The power of recommendations
September 30, 2016
Exploring the impact of ‘People who bought this also bought…’ and other recommender systems
One very powerful marketing tool has a big impact on the way you use the internet.
It’s called a recommender system, and you see examples every day, like Amazon’s “People who bought this also bought this” section, Netflix and YouTube’s viewing suggestions, Google News predictions and Pandora’s custom radio stations.
Each of these systems has the same common purpose: using information about users to predict items of interest. And the effects of recommender systems are strong.
“In fact, some studies in lab settings show that recommender systems can have an even greater impact than human recommendations can,” said Kartik Hosanagar at a recent University of Delaware lecture. “And there is pretty good evidence that they have a big impact on consumer choice, and an amazing economic impact on firms.”
For example, Hosanagar added, 80 percent of videos streamed on Netflix originate through personalized recommendations. At Amazon, about 35 percent of sales originate through personalized recommendations. This means billions of dollars in economic impact.
Hosanagar, a professor at the Wharton School of the University of Pennsylvania, an award-winning digital economy expert and one of the “40 under 40” (the world’s top 40 business professors under 40), studies these powerful systems in his research.
This week, he shared his latest research on the impact of recommender systems on consumer behavior and sales at a lecture hosted by the Institute for Financial Services Analytics at UD’s Alfred Lerner College of Business and Economics.
“These systems are a very interesting piece of marketing technology,” Hosanagar said. “They are a bit rare… because they offer very clear value to both consumers and firms. For consumers, they help us learn about new products, and they help us sort through very large choices.”
“For firms, they help in converting browsers to buyers, they help in cross-selling products, and they help increase customer loyalty by providing a personalized and custom experience on these websites.”
This usefulness means that most major online firms use recommender systems in some form. But a number of questions about recommender systems remain unanswered.
For example, “How do they change the distribution of sales?” Hosanagar asked. “In other words, what kinds of products gain versus lose market share because of recommender systems?”
One view is that since recommender systems help consumers to discover new products, they increase the diversity of products sold. Another view suggests that recommender systems reinforce the popularity of already popular items, creating a “rich get richer” effect.
“There has been no empirical study to resolve this theoretical debate that has existed for some time,” Hosanagar said. His recent study seeks to address this gap.
Hosanagar’s team partnered with one of the top five e-commerce retailers in Canada, which was not yet utilizing recommender systems. The team ran a field experiment over two-week period with a sample size of 846,000 users.
The study’s findings suggested that recommender systems did not increase sales diversity at the aggregate level.
When he shares this finding, Hosanagar said, people are often surprised and share personal stories of recommender systems helping them to discover new products that they never would have considered otherwise.
Interestingly, this phenomenon can be seen in the study’s data: The study also found that recommender systems can help to increase diversity at the individual level, just not at the aggregate level.
“It’s odd that sometimes sales diversity increases at the individual level, but when I add all these people up… the graph, at the collective level, is actually going downwards,” Hosanagar said. “So how do we reconcile this aggregate versus individual effects?”
Hosanagar’s explanation: “At the individual level, the recommender system is indeed exposing us to new items, and so we are individually clearly exploring. But our explorations are highly correlated… because now I’m being instructed or encouraged to explore in a direction others have explored.”
“’People who bought this also bought this’ is effectively telling me: ‘There’s all these ways in which you can explore,’” Hosanagar continued. “And indeed, I do explore in a new way. But the new way in which I explore is also the same way others have explored… So once you add it all up, it doesn’t show up at the aggregate level.”
Interestingly, what type of product you buy can also have an impact on the effect of recommender systems on your shopping experience.
For example, the recommender system has a greater impact on “hedonic” items, items that exist for pleasure (like perfume or movies), and a smaller impact on “utilitarian” items that fill basic needs (like soap or paperclips).
This means, Hosanagar explains, that a recommender system might have a greater impact on a wine seller than on an electronics seller.
He added that findings like these have implications for multiple industries.
For marketers, for example: “If these systems are effective in reducing consumer search costs and helping consumers to find better product matches, they are attractive marketing tools.”
For IT: “Do tacit design choices we make have an unknown side effect in terms of choices people make?”
And for the operations management industry, sales diversity is an important factor in terms of the product assortment and inventory levels that firms should carry.