How Are YouTube and Netflix Recommendations Decided? An Introduction to Recommender Systems

Watch YouTube and the next video flows on naturally; open Netflix and the home screen is filled with titles that match your taste. In a shop, “items viewed together with this one” quietly tags along. None of this is a coincidence. A recommender system is at work behind the scenes.

In this post, I’ll unpack, without any code, the principles by which recommendations are decided. I’ll look, at a non-developer’s level, at what a recommendation ultimately computes and how it grows to fit you better over time.

A recommendation is, in the end, a search for similarity #

What a recommender system does is, in a word, find similarity. There are roughly two directions. One is to find people whose taste is similar to yours and recommend what they liked; the other is to find content similar to what you’ve already liked and recommend that.

We do both all the time in daily life. We’ll watch, on faith, a movie recommended by a friend whose taste lines up with ours, and we reach for a film by the same director as one we enjoyed. A recommender system carries out these two intuitions automatically and at scale, across countless people and items.

Following similar people #

The first approach uses people whose taste overlaps with yours as the clue. It’s commonly called collaborative filtering. Among what people who’ve liked similar things to you enjoyed, it picks out and recommends what you haven’t seen yet.

The familiar line “customers who bought this also bought” is exactly this approach. Without analyzing the content of each item one by one, a fairly plausible recommendation emerges from patterns where people’s behavior overlaps. The more people’s choices pile up, the more accurate this approach becomes.

Following the similarity of the content itself #

The second approach looks at how alike the content itself is. For a film, it weighs features like genre, cast, and mood to recommend similar titles. If you watched some action movie to the end, it recommends other action movies of a similar genre and cast.

This approach has the advantage of working even for new content where people’s behavior hasn’t accumulated yet. Even if a freshly uploaded video has no views, it can be recommended to the right people if its features are similar. Real services mix the two approaches.

Clicks and views tame the recommendations #

A recommender system isn’t something you set once and finish. What you clicked, how long you watched, whether you quit partway — these behaviors pile up as data again, and that data refines the next recommendation further. That’s why it feels like it fits you better the more you use it.

Here the earlier topic connects. For recommendations to get smart, user behavior has to be recorded well as events and organized into metrics. A recommender system runs on the data gathered that way as its fuel.

Recommendations can be tamed the wrong way too #

Recommendations aren’t always purely good. If it keeps recommending only things similar to what you’ve watched, you gradually see only similar content and your view narrows. This is the phenomenon often called the filter bubble.

Also, if people tend to linger longer on sensational content, recommendations may lean toward the sensational, following that behavior. A recommendation is a mirror reflecting people’s behavior, so it follows even the biases embedded in that behavior. That’s why many services deliberately mix in new kinds of content to strike a balance.

Why this makes work easier for non-developers #

• You view recommendations from a step back. Knowing that what’s on the home screen isn’t an objective correct answer but a reflection of your own behavior lets you choose without being dragged along.
• You use it in planning. Knowing that recommendations run on data as fuel, you can design in advance which behaviors should be recorded to make recommendations better.
• You can explain the limits. You can calmly explain to colleagues why recommendations sometimes lean one way, and why new content is hard to recommend.

Wrapping up #

Today we looked at how a recommender system, with YouTube and Netflix as the prime examples, builds recommendations using similar people and similar content as clues, and grows more refined by taking our behavior as fuel again. The key is that a recommendation is ultimately a search for similarity, and its raw material is the behavior we leave behind.

If you’re curious how the AI used in recommendations learns, read The Big Picture of AI, Machine Learning, and LLMs; if you want to know more about how the data and metrics that fuel recommendations accumulate, read How Do You Analyze Product Data and Logs.