Listen to this article
For years, Netflix has been one of the easiest examples to point to when talking about algorithms.
You watch something, Netflix recommends what comes next. The more you watch, the better it seems to know you. Simple enough.
Except Netflix has spent years making decisions with data in places most people never notice.
Some of them shape what appears on your screen. Others happen long before you press play, or after Netflix has already made a choice and needs to know whether it was right.
So maybe the recommendation algorithm was never the most interesting part of Netflix’s data story.
The Algorithm Was Only the Beginning
Netflix’s recommendation story began with a relatively straightforward problem: how to predict what someone is likely to enjoy.
Streaming changed the scale of that problem. Netflix was no longer relying mainly on ratings; it could learn from what people actually watched and how they moved through the service. By 2012, Netflix said 75% of what its members watched came from some form of recommendation.
By then, recommendation was no longer simply a feature that helped people discover content. It was shaping a large part of what actually got watched.
And that made one prediction far less useful on its own.
By 2015, Netflix described building each homepage from thousands of potentially relevant titles and tens of thousands of possible rows, balancing relevance, diversity, position, and even the limitations of different devices. Its first personalised page-generation approach went on to perform significantly better online than the rule-based template it had previously used.
The question had expanded from What might you like? to How should an entire set of choices be built around you?
And once Netflix could decide what to put in front of you, another variable became difficult to ignore: how that choice looked when it got there.
Then Netflix Started Shaping the Choice
The right recommendation could still be easy to overlook.
Netflix found that the artwork representing a title could change what part of that title felt relevant to a viewer. Good Will Hunting, for example, could be presented through Robin Williams to someone whose behavior suggested an interest in comedy, while another viewer might see artwork emphasizing the film’s romantic side.
The movie stayed the same. The reason to notice it changed.
That pushed personalization into a more interesting territory. Netflix was no longer only trying to predict which choice fit a person. It was testing which version of that choice gave the title the strongest entry point. But that created another problem.
If Netflix could change the recommendation, its position and even the image used to sell it, how could it know which of those decisions had actually improved anything?
But Making a Decision Wasn’t Enough
That is where experimentation became part of the system.
Netflix built A/B testing into product decisions rather than assuming a more personalised experience was automatically better. Different groups could see different artwork, interfaces, or recommendation approaches, allowing Netflix to compare what actually changed in member behavior. And sometimes the difference was far from cosmetic.
Netflix reported that in some artwork experiments, changing the image associated with a title produced 20% to 30% more viewing for that title. That result did more than prove that thumbnails mattered. It showed why Netflix needed a way to test the decisions its own data was producing.
Prediction could suggest what might work. Experimentation could show whether it actually did. Data was no longer only an input into the decision. It had become the mechanism Netflix used to challenge the decision afterwards.
And once that loop proved useful inside the product, there was little reason to keep it there.
Then Data Moved Beyond the Product
One of the clearest examples came when Netflix moved into original content.
With House of Cards, Netflix did something traditional networks rarely did: it committed to two full seasons, 26 episodes, without asking for a pilot first. The deal was reportedly worth more than $100 million.
That confidence did not come from an algorithm inventing the show. Netflix had already seen overlapping audiences around political dramas, the original British House of Cards, David Fincher’s films, and Kevin Spacey. Ted Sarandos later explained that the data was used to estimate the potential audience, not to decide how the show itself should be made.
In other words, the data did not remove the creative risk. It gave Netflix a clearer idea of how much audience might already exist on the other side.
And the decision paid off. In 2013, House of Cards became the first original online series to receive major Primetime Emmy nominations, later becoming the first streaming series to win a major Emmy. That was a different use of data altogether.
Netflix was no longer using behaviour only to decide what to show people. It was beginning to use that behavior to decide which risks were worth taking in the first place.
What Netflix Really Built
What makes Netflix interesting is not that it found one algorithm smart enough to predict what millions of people want to watch.
It is that every answer creates another question. A recommendation led to questions about placement. Placement led to presentation. The presentation had to be tested. Testing created new evidence. Eventually, that same way of thinking moved beyond the screen and into much bigger decisions about content and risk.
That is what makes Netflix’s data-driven approach harder to copy than a recommendation engine. The value is not in one model, one dataset, or even one successful prediction. It is in the loop Netflix built around them: observe behavior, make a decision, test what happens, and use the result to make the next decision with a little less uncertainty.
Netflix became known for predicting what you might watch next. But its bigger achievement may have been learning how to question what it thought it already knew.
