How Netflix’s Choice Engine Drives Its Business

It’s Friday night after a long week, and you’re definitely going to relax and watch a movie. So you turn to Netflix, the world’s largest streaming service. It is also the prototypical choice engine: its goal is to help you find something to watch. It does not just passively present options, it tries to customize the set of things that you see, it gives you some control over what is presented, and it even helps you comprehend new options you might like.

In fact, Netflix’s entire existence depends on finding programs that you will want to watch from its large library of licensed content. Keeping you engaged and paying a monthly subscription is its major source of revenue. The Netflix stock price starkly reflects changes in the number of customers: announcing that it had lost 130,000 customers in 2019 (less than 0.2 percent of its total of 60 million) caused a 10 percent drop in the stock price in one day. Netflix needs to connect customers with content, so you think it will be able to find you something to watch.

To do this, Netflix’s landing page uses many choice‑architecture tools. Let’s look at some of them.

Plausible Paths: Netflix knows that decisions about how to make decisions, choosing a plausible path, happen quickly. Netflix does not keep the chooser waiting. It spends enormous effort preventing waiting, making the decision of what to watch very fluid. Netflix engineers brag about the technological innovations used to quickly load the landing page for customers, and how important preventing waiting is for keeping customers.

Defaults: By default, when you go to Netflix, a trailer, complete with audio, starts playing. Until early 2020, there was no way to turn it off, despite many complaints on social media.

Number of Options: Netflix has a catalog of almost 6,000 titles (about 4,000 movies and 2,000 TV series). They somehow reduce that flood of options to the 80 or so that they believe might be right for you to watch. This involves some serious AI magic, as we will see in a moment.

Ordering: Netflix sets the order of the rows. Does “Trending Now” come first, or “New Releases”? Once the order of rows is decided, Netflix needs to assign each program a position, at the beginning, middle, or end of the row. If you compare your landing page to anyone else’s, you will see both orders are specific to each user. Films on the top left of the initial screen are watched more often than those elsewhere.

Describing Options: Each of the rows has a heading, like “Critically Acclaimed TV,” “Trending Now,” or “Witty TV Shows.” Not only do the order of these differ for every customer but different options appear for different people. Each program has a still image. Does Netflix pick pictures guaranteed to draw attention to all content, or is it more targeted, attempting to increase the popularity of some content?

The goal of Netflix’s choice architecture is to find things that will make you happy inexpensively.

Describing options also includes selecting which scenes are in the trailer that runs when you hover over the picture. How does Netflix decide which moments to show from the 62 episodes of Breaking Bad? For the introduction to the American version of House of Cards, it developed three different trailers for different audiences, based on what it knew of their past viewing. One was for fans of the British version of House of Cards. Another featuring Robin Wright (Claire Underwood) and other female characters ran for viewers who had watched Thelma and Louise. A third was aimed at serious film buffs, because the producer, David Fincher, is well known in those circles for movies like The Social Network and The Girl with the Dragon Tattoo.

Netflix presents an attribute predicting how much you will like each title, using a 0‑to‑100‑percent scale. It also collects your ratings on a “thumbs‑up, thumbs‑down” scale. Why did Netflix make the two scales different? Netflix had used a 5‑star scale for both, but it believed that people were confused by the scale. People found that the thumbs scale was easier to use, and changing to the thumbs doubled the number of ratings collected. It also found that people tended to only rate highbrow and serious movies with a 5, but they were happy to give a binge‑watched situation comedy a thumbs‑up. The response scale made raters, it seems, less pretentious and perhaps more honest.

This entire set of tools is tuned through A/B tests, as many as a hundred a year, on every detail of the experience. With millions of viewers using the landing page each day, a lot can be learned.

So how does Netflix do it? As I developed one of the first university courses on choice architecture, I tried hard to listen to my students. They live more of their life online than I do. At the end of the class, I always ask them about their most and least favorite choice architectures. Most years, the winner of both the best and worst award is Netflix. When I ask why, I hear the following:

“I love Netflix because it finds me something to watch.”

“I despise Netflix because I can’t find what I want to watch.”

Listening to the conversation carefully reveals a disconnect between the goals of some choosers and those of Netflix. Netflix does not try to optimize customer satisfaction or, as they put it, “happiness.” Netflix tries to maximize efficiency—that is, “maximum happiness per dollar spent” on content, as a vice president of product engineering there once put it.

Some shows, such as The Crown, deliver happiness big‑time but are expensive to make, costing over $10 million an episode. A menu full of such programs might make people happy, but the subscription would be prohibitively expensive, and Netflix would disappear.

If you think of Netflix as a video Library of Congress, containing all the programs ever produced, you will be disappointed. If, instead, you are looking for an outlet that will entertain you, and provide that entertainment easily and efficiently, you have met your ideal video service.

Instead, the goal of its choice architecture is to find things that will make you happy inexpensively. According to Jenny McCabe, director of global media relations at Netflix, “We look for those titles that deliver the biggest viewership relative to their licensing cost.”

Whether Netflix shares your goals depends, as my students’ comments show, on what you are looking for. If you think of Netflix as a video Library of Congress, containing all the programs ever produced, you will be disappointed. If, instead, you are looking for an outlet that will entertain you, and provide that entertainment easily and efficiently, you have met your ideal video service.

But to do that, Netflix has to know you. Let’s look at how sites like Netflix do that.

One of the things that can be done in an interactive choice engine is customizing the choice architecture. This can lead to happier customers and more productive firms. Netflix was producing more than 33 million versions of its site as early as 2013. To do that, Netflix has to know something useful about its customers. Some of that knowledge comes from Netflix’s recommendation system. Some estimate that it adds $1 billion of value to the firm. We’ll get to the topic of recommender systems, but first I want to talk about a broader and sometimes simpler concept: a user model.

Whenever we customize a site to increase its usefulness to a chooser, it is because we believe we know something about that person. That knowledge, that picture of a person, drives the customization. While user models can sometimes be complex analytic systems, they can also be quite simple. Keep in mind the very first thing that Netflix asks when you log in—Who is watching?—alongside three buttons, typically your name, your partner’s name, and “kids.” Netflix ask this up front because the customization is different for each user.

More sophisticated methods exist, of course. A potentially more powerful method is collaborative filtering. It gathers data about what users have bought in the past and uses AI to predict what people are likely to buy in the future. These methods can use both explicit information, like a customer’s ratings of the options, and implicit information, like whether or not they finished a specific program on Netflix. Most famously, perhaps, collaborative filtering is used by Amazon in generating “People who bought this also bought . . .” listings. Collaborative filtering requires a large set of past user behavior to make predictions. This is the heart of suggestions made by Apple Music, “who to follow” suggestions on Twitter, and matches on Tinder. Yes, Tinder apparently changes the people it will show you based on your swipes. Swiping right will change who you see in the future.

It’s important to realize that, in its pure form, collaborative filtering doesn’t use in‑depth information about the options themselves. When Apple Music recommends a tune, it knows nothing about the song’s tempo, beat, lyrics, or instrumentation. It simply knows that people who are like you like that song too.

Instead of AI for making choices, maybe we need to think about IA, intelligent augmentation, where choice architecture assists choices.

Compare that to content-based filters, which require knowing the attributes of the options. Sometimes that is easy—for example, if we were on a website that sells men’s shirts. The description of a particular Oxford shirt contains a lot of information: its color, its material, the kind of collar it has, if it’s no‑iron, and so on. This is the same data that customers see when they make choices, and we can use it to predict choice. For other products, like music, knowing the product’s attributes is really challenging.

Companies that use content‑based filters ask users to rate options on dimensions, like the sportiness of a shirt or the tempo of a song. They then write algorithms to decompose the song from its digitized representation (for example, the mp3 file) into its attributes. Content‑based filtering is used by the online streaming service Pandora via its so‑called Music Genome Project. A trained musicologist spends twenty to thirty minutes listening to each song, rating it on hundreds of dimensions, or “genes,” as Pandora calls them. An algorithm uses those ratings to select similar songs. Here, unlike collaborative filtering, the algorithm knows much more about the song and less about the person. Pandora was purchased by SiriusXM in 2018 for $3.5 billion, and the technology is now used to select songs for some of SiriusXM’s stations.

As time went on, collaborative filtering and content‑based algorithms were used together in ensembles. Since they have complementary strengths and weakness, this makes sense. It is important to note, however, that a user model is not synonymous with fancy AI. If we want to know something about the customer, we can often do important customization by asking a quick question like, “What kind of a car are you shopping for?” or “How old are you?”.

Most of the buzz around recommender systems usually emphasizes replacing choice. Another view is that user models allow the designer to augment the choice architecture. Instead of AI for making choices, maybe we need to think about IA, intelligent augmentation, where choice architecture assists choices.

From The Elements of Choice: Why The Way We Decide Matters by Eric Johnson published on October 12, 2021, by Riverhead, an imprint of Penguin Publishing Group, a division of Penguin Random House LLC. Copyright © 2021 Eric Johnson.