Two people opening the same streaming service see different homepages, and the differences go deeper than the titles on offer. The order of the rows themselves is generated for each viewer.
Rows are ranked before the titles inside them
The system first decides which categories to show and in what sequence, drawing on what a viewer has watched, abandoned and browsed without selecting. Only afterwards does it choose which titles fill each row.
That two-stage structure matters because attention falls away sharply down the page. A title in a row placed fourth is seen by a fraction of the people who see the top row.
Moving a title up therefore means winning a row-level competition first, which is why the same film can feel prominent to one household and invisible to another on the same evening.
The artwork is selected as well
Services hold several images for a single title, each emphasising a different character, tone or genre cue, and they choose between them based on what a viewer has responded to before.
A drama with a romantic thread can be presented as either, and the framing changes how many people stop scrolling. The underlying programme is identical in both cases.
This is why a title someone dismissed months ago can reappear looking like something else entirely, without anything about the production having changed at all.
Licensing economics sit behind the ranking
Owned productions and licensed titles cost the service differently, and a licence often carries payment tied to how much it is watched within a defined period.
Promoting an owned title costs nothing beyond the screen space it occupies, while heavy promotion of an expensive licence can increase what the service owes for the same viewing.
Neither consideration appears anywhere on the page, but both shape which rows rise to the top when predicted engagement between two options is close.
Contracts can require visibility
Some agreements specify promotional treatment, guaranteeing a title a place on the homepage or in a launch banner for an agreed window after it arrives.
Those obligations override the personalised ranking for the duration, which is the usual explanation for a prominent recommendation that fits nobody's viewing history.
Once the guaranteed period ends, the title drops back into the ordinary competition and can disappear from view within days.
The system optimises for the session, not the year
Ranking models are trained to get a viewer watching quickly, because a session that ends without a selection is the outcome the service most wants to avoid.
That objective favours the familiar and the easily started over the demanding, and it compounds as each choice feeds back into the next set of predictions.
Deliberate browsing, searching directly rather than scrolling, is the only reliable way for a viewer to escape a loop the interface is designed to keep them inside.