Years ago, at a presentation in front of hundreds of financial analysts, Amazon.com founder Jeff Bezos decided to show off his company's...

Share story


Years ago, at a presentation in front of hundreds of financial analysts, Amazon.com founder Jeff Bezos decided to show off his company’s personal-recommendation feature by pulling up his own account.

To his embarrassment, at the top of the list was the DVD for the B-movie “Slave Girls From Beyond Infinity.”

Now, don’t start thinking the head of one of the Internet’s most successful retailers was a secret fan of obscure skin flicks. As Bezos later explained, the suggestion appeared because he had bought “Barbarella” the week before. A computer simply made the leap that if one likes a film with Jane Fonda in skin-tight clothes, one must like all movies with women in skin-tight clothes.

Most Read Stories

Unlimited Digital Access. $1 for 4 weeks.

It can be argued the process by which Web sites attempt to deduce your tastes has come a long way since then. It also can be argued it hasn’t come far at all.

In a mass-media universe, the ability to match consumers with what they like has become an important contributor to retailers’ bottom lines.

What began in the 1990s with versions of Amazon’s “Customers who bought this also bought” feature has evolved into complicated tools designed to determine whether you’re more White Zombie or Barry White.

Not everyone is on the bandwagon, however.

Only 7 percent of online retailers’ sites can adapt to show products based on a customer’s past purchases, while 58 percent have no personalization features at all on their sites or via e-mail, according to a survey of about 250 online merchants released in April by Chicago consultants The E-tailing Group.

Only about a quarter of those surveyed said their sites even greet repeat customers by name.

Amazon has spent $650 million on researching technology and content — including personalization features — over the past three years, a spokesman said. The logic is that the more you appeal to customers’ tastes, the longer they’ll use whatever service you’re offering.

It’s why the Seattle company offers recommendations for its books, music and DVDs; why Netflix does it with movies; why TiVo does it with TV shows; why Google does it with text ads.

But, as Bezos’ “Slave Girls” example points out, it’s a double-edged sword.

Recommendation systems are only as good as the resources devoted to them, so while they’ve become more sophisticated, they still have a way to go, said Patti Freeman Evans, an online retail analyst at New York-based Jupiter Research.

“If they recommend something that doesn’t make sense — even though it makes sense in the algorithm of the recommendation system — it actually puts a damper on the validity of the recommendation,” Freeman Evans said.

“Consumers take them less seriously because of that.”

Gauging sales

One way to measure how well they work is to gauge the amount of sales the recommending systems generate. Trouble is, few companies release such data.

Amazon said it’s able to track such purchases, but declined to release related data.

Netflix, the online DVD rental service, has said 40 to 50 percent of its rentals can be traced to its recommendation system. Its system uses a program called CineMatch, which relies upon having members rate movies on a five-star scale, then offers suggestions based on similar ratings of others.

But Mike Kaltschnee of Danbury, Conn., who runs a blog dedicated to Netflix and other online rental services, began noticing problems: The system couldn’t differentiate between the independent films he likes and the Barney movies his two young daughters prefer, resulting in a mishmash of suggestions no one wanted.

So Netflix configured the system to accommodate individual profiles.

There’s still a catch, though: You have to rate your rentals for the system to work.

“What you put into it is what you get out,” said Kaltschnee, 40, who prefers Netflix’s system to those offered by other online rental services.

Tracking metrics are the key to these systems, because without the data, they’re useless, said Jack Aaronson, chief executive of the Aaronson Group, a New York customer-relations consultancy, and the former director of personalization at barnesandnoble.com.

With his clients, which include ShopNBC.com, Dell and Fingerhut.com, Aaronson tracks such data as the number of products per order, how often users return to the site, the average order size over time and how many people discovered products they bought through the recommendation system.

As data are compiled, the systems are able to tweak variables and monitor the results.

Collaborative filtering

At their base level, many systems rely on what’s called collaborative filtering, the method of recommending a product to someone based on the buying habits of everyone else who bought that same product.

Its most recognizable example is Amazon’s “Customers who bought this also bought” feature. For instance, those who bought “Slave Girls” also purchased “Cannibal Women in the Avocado Jungle of Death,” “Sorority Babes in the Slimeball Bowl-O-Rama” and “Assault of the Killer Bimbos.”

Collaborative filtering is strictly an examination of the “what” process; with no “why” it works well with what Aaronson called “limited vertical product spaces,” such as books or music.

When cross-recommendations come into play — suggesting a DVD to go with those books you have in your virtual shopping cart, for example — companies have to come up with other technological solutions.

One such company working on that problem is Cambridge, Mass.-based ChoiceStream, which has worked with America Online, Yahoo! and Columbia House to try to match people with products they may like.

Part of the process deals with assigning subjective attributes to pieces of information, said David McFarlane, executive vice president of sales and marketing.

For example, besides detailing a particular movie’s director, stars and genre, ChoiceStream may categorize it as “romantic,” “thought-provoking” or “mindless fun.”

Still, analysts said it may take years of data for personalization systems to become an important fixture in the online retailing world, assuming companies can come up with the right way to make use of that data.

Issue of privacy

There is also the issue of privacy. Companies such as TiVo and ChoiceStream say they don’t store or pass along private information when coming up with recommendations.

In July, ChoiceStream released a survey that said 81 percent of 673 respondents would like to receive personalized content, and 56 percent said they would provide demographic data in exchange for the service.

Aaronson, the consultant, said that while privacy advocates have compared personalization to Big Brotherlike tactics, the two need not be mutually exclusive.

“The highest form of personalization is the ability to turn itself entirely off,” Aaronson said. “A highly personalized system should be smart enough so that I, the user, can say, `You know what? I don’t want you sending me things.’ “