While previous research has established that personalized content recommendations increase product sales and consumer engagement, there has thus far not been consensus on whether personalized recommendations will increase or decrease sales/consumption diversity. In this paper, we present results from a large-scale, randomized field experiment on one of the world’s leading platforms for streaming. In the experiment, both control and treatment users were given podcast recommendations. However, the recommendations provided to treatment users were personalized based on their previous activity, whereas control users were recommended the most popular podcasts among those in their demographic group. Consistent with previous studies, we find that the treatment increased the number of users streaming podcasts by 38.29%, and increased the average number of podcast streams per user by 30.67%. We also find that the treatment decreased the individual-level diversity of podcast streams by 2.83%, and increased the aggregate diversity of podcast streams by 3.04%. The treatment effects we observe are largely driven by changes in the content that users stream from the surface upon which recommendations are delivered, and do not persist beyond the conclusion of the experiment. Our results provide evidence of an “engagement-diversity trap”; while personalized recommendations may increase user engagement, they may also create costs for the firm due to unwanted public scrutiny and cause societal harm. Our findings also suggest that online echo chambers and radicalization pathways may, in large part, be attributable to recommendation algorithms, as opposed to user choice.
David Holtz is a doctoral candidate in the information technologies group at Sloan School of Management, MIT. His research interests span online marketplace design, causal inference, applied machine learning, and network science.