Jun 04, 2025·8 min
Zhang Yiming & ByteDance: Building a Global Attention Engine
How Zhang Yiming and ByteDance combined recommendation algorithms and content logistics to scale TikTok/Douyin into a global attention engine.
How Zhang Yiming and ByteDance combined recommendation algorithms and content logistics to scale TikTok/Douyin into a global attention engine.
Zhang Yiming (born 1983) is best known as the founder of ByteDance, but his story is less about celebrity entrepreneurship and more about a specific product belief.
After studying at Nankai University (moving from microelectronics toward software), he took roles that exposed him to search, feeds, and consumer internet scale: building at travel search startup Kuxun, a short stint at Microsoft China, and then founding an early real-estate product, 99fang.
Zhang’s core question was simple: how do you match the right information to the right person quickly, without asking them to do a lot of work?
Earlier internet products assumed users would search or follow portals and categories. But as content exploded, the bottleneck shifted from “not enough information” to “too much information.” His product thesis was that software should do more of the filtering—and do it continuously—so the experience improves with every interaction.
From the start, ByteDance treated personalization as a first-class product primitive, not a feature you add later. That mindset shows up in three recurring choices:
This is a breakdown of mechanisms, not mythology: how recommendation algorithms, product design, and “content logistics” work together—and what that means for creators, advertisers, and safety at global scale.
ByteDance didn’t start with short video. It started with a simpler question: how do you help people find useful, interesting information when there’s too much of it?
Zhang Yiming’s early products were news and information apps designed to learn what each user cared about and reorder the feed accordingly.
The breakout early product was Toutiao (a “headlines” app). Instead of asking users to follow publishers or friends, it treated content like inventory and the feed like a personalized storefront.
That framing mattered because it forced the company to build the core machinery early: tagging content, ranking it, and measuring satisfaction in real time.
Most consumer apps at the time leaned on a social graph—who you know determines what you see. ByteDance bet on an interest graph—what you watch, skip, read, share, and search determines what you see next.
That choice made the product less dependent on network effects at launch and more dependent on getting recommendations “good enough” quickly.
From the beginning, ByteDance treated product decisions as hypotheses. Features, layouts, and ranking tweaks were tested continuously, and winning variants shipped fast.
This wasn’t just A/B testing as a tool; it was a management system that rewarded learning speed.
Once the recommendation engine worked for articles, moving into richer formats was a natural next step. Video offered clearer feedback signals (watch time, replays, completion), faster content consumption, and a bigger upside if the feed could stay consistently relevant—setting the stage for Douyin and, later, TikTok.
For most of media history, the problem was scarcity: there weren’t enough channels, publishers, or creators to fill every niche. Distribution was simple—turn on the TV, read the paper, visit a few websites—and the “best” content was whatever made it through limited gates.
Now the bottleneck has flipped. There’s more content than any person can evaluate, even in a single category. That means “too much content” is less a creation problem and more a distribution problem: the value shifts from producing more posts to helping the right viewer find the right thing quickly.
Chronological feeds assume you already know who to follow. They’re great for keeping up with friends or a small set of creators, but they struggle when:
Follower-based discovery also favors incumbents. Once a few accounts capture attention early, growth becomes harder for everyone else—regardless of quality.
When content is abundant, platforms need signals that separate “seen” from “enjoyed.” Time spent matters, but it’s not the only clue. Completion rate, rewatches, pauses, shares, and “not interested” actions help distinguish curiosity from satisfaction.
In a broadcast model, scaling means pushing one hit to millions. In a personalized model, scaling means delivering millions of different “small hits” to the right micro-audiences.
The challenge isn’t reach—it’s relevance at speed, repeatedly, for every person.
ByteDance’s feeds (Douyin/TikTok) feel magical because they learn quickly. But the core idea is straightforward: the system repeatedly makes a guess about what you’ll enjoy, watches what you do next, and updates the next guess.
Think of the feed as a shop with millions of items.
Candidate generation is the “shortlist” step. From the huge catalog, the system pulls a few hundred or thousand videos that might fit you. It uses broad clues: your language, location, device, accounts you follow, topics you’ve engaged with, and what similar viewers liked.
Ranking is the “final ordering” step. From that shortlist, it predicts which videos you’re most likely to watch and enjoy right now, and sorts them accordingly. Small differences matter here: switching two videos can change what you watch next, which changes what the system learns.
The algorithm doesn’t read minds—it reads behavior. Common signals include:
Importantly, it also learns “negative” preferences: what you consistently skip, mute, or mark as not interested.
For a new user, the system starts with safe, diverse picks—popular content in your region and language, plus a mix of categories—to quickly detect preferences.
For a new video, it often runs a controlled “trial”: show it to small groups likely to be interested, then expand distribution if engagement is strong. This is how unknown creators can break through without an existing audience.
Short videos produce lots of feedback in minutes: many views, many swipes, many completions. That dense stream of signals helps the model update rapidly, tightening the loop between “test” and “learn.”
ByteDance can run A/B tests where different groups see slightly different ranking rules (for example, weighting shares more than likes). If one version improves meaningful outcomes—like satisfaction and time well spent—it becomes the new default, and the cycle continues.
ByteDance’s feed is often described as “addictive,” but what’s really happening is a compounding feedback system. Each swipe is both a choice and a measurement.
When you watch, skip, like, comment, rewatch, or share, you’re generating signals that help the system guess what to show next.
A single view isn’t very informative on its own. But millions of tiny actions—especially repeated patterns—create a clear picture of what tends to hold your attention. The platform uses those signals to:
This is the flywheel: engagement → better matching → more engagement. As the matching improves, users spend more time; the extra time produces more data; the data improves matching again.
If the system only chased “more of what worked,” your feed would get repetitive fast. That’s why most recommendation systems deliberately include exploration—showing content that’s new, adjacent, or uncertain.
Exploration can look like:
Done well, it keeps the feed fresh and helps users discover things they didn’t know to search for.
A flywheel can spin in the wrong direction. If the easiest way to win attention is sensationalism, outrage, or extreme content, the system may over-reward it. Filter bubbles can form when personalization gets too narrow.
Platforms typically balance satisfaction and novelty with a mix of diversity rules, content quality thresholds, and safety policies (covered later in the article), plus pacing controls so “high-arousal” content doesn’t dominate every session.
When people talk about ByteDance, they usually point to recommendation algorithms. But there’s a quieter system doing just as much work: content logistics—the end-to-end process of moving a video from a creator’s phone to the right viewer’s screen, quickly, safely, and repeatedly.
Think of it like a supply chain for attention. Instead of warehouses and trucks, the system manages:
If any step is slow or unreliable, the algorithm has less to work with—and creators lose motivation.
A high-performing feed needs a constant flow of “fresh inventory.” ByteDance-style products help creators produce more often by lowering production effort: in-app templates, effects, music snippets, editing shortcuts, and guided prompts.
These aren’t just fun features. They standardize formats (length, aspect ratio, pacing) and make videos easier to finish, which increases posting frequency and makes performance easier to compare.
After upload, videos must be processed into multiple resolutions and formats so they play smoothly across devices and network conditions.
Fast processing matters because:
Reliability also protects the “session.” If playback stutters, users stop scrolling, and the feedback loop weakens.
At scale, moderation is not a single decision—it’s a workflow. Most platforms use layered steps: automated detection (for spam, nudity, violence, copyrighted audio), risk scoring, and targeted human review for edge cases and appeals.
Rules only work when they’re implemented consistently: clear policies, reviewer training, audit trails, escalation paths, and measurement (false positives, turnaround time, repeat offenders).
In other words, enforcement is an operational system—one that has to evolve as fast as the content does.
ByteDance’s advantage isn’t only “the algorithm.” It’s the way the product is built to generate the right signals for the feed—and to keep those signals flowing.
A great recommendation system needs steady supply. TikTok/Douyin reduce friction with an always-ready camera, simple trimming, templates, filters, and a large sound library.
Two design details matter:
More creators posting more often means more variation for the feed to test—and more chances to find a match.
The full-screen player removes competing UI elements and encourages one clear action: swipe. Sound-on by default increases emotional impact and makes trends portable (a sound becomes a shared reference).
This design also improves data quality. When each swipe is a strong yes/no signal, the system can learn faster than in cluttered interfaces where attention is split.
Remix formats turn “creation” into “replying.” That matters because replies inherit context:
In practice, remixing is built-in distribution—without needing followers.
Notifications can re-open the loop (new comments, creator posts, live events). Streaks and similar mechanics can raise retention, but they can also push people toward compulsive checking.
A useful product lesson: favor meaningful prompts (responses, follows you asked for) over pressure prompts (fear of losing a streak).
Small choices—instant playback, minimal loading, a single primary gesture—make the recommended feed feel like the default way to explore.
The product isn’t just showing you content; it’s training a repeated behavior: open app → watch → swipe → refine.
ByteDance didn’t “translate an app” and call it international. It treated globalization as a product problem and an operating-system problem at the same time: what people enjoy is intensely local, but the machinery that delivers it has to be consistent.
Localization starts with language, but quickly moves to context—memes, music, humor, and what counts as “good” pacing in a video.
Local creator communities matter here: early growth often depends on a small group of native creators who set the tone others copy.
Teams typically localize:
As usage grows, the feed becomes a logistics operation. Regional teams handle partnerships (labels, sports leagues, media), creator programs, and policy enforcement that reflects local law.
Moderation scales in layers: proactive filters, user reports, and human review. The goal is speed and consistency—removing clear violations quickly while handling edge cases with local expertise.
Going global means living inside app store rules and device constraints. Updates can be delayed by review processes, features may differ by region, and low-end phones force tough choices on video quality, caching, and data usage.
Distribution isn’t a marketing footnote; it shapes what the product can reliably do.
Trends can appear and disappear in days, while policy writing and enforcement training take weeks. Teams bridge the gap with “temporary rules” for emerging formats, rapid enforcement guidance, and tighter monitoring during volatile moments—then later convert what worked into durable policy and tooling.
For more on how the feed is supported behind the scenes, see /blog/content-logistics-hidden-system-behind-the-feed.
ByteDance’s feed is often described as an “algorithm,” but it behaves more like a marketplace. Viewers bring demand (attention). Creators supply the inventory (videos). Advertisers fund the system by paying for access to that attention—when it can be reached predictably and safely.
Creators don’t just upload content; they produce the raw material the recommendation system can test, distribute, and learn from.
A constant flow of fresh posts gives the platform more “experiments” to run: different topics, hooks, formats, and audiences.
In return, platforms offer incentives that shape behavior:
Brands usually care less about viral luck and more about repeatable outcomes:
Recommendation allows niche communities to flourish without needing huge follower counts. At the same time, it can rapidly concentrate attention into mass trends when many viewers respond similarly.
That dynamic creates a strategic tension for creators: niche content can build loyalty; trend participation can spike reach.
Because distribution is performance-based, creators optimize for signals the system can read quickly: strong openings, clear formats, series behavior, and consistent posting.
It also rewards “readable” content—obvious topics, recognizable audio, and repeatable templates—because it’s easier to match to the right viewers at scale.
ByteDance’s superpower—optimizing feeds for engagement—creates a built-in tension. The same signals that tell a system “people can’t stop watching this” don’t automatically tell it “this is good for them.” At small scale, that tension looks like a UX issue. At TikTok/Douyin scale, it becomes a trust issue.
Recommendation systems learn from what users do, not what they later wish they’d done. Quick replays, long watch time, and late-night scrolling are easy to measure. Regret, anxiety, and compulsive use are harder.
If a feed is tuned only for measurable engagement, it can over-reward content that triggers outrage, fear, or obsession.
A few predictable risks show up across markets:
None of these require “bad actors” inside the company; they can emerge from ordinary optimization.
People often ask for a simple explanation: “Why did I see this?” In practice, ranking mixes thousands of features (watch time, skips, freshness, device context, creator history) plus real-time experiments.
Even if a platform shares a list of factors, it still won’t map cleanly to a single, human-readable reason for one specific impression.
Safety isn’t just moderation after the fact. It can be designed into the product and operations: friction for sensitive topics, stronger controls for minors, diversification to reduce repetitive exposure, limits on late-night recommendations, and clear tools to reset or tune the feed.
Operationally, it means well-trained review teams, escalation paths, and measurable safety KPIs—not only growth KPIs.
Policies about what’s allowed, how appeals work, and how enforcement is audited directly affect trust. If users and regulators believe the system is opaque or inconsistent, growth becomes fragile.
Sustainable attention requires not just keeping people watching, but earning permission to keep showing up in their lives.
ByteDance’s success makes “recommendations + fast shipping” look like a simple recipe. The transferable part isn’t any single model—it’s the operating system around discovery: tight feedback loops, clear measurement, and serious investment in the content pipeline that feeds those loops.
Fast iteration works when it’s paired with measurable goals and short learning cycles. Treat every change as a hypothesis, ship small, and read results daily—not quarterly.
Focus metrics on user value, not just time spent. Examples: “sessions that end with a follow,” “content saved/shared,” “surveyed satisfaction,” or “creator retention.” These are harder than raw watch time, but they guide better trade-offs.
Engagement-only optimization without guardrails. If “more minutes” is the scoreboard, you will eventually reward low-quality, polarizing, or repetitive content because it’s reliably sticky.
Also avoid the myth that algorithms remove the need for editorial judgment. Discovery systems always encode choices: what to boost, what to limit, and how to handle edge cases.
Start with constraints, not slogans:
Recommendations depend on content logistics: tooling, workflows, and quality control. Invest early in:
If you’re budgeting, price the whole system—models, moderation, and support—before scaling (/pricing).
A practical note for teams building software products: many of these “system” investments (dashboards, internal tools, workflow apps) are straightforward to prototype quickly if you can shorten the build–measure–learn loop. Platforms like Koder.ai can help here by letting teams vibe-code web apps through a chat interface, then export source code or deploy—useful for spinning up experimentation dashboards, moderation queue prototypes, or creator operations tooling without waiting on a long traditional build pipeline.
For more product thinking like this, see /blog.
ByteDance’s core product thesis can be summarized in a simple equation:
recommendation algorithms + content logistics + product design = a scalable attention engine.
The algorithm matches people with likely-interesting videos. The logistics system ensures there’s always something to watch (supply, review, labeling, distribution, creator tools). And the product design—full-screen playback, fast feedback signals, low-friction creation—turns every view into data that improves the next view.
Some important details remain unclear or hard to verify without internal access:
Rather than guessing, treat public claims (from the company, critics, or commentators) as hypotheses and look for consistent evidence across disclosures, research, and observable product behavior.
If you want to go deeper without getting overly technical, focus on these topics:
If you keep these questions handy, you’ll be able to analyze TikTok, Douyin, and any future feed product with clearer eyes.
Zhang Yiming’s product thesis was that software should continuously filter information for you, using behavior signals, so the experience improves with every interaction. In a world of content overload, the product’s job shifts from “help me find information” to “decide what’s most relevant right now.”
A social graph feed is driven by who you follow; an interest graph feed is driven by what you do (watch, skip, rewatch, share, search). The interest-graph approach can work even when you follow nobody, but it depends heavily on getting recommendations good enough early and learning quickly from feedback.
Most feeds do two main things:
Candidate generation finds “possible fits”; ranking decides the final order where tiny changes can reshape what you watch next.
Strong signals usually come from observable behavior, especially:
Likes and comments matter, but watch behavior is often the most reliable because it’s harder to fake at scale.
For new users, platforms start with diverse, “safe” popular content in your language/region to quickly detect preferences. For new videos, they often do a small trial distribution to likely-interested groups and expand if engagement is strong. Practically, this means unknown creators can break through without a large follower base—if early performance is good.
Exploration prevents the feed from becoming repetitive by intentionally testing adjacent or uncertain content. Common tactics include:
Without exploration, the system can overfit and create narrow loops that feel stale or overly polarized.
Runaway optimization happens when the easiest way to win attention is sensational or extreme content, so the algorithm unintentionally rewards it. Platforms try to counter this with diversity rules, quality thresholds, and safety policies, plus pacing controls that prevent high-arousal content from dominating every session.
Content logistics is the end-to-end pipeline that moves content from a creator’s phone to the viewer’s screen:
If this pipeline is slow or unreliable, recommendations suffer because the system gets less (and lower-quality) inventory and weaker feedback loops.
Low-friction creation tools (templates, effects, sound libraries, easy editing) increase posting frequency and standardize formats, which makes content easier to test and compare. Remix mechanics (duets/stitches) also act as built-in distribution by anchoring new posts to proven clips, helping the system understand context and interests faster.
A/B testing turns product decisions into measurable hypotheses. Teams ship small changes (UI tweaks, ranking weights, notifications), measure outcomes, and roll forward winners quickly. To keep it responsible, use metrics beyond raw watch time (e.g., satisfaction, saves/shares, “not interested” rate, complaint rate) so growth doesn’t come at the cost of user well-being.