Guy Aridor, Winston Chou, Nathan Kallus, Antoine Scheid, Allen Tran, and Kevin Zielnicki
When a recommendation system gets better, what happens to culture?
One familiar worry is that algorithms narrow our attention. A platform learns what succeeds, recommends more of it, and gradually funnels everyone toward the same handful of blockbusters. Another theory predicts something stranger: algorithms favor both the biggest hits and the most obscure niches, hollowing out everything in between.
Our paper examines these ideas using an experiment involving 8.5 million Netflix members. Some members received improved recommendations; others continued to use the existing system. This allowed us to study not merely whether better recommendations increased viewing, but how they changed what people chose to watch.
The improved system did increase consumption. Members also relied more heavily on recommendations, which is reassuring evidence that the recommendations were actually useful. But the additional viewing did not concentrate around Netflix’s largest “superstar” titles.
Instead, consumption spread toward the middle of the catalog: titles that were neither global sensations nor highly obscure. These middle-tail titles are plentiful and potentially appealing, but easy to overlook. A better recommendation system appears to help them find the people most likely to enjoy them. The deepest niches, meanwhile, saw relatively little change.
This distinction matters. A catalog is not just a collection of titles; it is also a problem of matching. Popular titles need little help being discovered, while extremely niche titles may have intrinsically small audiences. The middle is where improved information can have the greatest effect: many titles have plausible audiences, but those audiences do not yet know where to look.
Recommendation systems are often described as machines that predict taste. Our results suggest another description: they are infrastructure for distributing attention. When that infrastructure improves, the biggest beneficiaries may not be the hits or the hidden curiosities, but the broad and easily missed territory between them.