The marketplace gives consumers access to more than 300 products created using a 3-D printer.
(Page 2 of 3)
Staff at Scentiments were being forced to think up cross-sells. Prior to implementing product recommendations, the e-retailer regularly received questions from customers by phone and e-mail asking for fragrance suggestions based on other fragrances they liked. Now the recommendations system does the work, and it does so in a more precise and effective manner than humans can, says Wyner of Scentiments, which uses software-as-a-service recommendations from MyBuys.
“It’s about the customer experience-we wanted to set ourselves apart from other web sites and give shoppers input on other choices when choosing their own brand they like,” Wyner says. “Our customer is usually driven by their brand. They’re looking to refill their prescription, so to speak. But many see an ad in a magazine and hit the site to search, for example. The MyBuys engine mines the customer and product data and presents other products recommended just for you.”
A big question among retailers looking to buy product recommendations systems is which style to choose: one that bases recommendations primarily on the online behavior of an individual or one that bases them on the collective behavior of a retailer’s online shoppers?
The Knot selected Baynote in part because the vendor uses the group approach to recommendations. “Choosing to do cross-sells is a no-brainer, it’s a huge opportunity,” Savilia says. “But only the audience can make total sense of this since things are always changing and trendy. The audience behavior of the past couple of weeks, for instance, always keeps the cross-sells fresh. And recommendations continually increase cross-selling opportunities.”
The behavior of a community of shoppers can indicate purchase intent in ways the past behavior of an individual cannot, contends Scott Brave, chief technology officer at Baynote.
“In one client’s case, when someone clicks on a red washing machine, she is presented with other colorful washing machines, as opposed to just red washing machines or including basic white ones, too,” Brave says. “This is because the community shopping this client’s site shows that people looking at one color of washing machine intend on purchasing a washing machine in some color, not just any washing machine or a washing machine of the same brand.”
When it comes to gaining insights from the group, shoppers at TheKnot.com threw executives for a loop. Recently, when shoppers looking for a groomsmen gift clicked on a flask, they were presented with recommendations for cigar cutters-and bridesmaid gifts. At first this did not seem to make much sense. Digging into the data, Baynote made a discovery about the customers.
“The patterns of behavior of the community showed us it was the same person buying both gifts-the bride was doing the shopping,” Brave says. “That was a connection we hadn’t thought about. The intent of these shoppers was not simply to buy a groomsmen gift, it was to buy gifts for groomsmen and bridesmaids.”
The other approach to making recommendations relies on the individual’s online behavior. Before a launch, a vendor using this approach may either build a profile based on historical data from an e-retailer or let the system run invisibly in the background for a period of time, gathering information to build profiles.
“We look at everything: what they look at, what they abandon, what they check out with, their response to recommendations we make, whether they take action on abandoned items in their cart, how and where they navigate, what they search, and so on,” says Paul Rosenblum, vice president of products and strategy at MyBuys. “From all of that we derive attributes about the consumer. What is their price point sensitivity, are they a sale buyer, are they a trend follower? It’s a truly one-to-one approach, a unique profile for each and every consumer.”
MyBuys technology also includes a group-oriented algorithm, but it uses it primarily for shoppers new to a site, for whom the system lacks historical information.
“We have found that when you treat people as individuals, they respond better than when you treat them with a groupthink mentality,” says Lisa Joy Rosner, vice president of marketing. “If the name of the game in recommendations is relevancy, you don’t get more relevant than serving the individual exactly what they want rather than taking a guess with what the group might like.”
Then again, why not use both? ChoiceStream, which runs individual- and group-oriented algorithms, believes the best results come when recommendations are based on an individual’s behavior. However, group recommendations can come in handy, says Toffer Winslow, executive vice president of sales and marketing.
“In many cases there is an additional layer of sophistication to be had by filtering the wisdom of the crowd through the prism of a shopper’s unique preferences,” he says. “Let’s say the wisdom of the crowd says a particular pair of brown pants happens to be highly correlated with an orange shirt. But when I look at the shirt, a more sophisticated recommendation engine would know I have never bought or even looked at anything that is orange, but that I do happen to buy yellow things. So it may then recommend that same shirt to me but in yellow.”
There is no one-size-fits-all solution to product recommendations, Winslow adds, which is why ChoiceStream has built a platform that can accommodate different approaches.
While it may one day employ the individual profile method, Borders opted for the group approach and uses the ChoiceStream system to offer recommendations based on the cumulative shopping patterns of millions of shoppers, Winslow says.
“The algorithm is running customer segments to match a customer up with other similar people to give us a wider set of data to draw from,” says Ertell of Borders. “There’s just not enough data for any individual person. As much as we all think we’re individuals, in reality there are tens of thousands of people who are really close to us. By tying people together as cohorts, we’re able to leverage a lot wider data set to provide better recommendations.”
When the cookie jar is empty