How does a video app decide what to show next?

Answer first

A video app decides what to show you next by scoring videos. First it picks a much smaller list from its huge library. Then it scores videos using clues such as past viewing and the video’s title and length. The top scores are shown, and then the list is checked again.

In one line: A program that picks a short list of videos from a huge library, scores each one for a viewer and shows the top few next.

Last reviewed: 25 October 2026 3 min read

What decision does it make?

When a video ends or you open the app, it makes one choice: which few videos should go at the top of the list for this viewer? It cannot score every video in the library, so it narrows them down first.

Decision	Possible answers
What should this viewer watch next?	A short, ranked list of videos, with the highest-scoring ones at the top.

What does it look at?

Past viewing. The program learns from what viewers chose to watch.
The title. A title can be a clue when a program predicts clicks.
Video length. A video’s length can be a clue when a program predicts watch time.

How does it learn the rule?

Collect examples. The program can learn from how viewers clicked and watched before.
Make a short list. For each viewer, it picks a much smaller set of possible videos from the whole library.
Score each video. A model can predict clicks or watch time for videos on the short list, then rank them.
Check the list again. Before showing the list, the app can adjust it to add variety.

Where can it go wrong?

It rewards tricks. If it scores videos by clicks, a video with a shocking title can win even when people leave it straight away.
It favors long videos. If it scores by minutes watched, long videos can win over short ones that were just as good.
It shows more of the same. A list can have too many similar videos. Re-ranking can add variety.

Try it

Toy example with made-up scores: you are the app. Move the sliders to decide how much you care about each clue, and see which video goes to the top.

A toy example: four videos, each scored from 0 to 5 on three clues. The numbers are made up for practice.

Item	Values	Score at starting weights
Next episode of a cooking series you finished	Like what you watched (0 to 5): 5; Clicks it gets (0 to 5): 2; Minutes people watch (0 to 5): 3	23
Science experiment video	Like what you watched (0 to 5): 3; Clicks it gets (0 to 5): 5; Minutes people watch (0 to 5): 3	20
Shocking title, most people leave after a few seconds	Like what you watched (0 to 5): 3; Clicks it gets (0 to 5): 4; Minutes people watch (0 to 5): 0	13
Two-hour video on a topic you like	Like what you watched (0 to 5): 4; Clicks it gets (0 to 5): 1; Minutes people watch (0 to 5): 4	21

Check it

What does a video app decide?

Answer: Which videos to show you next. It ranks videos and puts the top scores at the top of your list.

The app only cares about minutes watched. Which video is likely to win?

Answer: A very long video. Long videos rack up more minutes, even if a short one was just as good.

Why does the app check the list again before showing it?

Answer: To add variety to the list. A last check can add variety so the list is not all the same.

For older readers

This is a recommendation system, often built in three stages. Candidate generation narrows the whole catalog to a small set. Scoring uses a model, trained on past viewing behavior, to predict how likely a viewer is to click on or keep watching each candidate, using clues about the video such as its title and length. Re-ranking then adjusts the list again, for example to add variety. Choosing what the model predicts, clicks or watch time, is a real design decision with side effects.

Questions

A ranked list can show many similar videos. Some apps re-rank the list to add variety.

It does not know you the way a friend does. It predicts clicks or watch time from past viewing behavior and clues about each video.

If an app scores videos partly by clicks, a shocking title can push a video up the list. That is why the people who build the app have to choose its clues carefully.