To test whether the initial engine was indeed able to factor in a shopper's intent, we searched for Apple. Understandably, the results were about iPhones, iPads and other hardware from that company. But we then searched for Oranges, Bananas, Pears and then again asked it about Apple, hoping that it would now understand our desire for the Garden of Eden edible type of apple, which Walmart.com does indeed sell. Nope, it didn't take the hint and continued to display mobile computers. (When we searched for "fruit," the desired apples did materialize in the results.)
Sri Subramaniam, vice president at WalmartLabs, said that the new search engine—called Polaris—factors in many variables that the earlier engine had not. But for the type of inquiry context where the engine considers all the shopper's questions and then changes its answers accordingly, that will have to wait for "the next couple of months." That is when the engine will ramp up its personalization efforts, he said.
That's understandable, but it's not what Walmart bragged about in its announcement on August 30, which claimed that the new engine "uses semantic search technology to anticipate the intent of a shopper's search to deliver highly relevant results for them" and that it "focuses on engagement understanding, which takes into account how a user is behaving with the site to surface the best results for them."
Right now, the context Polaris uses is based on the fact that the shopper is visiting Walmart.com. Using the apple example, he said, Walmart sells far more iPhones and iPads than whole, sliced or canned apples. As such, it's reasonable to assume that a Walmart.com visitor's apple search is for purposes of mobile, not meal.
In short, Walmart instructs its engine to first look at sales reports and to then factor product popularity into its analysis of what answer the user probably is seeking.
Specifically, Polaris decides what results to display based on product popularity ("last seven days of sales and last one year of sales"), the number of shoppers who have clicked or purchased that product, product user-generated ratings, the number of Facebook likes, what searches are popular from specific IP address ranges (suggesting geography) and other factors, Subramaniam said. "There are about 100 signals we use, not just text," he said. That click analysis is updated every few hours.Subramaniam, who used to work on the eBay search engine team, said this illustrates some of the differences in the type of search engine issues that exist for Google, Bing or Yahoo (everything from everywhere) to those that apply to Amazon or eBay (limited to E-Commerce inquiries, but for all of their and their third-party sellers' products) to those that apply to a Walmart or Target (limited to the items offered by a single merchant).
Subramaniam argued that the single-retailer engine is actually far more challenging because of the fewer products it can offer. That's somewhat counter-intuitive, in that mastering a smaller database would seem easier. But the Polaris goal is to correctly answer the question.
If a user, for example, is asking for a specific type of running shoe, Google can simply return any results referencing that shoe. But if Walmart.com doesn't carry that specific shoe, it needs to have a database of umpteen products it does not sell so that it can recognize those brand names and be able to offer a similar product it does sell. From a programmer's perspective, that's a much more difficult task.
"Even if we have relatively fewer items compared with Amazon or eBay, it's a harder problem for us to satisfy the end user. Google can find those exact pages," Subramaniam said. "The challenge is more on the interpretation of the user query, not on delivering literally what they ask for."
And doing that is more difficult because Walmart.com's control over the product descriptions it has is not that much stronger than Amazon. On the plus side, it only sells its own products. But it still has to deal with whatever product descriptions and information its huge number of suppliers delivers. Keeping those descriptions uniformly consistent—let alone comprehensive, delivering all the datapoints Walmart's new search engine wants—is not a lot of fun.
Other challenges exist, many of which are based on shopper psychology. When trying to personalize an engine, that becomes crucial but also very tricky. Let's say the system has learned over time that this shopper is a man who is an electronics geek. It then interprets inquiries for that person. But what if this geek is buying a gift for his girlfriend or wife? "You don't really know what the user means," Subramaniam said.
That's some of the rationale for holding off personalization for the engine's next version. "What we don't have yet is, 'What did you type in the session?' That is on our roadmap, factoring in the user's history," he said, adding that for the initial version of Polaris, "we decided to play it safe and keep it anonymous and aggregate at first."