The more accurate your product recommendations, the more likely you are to capture add-on sales. With collaborative filtering tools, you can drive personalization to the next level
Last year 800.com, a multichannel marketer of home entertainment equipment, increased its holiday sales 61%, bringing its revenue for the quarter up to about $22 million. At the same time, the Portland, OR-based firm cut its marketing expenses about 30%, says founder/president Greg Drew.
How did 800.com do it? Drew credits the implementation of several direct marketing and personalization tools, including a collaborative filtering application. “One of the things that we fundamentally believe is that if we’re going to be communicating with our customers on a one-on-one basis, we have to be as targeted as possible,” Drew says.
For instance, at the bottom of one of 800.com’s product pages, customers will see a box titled “You may also like.…” It’s here where the recommendations generated by the collaborative filtering engine show up.
“On the site are 61 categories of merchandise, from digital cameras to MP3 players to car audio equipment,” Drew notes. “It’s unproductive for us and the customers to be barraged by a set of products not related to the areas they’re interested in.”
Collaborative filtering helps 800.com and other online marketers home in on their targets. The software looks at purchases a customer makes and compares them to similar purchases by other shoppers. Based on the items that those buyers have purchased, the collaborative filtering tool makes a recommendation that’s apt to be in line with the new customer’s interests.
The recommendation will be narrower than any made by a rules-based tool, which typically suggests products by category. For instance, if a customer is looking at shoes, a rules-based tool may suggest that the shopper link to the page for socks.
But let’s say a customer buys a novel by John Grisham from a Website with collaborative filtering. The collaborative filtering engine will examine what other books customers who purchased Grisham novels also bought. If many also picked up books by writer Scott Turow, the system will recommend Turow’s books to the Grisham customer.
“Collaborative filtering compares what you know about someone’s preferences to the preferences of the customer base as a whole,” says Steve VanTassel, vice president of commerce solutions with Net Perceptions, an Edina, MN-based developer of personalization applications.
The concept behind collaborative filtering parallels what direct marketers have been doing offline for years, says Jim Hathaway, director of business development for Graphica, an advertising and marketing firm in Chester, NJ. “You’re taking snapshots of customers and seeing similar patterns among groups to predict what people will be interested in buying.”
Is it right for you?
Collaborative filtering works best for marketers with large numbers of products and customers. Rules-based tools, in contrast, work best with a limited number of products; as product choices grow larger, they become unwieldy. While it is difficult to state a minimum number of products at which collaborative filtering makes sense, VanTassel says many of Net Perceptions’ customers have at least 500 products, and several offer several hundred thousand items.
The technology is particularly effective with products in which taste plays a key role in the purchasing decision. “If I like a certain musician or movie, someone else like me will probably like something I like,” says Andrew Dorward, the New York-based director of personalization with Germany-based Bertelsmann Online (BOL).
BOL began working with Net Perceptions to offer collaborative filtering in June 1998, Dorward says. BOL sells books, movies, and music to customers in 14 countries. “We needed to find a solution that would work for a very large store with millions of customers worldwide,” Dorward notes.
Collaborative filtering on Bertelsmann Online analyzes both purchases that customers make and information that they volunteer, such as favorite authors or singers. Dorward says about 12% of customers contribute such information on themselves: “The collaborative filtering process works by correlating their explicit info against other users’ behavior.”
While Dorward won’t reveal how much BOL spent to implement collaborative filtering, VanTassel estimates that most customers will spend about $500,000 and several months to get it up and running.
800.com is using the Real-Time Personalization Server from E.piphany, a San Mateo, CA-based software company. Several employees, from both 800.com and E.piphany, spent about two months installing the system. Neither company will disclose the cost.
While it’s difficult to state exactly what sort of response an online marketer will get with collaborative filtering, VanTassel says that some of Net Perceptions’ clients have seen increases of more than 700% in the conversion rates of customers responding to cross-sell offers. While the firm doesn’t guarantee such impressive returns, i.merchants that implement collaborative filtering typically see a conversion-rate boost of at least 20%, VanTassel says.
Just one tool in the box
Despite the success marketers have had with collaborative filtering, the technology isn’t a cure-all. For one thing, the software may not realize when someone is buying a gift. If the application relies on that purchase to make a recommendation, the recommendation is likely to be irrelevant. Retailers can work around this by offering customers free gift cards or wrapping. Presumably, customers who accept the offer will be buying for someone else, says Graphica’s Hathaway. The application can then make note of that information.
And retailers with extensive product lines that aren’t related — such as general merchants that sell everything from auto produs to shampoo — may not find collaborative filtering especially effective, says Walter Janowski, research director with Stamford, CT-based research firm GartnerGroup. There just aren’t easily discernible relationships among the diverse product groups, he says.
At the same time, merchants selling products that have a functional relationship between them may also find collaborative filtering not effective enough to justify the investment. If you sell computer equipment, for instance, a customer’s choice of processor will influence the type of peripherals he should buy. In that case, a rules-based application probably would suffice.
Finally, collaborative filtering by itself doesn’t always effectively represent the merchant’s interests, VanTassel notes. If you want to show customers products they’ll like but also feature items with higher gross margins, collaborative filtering alone can’t make that sort of determination. You’d need to run a rules-based tool as well.
NetFlix, an online DVD rental firm, does just that, says Marc Randolph, executive producer with the Los Gatos, CA-based firm. While NetFlix’s home-grown collaborative filtering engine aids in suggesting movie titles to its 300,000 subscribers, its rules-based application makes sure that the site doesn’t recommend movies a customer already has seen or that are out of stock.
Another personalization tool would be a statistical application. This would analyze sales patterns and customer demographics to determine what offers and products to show a particular customer.
Predictive technology will be the next wave in personalization tools, says David Sims, senior product manager with software provider Manna, based in Wellesley, MA. Predictive technology uses information available on a customer, such as a clickstream pattern, to determine, for instance, whether he or she is likely to abandon the sale. If that’s the case, the application could present the customer with a gift for finishing the transaction.
In the meantime, Dorward of Bertelsmann Online says he’s happy with the return on BOL’s investment in collaborative filtering. Because the site incorporated collaborative filtering from the start, he can’t cite before-and-after performance figures. But one measure of its value, he says, is that customers who offer information on themselves, which the system can use to make recommendations, buy twice as much as customers who don’t provide such information.
Karen M. Kroll is a freelance writer based in Minnetonka, MN.
Sourcebook
The following companies offer collaborative filtering packages:
E.piphany
www.epiphany.com
San Mateo, CA 650-356-3800
Gustos
www.gustos.com
Laguna Hills, CA 760-598-8555
Macromedia
www.macromedia.com
San Francisco 415-252-2000
Manna
www.mannainc.com
Wellesley, MA 781-304-1600
Net Perceptions
www.netperceptions.com
Edina, MN 952-842-5000
Before you begin…
For collaborative filtering to work, the application needs data to work with. “You need the ability to understand past patterns,” says Mark Kanok, product marketing manager with software provider E.piphany. That is, the software needs customer purchase information before it can make recommendations that make sense.
But if you’re just implementing collaborative filtering, you can work around your lack of existing data. When Bertelsmann Online began selling music CDs, director of personalization Andrew Dorward obtained customer purchase information from a CD-rental firm to populate his own database. “I created some fake purchase data,” he says. “So when real customers came into the store, I could compare them to the false rating system.”
— KK