An exclusive excerpt from Every Screen On The Planet reveals how the social media app’s powerful recommendation engine was shaped by a bunch of ordinary, twentysomething curators—including a guy named Jorge.
This week, President Donald Trump announced that he and Chinese leader Xi Jinping had agreed to terms that would allow the Chinese tech giant, ByteDance, to sell TikTok’s U.S. operations to a group of American investors. One vital asset that won’t be fully transferred in the deal—the platform’s powerful For You algorithm.
When you first open TikTok, the app doesn’t know that much about you yet. It knows your approximate location, your language preferences, and whether you’ve showed up in any of its other users’ contact lists—but your interests are still largely a mystery.
So it shows you a preselected set of popular videos, and watches you react, recording how long each video holds your attention, and whether you like, comment, share, or text the video to a friend. Some people carefully perform their initial interactions, knowing the machine is watching. Through their swipes, likes, and texts, they try to express their preferences to it, to create the feed they want, or the feed they think they should want. They do this especially at the beginning, because they know that the algorithm has only a handful of data points, and that it will rely on them more heavily than when it has many thousands down the road.
In late 2018, ByteDance hired a small team of content curators in Mexico City to introduce TikTok to the Latin world. Among them was a twentysomething named Jorge Reyes, who would select cooking tutorials, dance routines, soccer highlights, and other short clips to promote to the app’s Spanish-speaking users. Because he chose many of the first videos featured on Latin American TikTok, Jorge’s tastes and instincts shaped Tik Tok’s For You algorithm, eventually determining what hundreds of millions of future users would see.
Jorge was one of four initial Mexican content operators that ByteDance hired to curate the TikTok feed for Latin America and Spain. Like other ByteDance cohorts around the world, the Mexico City group was based out of a WeWork office and reported to a team in China. Their job was in large part to be “cultural translators”: young, educated, “cool” Spanish speakers who could act as teachers to the team back in China, helping them learn which videos would resonate with Latin urban youth.
Jorge and his cohort also had another student, though, one even more important than their colleagues in Beijing. Initially, the For You algorithm was bad at recommending videos outside of China. It would push posts that were, for example, only two seconds long, or so blurry you couldn’t make out what was happening. To fix this problem, ByteDance relied on local teams like Jorge’s—every time they removed a bad video or boosted a good one, they gave the For You algorithm another data point to learn from about what its next recommendation should be.
The bluntest instrument Jorge could use to boost a video was a lever known as “heating,” an override of the normal recommendations system that ensured a video would receive a certain number of views. Employees could choose how many views they wanted the video to accrue—5,000; 50,000; 100,000; 500,000; 1,000,000; or even 5,000,000. Once they made their selection, the video would immediately be shown to users until it hit its mark. Some of those people would engage with the post, sharing it out to their followers, some portion of whom would share it again, catapulting lucky creators to what could feel like instant virality.
Heating was an open secret within ByteDance, but one the company really didn’t want its users to know about. If people knew that TikTok staffers were simply picking winners to blast out on the For You page, then the idea of an unbiased, tastemaking platform would fall flat. It was much better if aspiring creators believed in the opaque, meritocratic magic of the algorithm. TikTok’s algorithm, however, wasn’t actually meritocratic or magical.
Algorithms are equations—big math problems whose variables are a pile of preferences, incentives, and weights—but they’re written by humans. They encode the biases, both overt and implicit, of their creators. A former FTC commissioner named Maureen Ohlhausen once suggested that people thinking about algorithms should replace the word “algorithm” with the words “a guy named Bob.”
“Is it ok for a guy named Bob to collect confidential price strategy information from all the participants in a market, and then tell everybody how they should price?” she asked in a 2017 speech. “If it isn’t ok for a guy named Bob to do it, then it probably isn’t ok for an algorithm to do it either.”
Ohlhausen was right: algorithms were just tools programmed by their human creators, and the decisions they made about how to price products or target people with news deserved the same scrutiny as other human judgment.
In TikTok’s case, at least for Latin American users, the algorithm actually was, in part, a guy named Jorge, both in the literal cases where Jorge heated videos, but also in a more systemic way. ByteDance’s For You algorithm, like all algorithms, was just a big jumble of preferences. And the preferences of those early curators, each expressed as an override to the existing system, had helped train and retrain the For You page. Each heated video, each deleted one, brought the algorithm ever closer to curators’ own subjective judgments about which videos the platform should reward.
One of Jorge’s other big tasks as an early curator was broadening TikTok’s appeal. TikTok grew out of another Chinese app called Musical.ly, which ByteDance acquired in 2017. When the sale closed, ByteDance had claimed that it would become “the world’s largest short video social entertainment platform.” The company’s goal for TikTok looked less like Musical.ly and more like YouTube: one where everyone from bored investment bankers to gardening grandmas would tune in to learn how to tie a fisherman’s knot, or care for a tulip bulb, or play a song on the guitar.
When Jorge joined TikTok in 2018, it was nowhere close to being that platform. Videos of teens lip-synching and dancing were so pervasive on the app that if a person did literally anything else in their video, it would be labeled as “diverse.” Jorge told me that “Only 12 percent of videos were initially labeled as ‘diverse.’” Jorge was charged with reversing these numbers, by coaxing new types of creators into making videos, and making sure that when they did, those videos did well.
Heating was one tool Jorge could use to reward “diverse” videos, but there were others, too. Curators could promote hashtags related to current events, movies, songs, and other pop culture that would encourage users to post about them. (Jorge recalled promoting the Luis Fonsi song “Despacito.”)
Curators could also arrange the order of posts on hashtag-specific pages. They could add an “official” label that would show users that TikTok, the company, had endorsed certain videos. And then there was the Discover page—a tab that would show users a seemingly random selection of videos to help them explore new content. Jorge and his colleagues could choose and rank the videos on that page, too.
In deciding what types of videos to promote, Jorge’s team also received help from data science teams back in China, which sent them weekly reports about the topics that were performing best in each market, whether it was soccer in Colombia, or cooking videos in Spain. There were certain topics they were forbidden from pushing: politics, religion, and anything that might be considered “vulgar” for younger users. But aside from these restrictions, Jorge’s team was told to go forth and try things, aiming to maximize diversity of topics and user satisfaction.
The manual curation tools that Jorge and his colleagues used were a key part of TikTok’s early rise across the world, but the people who made and watched videos on TikTok didn’t know heating was a thing. Heating served multiple purposes for the company, but among the most important of them was the ability to woo celebrities, brands, and creators from other platforms. If Jorge could convince a YouTuber to give TikTok a try, he could immediately heat the person’s first posts and give them a first-tier slot placement on a hashtag page—the instant engagement would convince the YouTuber that the app was worth their time. Jorge wouldn’t reveal, of course, that the creator’s instant success had been manufactured.
At the beginning, Jorge and other TikTok employees turned regular TikTok users into social media stars, and they loved doing it. But as the field became more crowded, the chances of becoming a breakout star necessarily decreased, and before long, the days of picking individual winners were over. ByteDance began making deals with marketing agencies, promising them payment in cash if they could deliver a consistent stream of new users. But even that couldn’t scale quickly enough. So they adopted an idea from the competition.
One of TikTok’s key competitors during this period was an app called Kwai, run by the Chinese tech giant, Kuaishou. To spur downloads, Kwai had offered a small cash payment if a user referred a friend. Especially for users in less-wealthy communities, the payments could be significant, and Kwai’s downloads surged.
In 2019, ByteDance launched TikTok Rewards, also known as TikTok Bonus: a program where people would receive points, redeemable for cash, for referring new users to the platform. The downloads that TikTok gained through this program were, to put it gently, not always authentic. Sure, there were the people who received a referral from a friend, watched some funny videos, and referred the app on to others. But there were also scaled operations—such as agencies and entrepreneurial click-farmers, which set up systems to generate referral cash at scale. At TikTok, employees anticipated abuse within the program. Still, even if the referred accounts weren’t real people, they looked good in both internal and external metrics. Inflated numbers helped the platform look stronger in the battle for app store ratings and hype.
The TikTok Bonus program was only the beginning. Over the course of its first year, ByteDance invested enormous sums to drive downloads, spending almost $1 billion to market TikTok and its other apps. To most U.S. users back then, TikTok was an unfamiliar upstart, but it came with the budget of a tech giant, spending more on promotion alone than the annual budget of most major metropolitan police forces, nearly ten times the average value of an IPO.
The blitz amounted to an ad spend of nearly three million dollars per day. The lion’s share of the spending went to platforms that would become TikTok’s direct competitors: Facebook, Instagram, YouTube, and Snapchat.
The ads featured clickbaity videos that led users directly to a download page. They took user- generated posts that had performed well on TikTok and juxtaposed them with text and links instructing users to download the app. Some featured young teens—and even celebrities—who had posted videos not realizing they might later be used as ads.
The ads reflected the same strategy that had led ByteDance founder Zhang Yiming to buy Musical.ly in the first place. He believed in the strength of ByteDance’s technology but knew it would be unusually hard to convince people to download a strange app they’d never heard of, especially once they learned it was Chinese. So he backed up a truck full of money, “acquiring” new users with advertising to build on the substantial user bases he had bought with the purchase of Flipagram and Musical.ly.
Driving downloads was a necessary part of the puzzle, but it wasn’t the same thing as getting new users to stick around. The ads gave TikTok name recognition and an early brand identity. But converting someone who just downloaded the app into a regular user was another enormous task— and it’s one that fell largely to regional curators like Jorge Reyes.
ByteDance’s challenge—like Facebook’s, and Google’s, and Twitter’s—was to achieve universality. There was nothing inherently Chinese about personalized news or video recommendations, the same way there was nothing inherently Californian about search engines or microblogging. A global TikTok could give influencers a worldwide audience, potentially multiplying their popularity by the company’s many markets around the world.
But teams like Jorge’s quickly learned that virality wasn’t universal. The company tried to prioritize content that was “not too Chinese,” but they soon found that humor, memes, and internet culture were unusually tricky to translate.
One way ByteDance tried to engage with foreign users was by showing them content familiar to them. Back in 2017, Yiming had approached Jonah Peretti, the founder of BuzzFeed, about licensing the American company’s catalogue of entertaining viral videos. “I asked what kind of content and he said it didn’t matter, he just needed tens of thousands of videos each day,” Peretti later wrote. “He just needed raw tonnage of content so the AI could create a personalized experience and get the flywheel going.”
By 2019, TikTok was among the most downloaded apps in the world. Yiming’s billion-dollar ad bet had paid off, and an influx of new users gave the company a flywheel of new data about US and Mexican and Brazilian culture. A stream of videos, posts, and comments from new users fortified the For You algorithm, giving it the predictive strength across the Americas that it had previously lacked.
TikTok—like Musical.ly—also began to emerge as its own, distinct experience. Its main competitors— Facebook, Instagram, YouTube, and Snap— were apps where you could choose your own adventure: they were driven by friend requests, search queries, likes, comments, shares, and other expressed preferences. They used users’ vacation photos and fundraising birthday posts to infer other things about behavior, but for the most part, they still kept the user in the driver’s seat.
TikTok was different. Users would open the app and the show would begin, placing the viewing experience on autopilot. Sure, you could like and comment, and sometimes you did, but you didn’t have to—because TikTok ran on your revealed preferences, rather than your expressed ones: it knew that you always lingered on videos about bisexuality, or alcoholism, or divorce, even if you never liked or shared or commented on them.
Connie Chan, a partner at the renowned investment firm Andreessen Horowitz, described TikTok as “the first mainstream consumer app where artificial intelligence is the product.” The app, she said, “never presents a list of recommendations to the user (like Netflix and YouTube do), and never asks the user to explicitly express intent—the platform infers and decides entirely what the user should watch.”
Chan suggested that TikTok could use this editorial power to “optimize the video feed for happiness,” apparently without concern that it might choose to instead optimize the feed for other things—like the desire to spend more time on TikTok, or more money on the products advertised there. “In fact,” she said, “the entire vibe of the platform is largely under TikTok’s control, because they, not users, decide which videos to display.”
Adapted from Every Screen on the Planet: The War Over TikTok by Emily Baker-White. Published by arrangement with W.W. Norton. Copyright © 2025 Emily Baker-White.