A version of this article originally appeared in Quartz’s members-only Weekend Brief newsletter. Quartz members get access to exclusive newsletters and more. Sign up here.
Netflix $NFLX gives itself 90 seconds. That's how long research shows the average subscriber will browse before losing interest and drifting to a competitor. In that window, the company's recommendation engine must surface something compelling from a catalog of thousands. Get it right, and the subscriber stays. Get it wrong often enough, and they cancel.
Back in 2016, when Netflix had about 80 million subscribers, company executives valued this algorithmic matchmaking at $1 billion per year in retained customers. A decade later, the streaming giant now has 325 million subscribers worldwide. While Netflix hasn't updated that figure publicly, the math suggests its recommendation system has become one of the most valuable pieces of software in entertainment.
Now, as Netflix pursues an $83 billion acquisition of Warner Bros. Discovery, a century-old studio that helped invent Hollywood, the algorithmic approach that built the streamer's dominance is poised to absorb the old guard entirely.
Netflix learned to watch you back
The company's early recommendation system relied on star ratings — what users said about movies after watching them. But in 2017, Netflix abandoned that approach for something more revealing: behavioral data.
What you actually click on. How long you watch before abandoning a title. What time of day you're viewing, and on which device. What you scroll past without selecting. This implicit feedback proved far more valuable than explicit preferences. People's stated tastes, it turns out, are unreliable narrators.
Today, Netflix logs hundreds of billions of these micro-interactions annually, feeding them into a system of interlocking algorithms that personalizes nearly every element of the viewing experience. The same movie might appear with different thumbnail images for different users, emphasizing romance for one viewer and action for another.
Even the order of rows on your homepage is calculated specifically for you. Behind the scenes, teams of "taggers" watch every title and assign granular attributes — whether a show features an ensemble cast, is set in space, or stars a strong female lead — that machine learning systems use to sort viewers into thousands of "taste communities."
The efficiency of this approach created a new category of entertainment that critics have dubbed the "algorithm movie" — films designed to appeal to the broadest possible audience by combining familiar, data-validated elements.
Think of Netflix's $320 million sci-fi flop The Electric State, which critics described as a mashup of Spielbergian childhood quests, Mad Max wastelands, and retro-futuristic aesthetics. Or the Ryan Reynolds vehicles that reliably surface on autoplay, their titles often telegraphing exactly what's inside: Tall Girl, Murder Mystery, Red Notice.
These productions typically have what one industry source described as "easy-to-follow story beats that leave no viewer behind." Screenwriters working with Netflix have reportedly received notes asking them to have characters announce what they're doing, so that viewers watching while scrolling on their phones can follow along.
The sound mixes are flat because they need to work across environments, from VR headsets to cracked phone screens. The lighting stays bright but low-contrast, engineered not to jar anyone out of a Netflix-and-chill stupor.
Netflix executives deny reverse-engineering content from data, with co-CEO Ted Sarandos claiming commissioning is "70% gut and 30% data."
But the company's influence extends beyond its own productions. Its global distribution model, which demands worldwide rights rather than territory-by-territory licensing, has restructured how independent films get financed. The old system of pre-selling distribution rights in individual markets has largely collapsed.
What's left is a system where the movies most likely to get made are the ones most likely to get recommended.
The AI acceleration is already here
That system keeps evolving. Now Netflix is layering generative AI onto its algorithmic foundation. The company already uses machine learning to select which frames from a show might work best as promotional images, to generate personalized artwork, and to assist with visual effects.
Netflix frames these tools as enablers for human storytellers, not replacements. But if the Warner Bros. acquisition succeeds, the company won't just be shaping new stories — it will control a library of old ones made long before algorithms had any say.
That includes Casablanca, a film famously rewritten on set, its ending unfinished until days before filming. That kind of creative chaos is hard to imagine surviving a system designed to minimize risk and maximize completion rates.
But the algorithm will still have 90 seconds to convince you to press play.