Product recommendation is one of the most elusive–and potentially profitable–forms of merchandising online, where consumer behavior can bring in reams of data about the way people shop. But figuring out what to do with that data, and how to present what you learn, is a psychological challenge all its own.
No company has tackled personalized recommendations on a larger scale than Amazon, which has a 100-person team devoted solely to suggesting other products its customers might like. So the small-but-influential world of personalization paid attention when Amazon’s head of personalization from 2003-2004, David Selinger, left for Overstock, and learned to leverage impulse deal-buying as well as a big back catalog of products. After over a year at Overstock, Selinger left to co-found a company called RichRelevance in 2006, which would take what he had learned and turn it into an algorithmic tool he could sell.
The problem is that online customers are a particular and fickle bunch: even if you serve them up interesting product suggestions, you can still fail to get a good conversion rate if you don’t present the suggestions in a palatable but unobtrusive way. Do it right, and you can increase sales by 25%, as Amazon did under Selinger. Do it wrong, and you can alienate people or even piss them off.
“The user interface can be more important than the algorithm itself,” Selinger says. “For example, a lot of vendors don’t tell you where a recommendation comes from. At Amazon, we did a test: we began telling people why we recommended things to them in our emails, and we did this with a statistically significant group of people. We saw a 60% increase in conversion just from adding those words, and a 99% decrease in customer service contacts” that complained about being pushed products.
As with online advertising, transparency in product suggestions can create a premise of trust that makes customers more responsive and tolerant to recommendations, even if the recommendations themselves aren’t very good. In fact, says Selinger, the best recommendation systems are often the ones that don’t push the products that are a 100% match with the data.
“You have to build a system that will try something, and then react,” he says. “You build into your algorithms not only good suggestions, but this closed loop that says that if something doesn’t work, back off and try something else.” Selinger says that his systems actually try odd recommendations from the get-go with a small percentage of customers, just to see how they react. “If that recommendation starts to get more traffic, let’s hone in on that and change our original assumptions” about what makes a 100% match. “Our system is actively trying to learn.”
The try-anything approach has made RichRelevance the Web’s largest vendor in e-commerce recommendation, serving more gross sales than any other like company. Their newest product, called MyRecs, is a turn-key solution for major vendors: they give RichRelevance a basic page with the company’s navigation bars, and MyRecs fills in the page with a customized catalog based on an individual shopper’s habits. But instead of just tacking banners with related stuff to a list of like-tagged products, RichRelevance has tried to make their product suggestion engine a centerpiece of customer interaction, not a sidecar.
One of their largest customers, Sears, has begun using a search product Selinger’s company calls ClickSee, which creates a tiled array of like-product images around whatever you click. In the past, searching for something in a Web store meant getting a list of results and scrolling through pages and pages of products. That might work fine with Web search, which is mostly text-based–but begin to include detailed photos, color selection and pricing, and online shopping quickly becomes as stressful as its real-life counterpart. Selinger hopes that ClickSee can make image-based shopping quicker and more relevant, in a way that treats customers with the same delicacy and precison that quality online advertisers do.
But as shopping becomes more of a community exercise–with buyers rating, reviewing and recommending products through crowd-sourcing–using an algorithm to power product suggestion seems increasingly like over-engineering. RichRelevance is responding by letting vendors “boost” certain items that might have special relevance now–say, a book that was just put on Oprah’s reading list.
As for customer-driven boosts? “That’s something that is a little more noisy,” Selinger says. While they have feedback mechanisms that let individuals customize what kinds of suggestions they get from the computer, getting useful crowd participation is harder. “We haven’t quite figured out the user interface to ask a consumer ‘is this recommendation good or not?'” says Selinger. “We’re working on that.”