Steering via algorithmic recommendations

Abstract

This article studies self-preferencing in algorithmic recommendations on dominant platforms, focusing on Amazon's dual role as platform owner and retailer. We find that products sold by Amazon receive substantially more “Frequently Bought Together” recommendations across popularity deciles. To establish causality, we exploit within-product variation generated by Amazon stockouts. We find that when Amazon is out of stock, identical products sold by third-party sellers face an eight-percentage-point decrease in the probability of receiving a recommendation. The pattern can be explained by the economic incentives of steering but not explained by consumer preference. Furthermore, the steering lowers recommendation efficiency.