One of the surprise hits of 2025 was the animated film K-Pop Demon Hunters. Though it initially received only a small theatrical release in June, buzz around the movie grew so much while it was streaming on Netflix that a wider theatrical run in August topped the U.S. box office. Netflix later declared it the most popular film to ever appear on its platform, the soundtrack album went platinum, and the characters became the year’s hottest Halloween costumes.
The movie’s success was another win for the Netflix strategy of using its recommendation system to bring a broad range of content to potentially interested viewers. In the best-case scenario, movies and shows like K-Pop Demon Hunters, Squid Game, or Stranger Things become must-see entertainment, driving a wave of new subscriptions.
But it’s hard to say what made these particular movies and shows click with viewers. Was it the allure of the content itself? Or was it the boost provided by Netflix’s personalization algorithm?
It’s a tricky question—one that Guy Aridor, an assistant professor of marketing at Kellogg, has been working to address with Netflix coauthors Kevin Zielnicki, Aurelien Bibaut, Allen Tran, Winston Chou, and Nathan Kallus. In a working paper, they present a model that disentangles the influence of the platform’s recommendation service from the underlying value of the content itself.
In addition to helping Netflix determine how many additional viewers different shows and movies attract, the model also offers a data-driven perspective on the types of content that recommendation systems help the most.
“What we find is that the middle-tier titles that are moderately popular, that have very strong niches—these are the ones that are actually really benefiting from the recommendations,” Aridor says.
The power of recommendation
Netflix regularly boasts of having the largest streaming catalog in the industry. But that volume isn’t as valuable if subscribers don’t find what they want to watch.
“You can’t produce and acquire a lot of titles and make them stick unless you can target them effectively,” Aridor says. “A lot of titles, if they’re not recommended to the right people, they likely won’t get much viewership.”
So, Netflix invests heavily in their recommendation system. In 2006, they offered a $1 million reward to the first team that could improve their recommendations by 10 percent. A group of computer scientists won it in 2009 by adapting a mathematical approach called matrix factorization.
Netflix didn’t stop there. Using data from their over 300 million subscribers, the company has continuously refined its system so that every user sees a personalized menu of content every time they open the app.
But this success complicates other parts of the Netflix business. It’s hard to measure how much viewership is driven by that targeting relative to the quality of the content itself. Are people just watching a show because Netflix suggested it, or would they seek it out regardless of the algorithm?
The subscription-based model further muddies the water. Competitors like Amazon and Apple can measure the value of a given film based on rentals or sales at different prices. But Netflix customers pay a flat fee to access the entire catalog, so the company can’t easily judge the replacement value of each title.
Measuring value in the attention economy
Aridor and his coauthors tackled this problem by building a new model for content valuation. The model enables them to simulate hypothetical scenarios that allow Netflix to ask the questions it cares about.
For example, if the current recommendation system was replaced with a random selection of content or simply the most popular shows and movies, how would user engagement change? If a viewer who would probably like Emily in Paris couldn’t find it, what would they choose to watch instead? Would they just turn off the app and do something else?
The researchers were able to build this model because of intentional amounts of randomization in the Netflix algorithm. In order to better learn user preferences, the company routinely runs subtle experiments that randomize different elements of the Netflix homepage. The researchers used what viewers choose to do under these different conditions to train and validate their model.
“This is a general problem in the so-called attention economy,” Aridor says. “There’s limited or no price variation, and so by inducing variation in the set of shown titles, that teaches us something about how you substitute between products.”
Surfacing the middle tier
Once built, the researchers could use the model to measure incremental viewership—how much each individual show or movie drives engagement on Netflix. They could even assess the value of the recommendation system itself, as compared with other algorithms.
Unsurprisingly, the current Netflix recommender beat alternatives such as random suggestions, showing only the most popular content, and a matrix factorization system similar to the winner of the Netflix Prize in 2009.
But the current system also performed best on another measure that Netflix values: increasing content diversity, or the overall variety of shows and movies that users watch.
“Research from other streaming platforms shows that more diverse consumption is strongly correlated with good long-run outcomes from a consumer-satisfaction point of view,” Aridor says. “So it’s important that the recommendation system isn’t just inducing people to all watch the same types of titles.”
The model also revealed that proven hits like Emily in Paris and Stranger Things don’t need much additional promotion, and that obscure shows and movies don’t connect outside of very specific audiences. Instead, it’s the shows and movies in-between that get a lift from the recommendation system.
“This is really where the bread and butter of the recommender system and the Netflix business model go hand in hand, which is that you can only really have this kind of deep catalog if I am able to efficiently match these viewers with those titles,” Aridor says.
Cultural effects
Beyond its utility to Netflix, the findings contribute to an extended debate about media and recommendation systems: Do they cause audiences to cluster around a smaller segment of entertainment, or do they surface art that would otherwise be neglected?
“Maybe older recommender systems like the ones from 10 years ago that were deployed on these types of platforms did have this clustering problem,” Aridor says. “But in terms of diversity, today’s system performs much better.”
And while the researchers’ new model may help Netflix make decisions about which content to add, it’s unlikely to capture the organic, off-platform factors that elevate a solid performer into the next cultural juggernaut.
“When a midsize title gets recommended to the right set of people, then, sometimes, that audience is sizable enough to talk about it online and have conversations with each other, and then it snowballs,” Aridor says. “But it’s very rare, and it’s very hard to predict.”