Netflix Knows What You Want to Watch Before You Do — And It Has Nothing to Do With Stars or Rating

Hi everyone, I am Trends 24/7 and in this blog I’m going to walk you through something that genuinely surprised me when I first looked into it: the actual engineering behind why Netflix feels like it reads your mind.

Not the surface-level “oh it uses AI” explanation you see everywhere. The real thing.

The 90-second problem Netflix is quietly solving every time you open the app

Here’s a number that changed how I think about Netflix as a product: 90 seconds.

That’s how long you have before an average user gives up, closes the app, and finds something else to do. Not an hour. Not ten minutes. Ninety seconds of scrolling before the decision becomes “forget it.”

Netflix knows this. Their engineers have known it for years. And the entire recommendation system, all the machine learning, all the graph engineering, exists to solve this one problem inside that 90-second window.

80% of what gets watched on Netflix comes from algorithmic recommendations, not search. Users rarely type anything into the search bar. They scroll, they get served something, they click or they don’t. That 80% figure is why the company estimates the recommendation engine saves them over $1 billion a year in subscriber retention. Not revenue from recommendations directly but revenue from people not cancelling.

When you frame it that way, the whole system starts making more sense.

Why a simple “people who liked X also liked Y” approach doesn’t cut it

The obvious version of a recommendation engine works like a spreadsheet. You watched Stranger Things. Other people who watched Stranger Things also watched Dark. Therefore: Dark.

That’s called collaborative filtering, and Netflix still uses it as one input. But it breaks down fast.

What if you watched Stranger Things because your kid wanted to, and you’re personally more interested in French political dramas? What if the show you actually loved was one you abandoned after episode 3 because life got busy, and the system logged that as a non-completion? What if you watch completely different things on Tuesday nights versus Friday nights, and the system is mixing those signals together?

Collaborative filtering alone can’t separate any of that. It just sees the co-occurrence: user A and user B both watched these ten things, so serve them from the same bucket.

The graph is what changes this.

What a graph actually means here (and why it matters)

When Netflix talks about graph neural networks, which they’ve been moving toward increasingly since around 2020, they’re treating their entire catalog and user base as a connected web of relationships.

Not a table. A web.

Every title is a node. Every user is a node. Every time someone watches something, pauses it, rewinds it, or finishes it in one sitting, that’s an edge connecting those nodes. But then Netflix adds another layer: “co-engagement.” If thousands of users who watched The Bear also spent significant time with Succession, those two titles develop a strong edge between them even if they share almost no cast, crew, or plot.

Then add semantic links: similar settings, similar emotional tone, similar pacing. Directors who worked together. Writers who came from the same writers’ room. Actors who tend to pull the same audience demographic.

Now you have a graph where nodes that seem unrelated on the surface are actually three or four hops away from each other through real behavioral data. A graph neural network learns to traverse those connections and surface titles that sit in the right neighborhood for you, specifically, at this moment.

This is why Netflix can recommend a Korean thriller to someone who has only ever watched American sitcoms, if the behavioral graph says the path there is short enough and the signals are right.

The thumbnail you’re seeing is not the thumbnail your friend is seeing

This part I find genuinely strange every time I think about it.

Netflix generates multiple thumbnails for every title, and serves different ones to different users based on what the system thinks will make you click. If you’ve watched a lot of films where Uma Thurman has a significant role, you might see a Pulp Fiction thumbnail centered on her. Someone else’s algorithm decides John Travolta is the draw. Same movie. Different first impression.

For a show like Money Heist, there are roughly 86,400 possible frames in a single one-hour episode. Netflix isn’t randomly picking one. The system runs A/B tests continuously, measures which images produce clicks and completions (not just clicks, completions matter more), and rotates the winners.

Personalized thumbnails have been shown to increase click-through rates by around 30% in Netflix’s own testing. At their scale, 30% is not a small number.

What I find uncomfortable about this, honestly, is that you don’t know it’s happening. You think you’re making a neutral choice about whether a show looks interesting. You’re actually being shown a version of that show designed specifically to appeal to your particular watch history. The choice is real but the options have been pre-optimized.

The shift away from stars (and why it was smarter than it looked)

In 2016, Netflix replaced their 1-to-5 star rating system with a thumbs up/thumbs down. At the time, a lot of people called it dumbing down the product.

It wasn’t. It was a data quality fix.

Star ratings introduce noise. A 3-star rating means different things to different people. One viewer gives 3 stars to a film they’d happily rewatch; another gives 3 stars to something they considered switching off. The system can’t tell the difference.

A thumb rating is binary, which means it’s cleaner, but more importantly, Netflix weights it against behavioral signals. If you thumb-up something but only watched 40% of it, the algorithm adjusts the weight accordingly. If you thumb-down something but rewatched a specific scene three times before turning it off, that gets noticed too.

Netflix reported that the binary system improved recommendation engagement accuracy by around 200% compared to star ratings. I was skeptical of that number until I thought about how much noise they were removing from the input data.

What the system is actually tracking while you watch

It’s more granular than most people assume.

When you pause, the system notes it. When you rewind the same 30 seconds twice, it notes it. When you switch from watching on your TV to your phone at 11pm, it notes the context shift. When you start a series, watch episode one, then come back to episode two within 24 hours, that completion signal carries different weight than watching two episodes a week apart.

Netflix’s models learn that certain users are more likely to finish a series if they hit episode two quickly. For those users, the algorithm may prioritize shows where episode two has a strong hook, not episode one.

Time of day is a real factor. The system learns that you watch differently on Sunday afternoons than on Friday nights. Light comedies versus intense crime dramas. It doesn’t generalize across your whole profile, it builds context-specific sub-profiles and serves recommendations accordingly.

This is where reinforcement learning enters the picture. The system is trying to maximize not whether you click, but whether you remain a subscriber six months from now. Optimizing for immediate clicks actually produces worse long-term retention. Netflix’s research into what they call “reward innovation” focuses specifically on making recommendations that lead to satisfaction, not just engagement. Those are meaningfully different outcomes.

The part that doesn’t get talked about: the cold start problem

Every recommendation system hits the same wall when a new user signs up or a new title drops.

No behavioral data. No edges in the graph. Nothing to go on.

For new users, Netflix uses the brief preference selection at signup, plus whatever demographic and device signals are available, to make initial guesses. But those guesses are rough, and the first few weeks of a new account are frankly the weakest period for recommendations.

For new titles, Netflix uses content-based signals, genre tags, cast connections, director relationships, and metadata similarity to place the title into the right neighborhood in the graph before anyone has watched it. As viewership accumulates, the behavioral edges build out and the recommendations sharpen.

This is why a show can feel like it came out of nowhere three months after launch. The graph finally has enough data to route it to the right people.

So what does this actually mean for you as a viewer?

A few practical things.

Your recommendation profile is not static. If you watch a run of documentaries over two weeks, the algorithm shifts. If you share an account and watch different things on different profiles, those profiles are developing separately. The “continue watching” row is ordered by the system’s prediction of what you’re most likely to finish, not just recency.

The rows on your homepage are not arranged randomly. Their positioning, the titles within each row, and the thumbnails you see are all personalized outputs. Two people in the same household on different profiles can have genuinely different homepages for the same catalog.

And the shows that feel like they were made for you sometimes were, in a weird indirect way. Netflix’s greenlight decisions factor in recommendation data. If the graph says a large enough audience segment is underserved in a particular genre or tone, that informs what gets produced. The content strategy and the recommendation strategy are connected loops, not separate departments.

The honest version of what this system gets wrong

It over-fits to your recent history. Watch three thrillers in a row and the algorithm leans hard into thrillers. It can take a while to recalibrate if your taste shifts.

It has a diversity problem at the margins. The graph is very good at finding titles that sit in familiar territory. It’s worse at routing you to things that are genuinely outside your history but that you’d actually love. The exploration-versus-exploitation tradeoff in the reinforcement learning is real, and Netflix acknowledges this.

And the thumbnail personalization, while effective at driving clicks, occasionally feels manipulative in a way that’s hard to pin down. You’re being shown a curated version of a decision, not a neutral one.

None of this means the system is bad. It’s genuinely impressive engineering. But it’s worth knowing how it works, so you’re making choices with some understanding of what’s shaping your options.

That’s the actual system. Not magic, not simple “people also watched” logic. A graph that spans millions of nodes, updated in near real time, running multiple models on top of it simultaneously, wrapped in a UI designed to close the deal in 90 seconds.

The next time Netflix recommends something and you think “how did it know?” — now you know.