Why That Video? Decoding the Recommendation
How a video recommendation is computed: the score, the candidates, and the reason one video wins.
A complete interactive classroom, not just a preview.
Start when you are ready to enter this Stage's 7 scenes and explore, respond, and learn as you go.
What is happening when a video recommendation appears?
Every time you press play, an invisible judge has already ranked millions of videos.
Is it chance, popularity, or something about you?
A side-by-side simulation showing how a few signals decide the winning video.
The recommendation is a ranked score: your behavior plus video signals, not pure chance.
It feels random or driven only by global popularity.
- platform-specific algorithms
- training data pipeline
- policy and fairness interventions
- 01The Invisible Auto-PlayslideQuestion
You open an app and a video starts playing. You didn't search for it. Where did it come from? This investigation will turn that instant into an observable process.
- A recommendation happens in milliseconds
- It looks like magic, but it is a computation
- We will reverse-engineer it
- 02Your First GuessquizPrediction
Before we inspect the process, make a prediction. Which factor do you think has the biggest influence on whether a video recommendation appears?
- Commit to one explanation
- We will test it with evidence
- 03The Candidate Pool and the ScoresinteractiveEvidence
Imagine four videos are waiting in your feed. Each has a score for how well it matches you, a score for engagement, and a score for freshness. Move the sliders to see how changes in those signals reorder the list. One of these rankings is what the real recommendation likely used.
- Four candidates, three signals
- A small change can flip the ranking
- The top slot is earned by a score, not by chance
- 04The Two-Stage SecretslideExplanation
A recommendation is not one magic decision. First, the system pulls a small set of plausible videos from a massive library. Then a ranking model gives each video a score based on your history and the video's features. The highest score wins the screen.
- Stage 1: candidate generation
- Stage 2: personalized ranking
- The score combines match, engagement, and freshness
- 05What the Score Is NotslideBoundary
This score measures predicted engagement, not truth or quality. It can prefer a video that is mildly relevant and very popular over a perfect niche video. It also ignores many deeper questions about why you watched.
- Score is not 'best'
- Popularity can outweigh match
- Only visible signals are used
- 06Predict for a Different ViewerinteractiveTransfer
Change the user profile from 'new to cooking' to 'daily sports fan' and watch the same candidate videos re-rank. The algorithm's logic stays the same, but the output changes because the user history changes.
- Same candidate pool, different user
- User profile shifts the scores
- You can predict likely recommendations
- 07The Answer: It Was Ranked, Not Chosen at RandomslideResolution
The recommendation appeared because a ranking model scored millions of videos and chose the one with the highest estimated value for you. It used your past behavior, the video's features, and engagement patterns. That is what is happening when a video recommendation appears.
- A personalized score decides
- Your behavior is the input
- The top-ranked video wins the screen
Discussion threads for a Stage aren't available yet.
