Watch what shows up on your home screen when you open Netflix tonight. Very little of it is editorially curated in any meaningful human sense, including the shows in the top row, the arrangement of the thumbnails, and the particular image used for something you nearly watched last week. You are viewing a prediction about yourself that was created in the seconds that passed between opening the application and the site loading. Millions of users who pause at the same times, skip the same scenes, and stop watching the same kinds of episodes at around the same point in episode two were used by the system to build those predictions based on everything you’ve done on the platform.
The only truly honest explanation of how the system functions is seen in Netflix’s engineering publications, where the company has been openly forthright about the scope of this. Instead of active search, the platform’s recommendation engine accounts for more than 80% of the content that users watch. Once you sit with that figure, it’s stunning. The great majority of the content that Netflix’s 270 million users see on average each week was, in some way, chosen for them by an algorithm prior to their opening the app. Those who understood exactly what they were looking for and conducted direct searches make up the remaining fraction.
There are multiple levels operating concurrently in the technical architecture. You are paired with what Netflix refers to as “taste twins”—users whose behavioral patterns closely reflect yours—through collaborative filtering. The way they engage with content—where they pause, what they replay, which parts they skip, and how far into a series they stop before stopping—rather than their declared preferences, demographics, or even ratings. Your list of suggestions paints a sort of general picture of what people who act similarly to you ultimately watched and completed. The reason for your preferences is unknown to the model. It applies its understanding of what individuals who engage in similar activities often watch next on a large scale.
The aspect of the system that most people are unaware of and would likely find most disconcerting if they did is the personalization of the artwork. For a large percentage of its repertoire, Netflix creates a variety of thumbnail images, such as distinct visual cropping, different actors foregrounded, and different moments frozen from the episode. Your data determines which version of the thumbnail shows up on your screen. The thumbnail for an unrelated show that an actor appears in will display the actor’s face instead of a landscape or a scene that is pertinent to the plot if you frequently click on content that features that actor. It has the same content. The wrapper that surrounds it was created just for you and is optimized to increase the likelihood that you will click.
The algorithm’s reach beyond individual episodes began in 2012 with the introduction of autoplay. It was determined that the 15-second countdown before the next episode begins—the fleeting moment when you would otherwise choose to pause, check your phone, or go to bed—was a high-dropout moment and was removed from the typical viewing experience. Instead than deliberately choosing to keep going, you must actively choose to stop. The reasoning behind every design choice Netflix has made regarding how the interface presents its content is the same: eliminate the points where viewers pause to assess, lessen the difficulty of continuing, and create the path of least resistance toward the next piece of content. This is a subtle but significant inversion of the default.
It’s important to keep in mind the competitive background of all of this. In 2006, Netflix offered $1 million to any research team that could increase its suggestion accuracy by 10%. This was known as the Netflix Prize, and it attracted thousands of competing teams for three years before a coalition known as BellKor’s Pragmatic Chaos emerged victorious. The award propelled the field of collaborative filtering, spawned a substantial amount of scholarly research, and made it abundantly evident that Netflix valued their recommendation system enough to invest $1 million to improve it by a tenth. Since then, the reasoning behind that investment has only grown. With an annual content budget of almost $17 billion, the recommendation system is currently the company’s top engineering priority. In essence, what the algorithm presents to each subscriber is a distribution decision on the portion of that investment that they will ever come across. For the most part, most people are unable to see the majority of the library.
