How the TikTok algorithm actually decides what you see

Few systems in modern computing have been mythologized more than the TikTok For You algorithm. Most of the popular explanations are wrong — either oversimplified to the point of being useless, or invented from scratch by self-styled gurus. This article walks through what is actually known, drawing on the partial source-code leak from 2022, public talks by ByteDance engineers, court filings from the EU Digital Services Act discovery process and behavior we can observe directly by feeding test accounts.

The two-stage architecture

The For You ranker is not a single model. It is a two-stage cascade.

The first stage is **candidate generation.** From a pool of roughly tens of millions of recent and resurfaced videos, the system selects a few thousand candidates per user. This stage is fast, low-precision and runs entirely on lightweight embeddings. The goal is not to be right; it is to not miss anything that could be right.

The second stage is **ranking.** The few thousand candidates are scored by a heavier model that combines hundreds of features: predicted watch time, predicted engagement, content category match, freshness penalty, diversity boost. The top scores are stitched into the next batch of videos shown to the user.

This split matters because most of the public discussion conflates the two stages. When a creator says “I was suppressed,” what usually happened is that they passed candidate generation but failed ranking. That is a very different problem from being filtered out at the candidate stage, and the fixes are different.

What features actually feed the ranker

The leaked 2022 internal doc listed 234 features used in the production ranker at the time. The 2026 version is presumably larger, but the public structure has not changed dramatically. Features fall into five buckets.

**Content embeddings.** A neural network reads the video’s visual frames, audio waveform and on-screen text, and produces a dense vector that captures “what is this video about.” Embeddings allow the system to find similar videos for new users without explicit category tagging.

**User embeddings.** A parallel network builds a vector for each user from their full history of watches, likes, comments, shares and skips. Two users with similar embeddings will tend to see overlapping FYP feeds.

**Engagement predictions.** Per-video, the model predicts the probability that this user will watch to completion, like, comment, share, follow, save and skip. Each prediction is a separate sub-model that has been trained on hundreds of millions of labeled events.

**Behavioral features.** These are explicit signals: time of day, device type, network speed, geographic region. They modify how the engagement predictions are interpreted (a 30-second video on a slow connection has a higher skip probability for non-content reasons).

**Diversity and freshness.** A regularization layer prevents the FYP from collapsing into a single topic. It penalizes videos that are too similar to what the user has just watched, and it boosts very recent videos.

The myth of the “500 viewer test”

The most repeated piece of TikTok algorithm folklore is that every video gets a fixed test pool of 500 viewers. This is a useful simplification but it is not how the system works.

What actually happens is that a new video starts in the candidate pool of users whose embeddings match the video’s embedding. Distribution expands or contracts based on early engagement, but the “pool” size depends on the embedding cluster: a video targeted at a tight niche may only have a few hundred candidates, while a video that matches a broad interest cluster may start with tens of thousands.

The 500-viewer figure comes from the median test pool size for the most common embedding clusters around 2020. It became gospel because the order of magnitude is still roughly right, but treating it as a hard threshold leads creators to make wrong decisions about why a specific video underperformed.

Why retention is the dominant signal

Of all the engagement predictions, the one with the heaviest weight in 2026 is **predicted watch-through rate (WTR).** This is a deliberate product choice, not an accident of the model. ByteDance optimizes the FYP for total session length: how long a user stays in the app per visit. WTR is the single feature most correlated with session length.

Likes, in contrast, have a very low weight. The model treats a like as confirmation of an already-positive watch signal. A like without a strong watch is essentially ignored.

This is why videos with low like counts can still go viral on TikTok — if the watch-through is high, the algorithm pushes them regardless. It is also why “like for the algorithm” appeals from creators are essentially useless. A like is a downstream metric, not an input the algorithm waits for.

The follow loop

Following an account is not as binary a signal as it appears. The follow itself is a strong positive signal at the moment it happens. But the follow then opens a per-creator engagement channel that the algorithm tracks separately.

If you follow a creator and never watch their next 10 videos to completion, the algorithm gradually demotes their content in your feed even though you are still officially following them. This is sometimes called “silent unfollow” and it is the reason creators often complain that their reach drops even as their follower count grows. The follower count is real; the active follower fraction is much smaller.

What signals creators can actually influence

Given the architecture above, the levers a creator has are narrow:

**Hook design.** Influences the first 1-3 seconds, which dominate the WTR prediction.

**Topic consistency.** Tightens the user-content embedding match, increasing the candidate pool size for the right people.

**Posting cadence.** Affects how the freshness boost compounds across multiple videos. Posting less than once every 5 days reduces the recent-engagement priority.

**Reply behavior.** Replies to comments inside the first hour after posting trigger a quality signal that propagates to ranking.

**Cross-video coherence.** The algorithm builds a per-creator embedding from their last several videos. Coherent embeddings rank better than scattered ones.

Almost everything else — hashtags, trending sounds, posting time of day — is a tiny effect at best.

What the algorithm cannot do

Two persistent claims are technically false.

It cannot detect the political opinions of viewers and segregate them. The leak made clear that political-topic clustering exists, but the system is forbidden from using protected attributes (perceived ethnicity, religion, political affiliation) as ranking inputs, and the EU DSA filings confirm this constraint is enforced at the feature-engineering layer.

It cannot “listen through your microphone.” This claim has been debunked repeatedly by independent network monitors. The reason ads sometimes feel uncannily relevant is that the embedding system models you accurately enough to predict interests you have not explicitly expressed. That is more impressive than ambient surveillance, but less alarming.

Closing thought

The TikTok algorithm is not magic and it is not a black box. It is a well-engineered ranking system optimizing a single business metric — session length — through a small number of dominant features. Most of the mystique around it is a side effect of opaque communication from ByteDance, not actual technical novelty. Once you understand the two-stage cascade and the WTR-dominant ranker, the rest of TikTok’s behavior becomes predictable.