Why recommendation systems narrow taste while widening catalogue

A service adds forty thousand titles in a year. Individual users report seeing the same two hundred. Both facts are true, and they are produced by the same system: one number describes what exists and the other describes what is reachable from the home screen.

What everyone sees

A larger catalogue is presented as more choice, and the presentation is literally accurate. Users describe the service as having everything and simultaneously complain that there is nothing to watch. The contradiction is usually attributed to indecision rather than to the ranking layer sitting between the catalogue and the user.

What is actually happening

Collaborative filtering recommends what similar users engaged with, so items with more interaction data are easier to recommend confidently and new items begin with none. Fleder and Hosanagar showed that recommenders can reduce aggregate diversity even while helping individuals, and Celma and Cano found popularity bias in music recommendation graphs. The catalogue grows in the database while the reachable set is shaped by the confidence the model has, which is a function of existing attention.

Why it stays hidden

The narrowing hides because the interface reports abundance. The catalogue count is advertised, the reachable set is not measured, and no user can see the titles they were never shown. Search offers an apparent escape and mostly does not provide one, since it requires knowing the name of something the recommender did not mention.