Why a profitable product recommendation strategy requires more than just relevance
By Mary Wagner
E-retailers are better than ever at identifying the type of customer visiting an e-commerce site by browsing and purchasing behavior. And, by comparing her behavior with that of many others, that means they can serve up ever-more-specific, personalized recommendations of products likely to prompt her to click and buy.
But what happens when presenting the products she is most likely to respond to conflicts with the retailer’s business objectives of the moment? While it’s hard to argue that higher conversion rates are anything but good for an e-commerce site, profitability demands that e-retailers take into account other factors as well—for instance, the need to move 200 yellow sweaters that remain after the identical sweater in pink has sold out, or the desire to boost margins by selling more of the products the retailer purchased at favorable prices.
Personalized content can be pressed into service toward these goals, but retailers must carefully balance how they tweak personalization—and relevance—to accommodate business objectives.
“It’s a very interesting tension that exists between trying to feed customers what will optimally convert the highest versus the other things that are in your interest,” says Eyal Gutentag, principal at BestBuyEyeglasses.com.
And that balancing act extends beyond the recommendations e-retailers present on their sites to every form of marketing, down to the language of search engine ads.
Gutentag has first-hand experience with adjusting recommendations to meet business goals. BestBuyEyeglasses.com uses the product recommendation engine of vendor Richrelevance, which allows the retailer to determine which items are most likely to appeal to a customer, and then adjusts how those recommendations are presented according to business objectives.
Boosting AOV
Recently, for example, it has successfully used that method to increase average order value. In automating product recommendations based on customer data, it steps slightly outside what the data would suggest showing to a shopper viewing a “customers who viewed this also viewed this” display to favor higher-margin products that still fall within the criteria of what’s relevant to that customer.
During a recent test over four to six weeks the retailer increased sales of one of its higher-margin products—Dolce & Gabbana frames—by 21% using this strategy, Gutentag says.
To remain effective, the tweaked product recommendations must still be personalized to a shopper, based on what the retailer knows about her behavior and that of customers like her. Gutentag says that over-representing how frequently Dolce & Gabbana frames appeared among product recommendations by 5% to 10% was enough to drive the sales increase.
“Say you had a product on which you’d built up a significant amount of inventory. You might look at over-representing it not by 5% to 10%, but 25% to 30% in the recommendations. But you have to be careful because if you skew too far in that direction, you lose relevance and you’ve shot yourself in the foot. It’s a fine balancing act,” he says.
The increase in the number of D&G frames is automated, a business rule that rests on top of the product recommendations generated by the behavior of a substantial amount of traffic—the site receives hundreds of thousands of visits a month, providing statistically meaningful numbers for the Richrelevance engine to mine.
Gutentag adds this strategy has produced such positive results that the site may allocate more real estate to it in the future. Richrelevance offers pay-for-performance pricing based on sales driven by its recommendations, according to the vendor.
Rules rule
Many personalization products now make it possible to override the automated recommendations the systems produce, says Sucharita Mulpuru, principal analyst for retail e-business at Forrester Research Inc.
“Most of the personalization vendors have recognized what a good tool needs to be able to resolve, which is, what happens if the merchandiser wants to promote something, or doesn’t want to promote something, that is counter to what the algorithm says,” she says.
But merchants are still figuring out what works and how far they can go against what technology suggests. While BestBuyEyeglasses.com found that stepping outside the engine’s automated recommendations by no more than 10% was effective, there are no hard and fast rules.
“It varies from site to site. I think the bar now is just to be able to populate pages with relevant cross-sells that aren’t going to annoy vendor partners or present something to your customers that is going to make them suspicious because it’s something completely incongruent,” Mulpuru says.
While she says the personalization technology available has come a long way from the days of generic product recommendations, she adds, “we are still at the point where even automated cross-sells and upsells haven’t penetrated every sector of e-commerce.”
In personalizing product recommendations for its customers, Indigo Books and Music Inc., Canada’s largest multichannel seller of books, DVDs and CDs, looks for a balance between presenting the products its data suggests will be most interesting to customers and what it can deliver profitably.
“We do try to start with what is relevant and important to the customer—that’s the first filter we put on. The next is an inventory filter, and we can even get as sophisticated as to filter through certain price points or profit margins as well,” says Deirdre Horgan, executive vice president of marketing.
Horgan says the highly subjective judgments consumers make about products like books and music ruled out an out-of-the-box solution for personalization technology. Instead, the company built its own product recommendation engine and works with the computer science faculty of the University of Toronto on its ongoing development.
“Because someone bought one fiction book, it doesn’t necessarily mean they will like another genre of fiction as well,” Horgan explains. “You need to get granular with the content of the product so you can make recommendations that much more meaningful.”
To that end, Horgan says Indigo’s developers are looking for technology that would map the content of books so that the retailer can make more precise suggestions.
Indigo also is improving its ability to present personalized product recommendations with a preference center just added to its site. Customers can register the kinds of recommendations they are looking for—for instance, titles for children under age eight. “It takes some of the guesswork out of it for us.” Horgan says.
The products that are served by the recommendation engine also reflect where the inventory will be sourced from, with more profitable products served up first. While Horgan says the company still is in the “test and understand” phase of gauging consumer reaction to recommendations, the company’s goal is to deliver product recommendations that are both more profitable—taking into account inventory levels, re-order time and other factors—as well as relevant to its customers.
To date, Indigo has delivered its recommendations mostly through e-mail to its subscriber list. Horgan says that e-mail containing recommendations based on a customer’s past purchases or purchase behavior that is like that of similar customers has produced double the revenue of non-targeted e-mails.
Indigo also does some personalization on the site, with ongoing real-time customer data underlying a “customers who bought this, also bought this” recommendation attached to every item. And it’s leveraging the data gathered about its customers both online and offline for the print catalog it mails twice a year. The first several pages of the catalog vary, presenting products that are personalized to between 30 and 40 customer segments, Horgan says.
And over the next 12 to 18 months, Indigo hopes to enhance the kiosks recently added to its 300 stores to allow its registered customers to swipe their irewards loyalty program card at the kiosk and instantly receive a set of personalized recommendations.
“Over the last several years we have worked hard to try to improve and make recommendations even more relevant to the customers,” Horgan says. “So there is e-mail, there is the bi-annual publication, personalized recommendations are ongoing on the site, and the future vision is the brand-new kiosk program we have just launched in our stores.”
Personalized ads
PrezzyBox.com has found yet another way to make use of personalization technology. Using MarketMaestro personalization software from 7 Billion People Inc., the United Kingdom-based gift site takes personalization out on the web to its search ads.
In creating the ads, it uses language most appealing to the customer profiles the software’s analysis shows are most likely to be interested in what the retailer wants to promote for any reason—inventory levels, margin, or in a recent instance, products the retailer wanted to push for a Valentine’s Day promotion.
“We communicate an overall message—we don’t have specific messages for specific products,” says founder and general manager Zak Edwards. “We have tailored the marketing messages to appeal to as many profiles as possible.”
The software creates profiles of the types of customers that visit the site, and this guides writers in creating search ad copy designed to appeal to the top two or three profiles for a given item. Here’s an example: “Get her what every woman wants—the greatest choice of Valentine’s gifts at Prezzybox.”
That search ad would appeal to two types of shoppers, explains CEO Mark Nagaitis. One is the referential shopper who believes that because other women like Prezzybox products, so will the women he’s buying the gift for. The other is the shopper to whom range of choice is most important. In tests, Prezzybox has found it can double its conversion rate on search ads personalized by profile in this way, Edwards says.
7 Billion People also offers another product, WebLegend, which takes site personalization beyond product recommendations. It adapts how a web site’s pages are presented to site visitors in real time, based on how individual visitors like to receive information as determined by how they click through the site.
The software captures visitors’ clickstream data as they enter and move through the site, to create a portrait for each visitor. It uses behavioral psychology to analyze the portraits to identify each visitor’s goals, decision-making methods, buying behaviors and communications preferences.
Within three to five clicks, the software gathers enough information on the visitor to determine a behavioral tendency on at least one of 15 attributes, the vendor says. At that point, it begins to adapt the presentation of site content to fit the visitor’s preferences.
For example, one type of online shopper looking for a digital camera might enter the model name and go directly to the product page, then click straight to the technical specifications. If the shopper comes back to the home page the software would adjust display of the page to his personality, showing a simple search bar with minimal clutter and marketing offers. If this shopper goes to the product page, the technical specs would be prominently displayed and the checkout process would be straightforward without upsells.
By contrast, a shopper who browsed through a few cameras and clicked on a discount link would be presented a different experience, featuring user recommendations, deals and upsells in the checkout process, a style to which this type of shopper is more receptive, the vendor says.
“The deep personalization made possible by WebLegend is allowing our early clients to increase their profits by better serving the customers already coming to their web sites,” Naigitis says. The software already has helped one online marketer, United Kingdom-based car rental agency 121carhire.com, boost conversions, defined as online reservations, by 50%, according to the company.
Both products are delivered as software as a service, hosted by the vendor and accessed via the web, with MarketMaestro starting at $7,000 per month and WebLegend starting at $15,000.
That kind of technology investment is easier to justify if it helps a retailer not only identify consumers who like a particular style, but who are most likely to snap up those 200 slow-moving yellow sweaters that would otherwise have to be sold at a deep discount.
mary@verticalwebmedia.com
Click Here for the Internet Retailer Personalization Products & Services Guide
