How Streaming Services Decide What You See
No two people see the same Netflix, YouTube or Spotify home screen. Behind every row and thumbnail is a recommendation system learning from what you watch and skip. Here’s how it works, and how to take back some control.
Daily Life Buzz Editorial 5 min readExplainer
In this article6 sections
Open Netflix on your phone and on a friend’s, and the screens look completely different. Different rows, different order, sometimes even different pictures for the same film. On YouTube and Spotify it’s the same: your home page is built for you, and only you.
None of that is chosen by a person. It’s the work of recommendation systems, some of the most valuable software these companies own.
Why streaming services care so much
A streaming service has far more to offer than anyone can browse. If you can’t find something you want quickly, you might give up, and eventually cancel. So the home screen is designed to show you, within a few seconds, things you are likely to enjoy.
It works. In a 2015 paper, two Netflix executives, Carlos Gomez-Uribe and Neil Hunt, wrote that about 80% of the hours streamed on Netflix came from recommendations rather than search. They estimated that personalisation and recommendations saved the company more than a billion dollars a year by keeping subscribers from leaving.
The home screen is designed to show you, within a few seconds, things you are likely to enjoy.
What the system learns from
Recommendation systems mostly learn from what you do, not what you say you like. Signals can include:
- What you play, and whether you finish it or stop after a few minutes
- What you skip or scroll past
- What you search for
- Ratings, such as thumbs up or down, likes and saves
- When and how you watch: time of day, which device, how long your sessions last
- What the content is: its genre, cast, mood, language and other labels, often added by people or by software
Behaviour usually counts for more than ratings, because people often rate serious documentaries highly but actually watch light comedies.
How the predictions are made
People like you
The classic method is called collaborative filtering. If you and thousands of other people watched and finished many of the same shows, the system assumes you have similar taste. It then recommends things they enjoyed that you haven’t seen yet. It doesn’t need to know why you like something, only that people like you do.
Things like this
A second approach, content-based filtering, looks at the features of what you watched. If you finish a lot of slow-burning Scandinavian crime dramas, it will suggest more titles with those features.
Picking and ranking
Real systems combine both methods, and many more, using machine learning. They typically work in two stages. First they pick a few hundred possible titles from a huge catalogue. Then they rank those candidates by how likely you are to watch and enjoy each one. YouTube engineers described this two-stage design in a 2016 paper.
The ranking also decides the layout: which rows you see, in what order, and which titles go first in each row, because things at the top left of the screen are seen most.
Why the picture changes too
Netflix has explained that it can show different artwork for the same title to different people. Someone who watches a lot of romance might see a picture of the two leads together; someone who watches comedy might see the film’s funniest actor. The aim is to show you the reason you, specifically, might want to watch it.
The downsides
- Narrow choices. Systems that learn from your past can keep showing you more of the same, making it harder to discover something different.
- Shared accounts get confused. A child’s cartoons and an adult’s thrillers on one profile produce strange mixtures.
- What you see isn’t neutral. Services also want you to watch their own new shows, so recommendations reflect business goals as well as your taste.
How to take back some control
- Use separate profiles for each person, and a children’s profile for children.
- Rate things, especially things you didn’t like. Thumbs down is a strong signal.
- Remove titles from your viewing history if you watched something once and don’t want more like it. Most major services let you do this in your account settings.
- Search and browse deliberately now and then. What you choose feeds back into future suggestions.
- In the EU, the Digital Services Act requires very large online platforms, such as YouTube and TikTok, to offer at least one way to see content that isn’t based on profiling you.
Frequently asked questions
Do streaming services listen to my conversations to recommend things?
There is no good evidence that the major streaming services do this, and they don’t need to. What you watch, search for and skip already tells them a great deal.
Why do I keep seeing things I’ve already watched?
Systems sometimes recommend rewatching because many people do rewatch comfort shows. Rating or removing a title from your history usually helps.
Is the algorithm the same for everyone?
The software is the same, but because it learns from each person’s behaviour, its output is different for everyone. That is the whole point.
Sources
- Gomez-Uribe, C. A. and Hunt, N. (2015). The Netflix recommender system: algorithms, business value, and innovation. ACM Transactions on Management Information Systems, 6(4), Article 13.
- Covington, P., Adams, J. and Sargin, E. (2016). Deep neural networks for YouTube recommendations. Proceedings of the 10th ACM Conference on Recommender Systems, 191–198.
- Netflix Technology Blog: Artwork personalization at Netflix (2017)
- Netflix Help Center: How Netflix’s recommendations system works
- European Commission: The Digital Services Act
Fact-checked against the sources listed above. Spotted an error? Tell us. Read our editorial policy.