and grouped into broader categories that allow retailers to tailor (and even personalize) offers to people as members of a class but not as specific individuals. Privacy concerns are at the top of the list for retailers, says Hopping, of IBM’s Retail Store Solutions Division, “because if privacy blows up, that’s the kiss of death for a retailer.”
Retailers must also make the gathering of information about shoppers worthwhile to the shoppers themselves. “To get the customer to opt in, the retailer’s going to have to give them something,” says Hopping. That can be a discount, he says, but often it is something else—special parking or other services, or a piece of technology like Shopping Buddy that makes shopping easier.
One of the most intriguing areas of research involves figuring out which customers are worth the trouble of wooing in the first place. In its Haifa, Israel, research lab, IBM is designing advanced statistical and machine-learning models that will differentiate customers according to their future value.
Researcher Amit Fisher developed one such model by studying a year’s worth of activity at one of Israel’s leading e-auction sites. From such factors as the number, frequency, and value of their transactions, Fisher was able to classify Internet users into different categories—along the lines of bargain hunter, repeater, one-timer, defector, valuable customer—and assign an economic value to each category. The model then sought to predict from just a few visits where a new user was likely to end up a year or more down the road. “We compared the customer ranking that was generated by our model to the true ranking of the customer according to their purchases,” says Fisher. In trials involving groups of more than 1,000 users, he notes, the model correlated almost perfectly with actual data collected from the auction site. Programs that seek to assess a customer’s lifetime value are not new; however, IBM says that Fisher’s model, which is being developed for commercial use through several of IBM’s businesses, is the first to make an accurate assessment of a customer’s future value based on just a few visits. What’s more, such a model is “domain adjustable,” Fisher says. It could be used in banking to determine whether to issue a loan or a credit card. Or it could be employed by retailers to target promotions to potential best customers and give priority to those customers during times of peak demand.
Fisher’s model works with minimal personal data and takes into account only a few variables. But for retailers bent on amassing much more complete data about their customers—and then using that information to “maximize” lifetime customer value through highly targeted ads and promotions—the data-mining challenge is far trickier, says Edwin Pednault, a staff researcher in IBM’s data analytics research group. “Now I want to take much more information into account,” says Pednault—who has been working on a model that would do just that.
Instead of looking at effects of marketing campaigns separately, as traditional data mining has done, Pednault’s model examines the patterns of a customer’s activity, such as the types of products she likes, how she responds to promotions, and her price sensitivity. When a company has that kind of information about its customers, says Pednault, it can begin to ask, “How are my actions motivating them to change from one [buying] state to another?” In studies of one major department store chain, IBM showed that using Pednault’s model to predict the effects of snail mail marketing—alerting customers about sales, store events, and new items from their favorite product lines or styles—resulted in a 7 to 8 percent increase in store revenues.