Recommendation systems select what appears in feeds, catalogues and search results. Their design produces a narrowing effect that is a consequence of the method rather than an intention behind it.
The system predicts engagement, not quality
A recommender estimates the probability that a given user will act on a given item, using signals such as clicks, watch time and completion.
It has no representation of whether something is accurate, useful or worth a person's time, because those properties are not present in the training signal.
What it optimises is therefore a measurable proxy, and any gap between that proxy and actual value is inherited directly by the output.
Similarity between users does the work
Collaborative filtering identifies users whose behaviour resembles yours and recommends what they engaged with and you have not seen.
This works well because tastes cluster, and it requires no understanding of the items themselves beyond who interacted with them.
It also means recommendations reflect the behaviour of a group, so unusual interests are served poorly until enough similar users exist.
Feedback loops concentrate attention
The system learns from interactions with items it chose to show, so items never shown generate no evidence that they would have performed well.
Popular items accumulate more data, which makes predictions about them more confident, which leads to them being shown more often.
This self-reinforcement is why catalogues with vast inventories still direct most attention toward a comparatively small set of items.
Exploration is added deliberately
To counter this, systems reserve a portion of recommendations for items the model is uncertain about, accepting weaker short-term performance for better information.
Exploration is what allows new items to enter circulation at all, since without it nothing outside the established set is ever tested.
The balance between exploring and exploiting is tuned continuously, and it determines how quickly a catalogue's long tail is discovered.
The narrowing is measurable and adjustable
Operators track diversity of what is served, coverage of the catalogue and how much any individual's recommendations change over time.
Adjusting for these has a cost in the engagement metric, which is why the trade-off is a policy decision rather than a technical one.
Where a system sits on that trade-off is generally not visible to users, which is why identical interfaces can produce very different experiences.