Two people opening the same streaming app can see completely different images for the same film. The artwork is chosen by the system, from a set prepared in advance, based on what each viewer has watched.
Artwork is the last decision point
By the time a title appears on screen, the service has already decided to recommend it. The image determines whether the recommendation converts into a play.
Browsing happens quickly, and most rows are scanned rather than read. The image does nearly all the persuading before a title or description registers.
Because that moment is so short, small changes in artwork produce measurable changes in what a catalog gets watched. Services treat the image as a lever with a direct effect on viewing hours.
Multiple images are commissioned per title
Marketing teams supply many variants for a single title, each emphasizing a different aspect: a romantic pairing, an action set piece, a comedic performer, an ensemble.
These are not random crops. Each is built to represent a genuine element of the film so that the image does not misrepresent what follows.
The variants are then tested against viewer segments, and the system learns which image performs for which kind of viewing history.
Personalization is inferred, not declared
The service does not ask what appeals to a viewer. It infers from what that account has finished, abandoned and returned to.
Someone with a history of thrillers may see the tense frame from a film that also has a love story, while another account sees the couple.
Both images are accurate. They differ in which true thing they lead with.
The risk is a mismatch in expectation
Artwork that oversells one element attracts viewers who wanted that element and finds them a film that is mostly something else.
Those viewers stop watching early, and abandonment is tracked as carefully as plays. An image that wins clicks and loses completions is a bad image by the service's own measure.
That feedback loop is what keeps the practice from drifting into outright misrepresentation, at least for titles with enough viewing data to correct it.
Rows are personalized the same way
The same logic governs the ordering of rows and the labels attached to them, which are generated to describe clusters of similar titles rather than being fixed genres.
A viewer's home screen is therefore assembled fresh rather than retrieved, and no two accounts see the same interface. Even the same account can look different on a second visit the same evening.
What looks like a catalog is closer to a rendering of one, produced for a single person at a single moment.