How TikTok Recommends Videos: The Algorithm Explained

Deep Dive

1. 01

   The Interest Graph vs. Social Graph

   Core Architecture

   Unlike Instagram or Facebook which rely on social connections (who you follow), TikTok uses an interest graph — it maps what you're interested in based on behavior, not who you know. This is why a video from someone with 0 followers can reach millions. The algorithm doesn't care about follower count; it cares about content-audience fit.

   Example

   A new creator posts their first video about indoor gardening. The algorithm identifies the topic (gardening, plants, home decor) and tests it with users who've shown interest in similar content — regardless of whether they follow the creator.

2. 02

   Watch Time Is the #1 Signal

   Ranking Signals

   Average watch time (as a percentage of video length) is the single most important metric. A 15-second video watched to the end (100% completion) gets a stronger signal than a 60-second video watched for 30 seconds (50% completion). The algorithm also values re-watches — looping a video signals extremely high quality.

   Example

   A 7-second video that users watch 3 times (300% watch time) gets far more distribution than a 60-second video with 40% average watch time — even if the longer video has more total seconds watched.

3. 03

   The Engagement Cascade: Shares > Saves > Comments > Likes

   Ranking Signals

   Not all engagement is created equal. TikTok weights different engagement types differently: shares indicate the highest value (you're telling friends about it), saves indicate lasting value, comments indicate conversation-worthy content, and likes are the weakest signal. Optimizing for shares and saves outperforms chasing likes.

   Example

   A video gets 10K likes but only 100 shares. Another video gets 3K likes but 2K shares. The second video gets significantly more algorithmic distribution because shares are weighted ~5x more than likes.

4. 04

   Content Classification Through AI

   Content Analysis

   TikTok's AI analyzes every frame of your video — objects, text, faces, scenes, movements — plus audio transcription, captions, hashtags, and sounds. This multi-modal analysis creates a rich content profile that the algorithm uses to match your video with the right audience clusters.

   Example

   You film a video in a gym with dumbbells visible, say 'bicep curl technique,' and use the hashtag #fitness. The AI classifies this across multiple signals: visual (gym equipment), audio (fitness instruction), text (workout), hashtag (fitness community).

5. 05

   The Batch Testing System

   Distribution System

   Every video goes through a batch testing process. First, it's shown to ~500 users in your region. If engagement exceeds a threshold, it's shown to ~5,000. Then ~50,000. Then potentially millions. Each batch must pass an engagement threshold to reach the next level. Videos that fail at any stage stop expanding.

   Example

   Your video is shown to 500 users → 8% engage (above 5% threshold) → expanded to 5,000 → 6% engage → expanded to 50,000 → 4% engage → expansion slows. Each batch is a new test with higher bars.

6. 06

   Negative Signals Matter Too

   Ranking Signals

   The algorithm tracks negative signals: 'Not Interested' taps, rapid scroll-aways (watching less than 1 second), hiding a creator, and reporting. These negative signals are weighted heavily — a few 'Not Interested' taps can significantly reduce a video's distribution.

   Example

   If 2% of the test audience taps 'Not Interested' on your video, the algorithm dramatically reduces distribution — even if the other 98% had moderate engagement. Negative signals are amplified.

7. 07

   Sound and Music as Discovery Vectors

   Discovery Mechanics

   TikTok's algorithm treats sounds as content categories. When a sound trends, the algorithm actively promotes videos using that sound to users who've engaged with other videos using the same audio. Using a trending sound is like getting a distribution boost from the sound's existing audience.

   Example

   A sound has 50K videos. Users who've watched and liked videos with this sound are more likely to be shown your video using the same sound — the sound itself acts as a distribution channel.

8. 08

   Time Decay and Recency

   Timing

   Newer content gets a distribution advantage. The algorithm prioritizes recently posted videos over older ones. However, TikTok also has 'delayed virality' — if an older video suddenly gets engagement (from a share or being rediscovered), the algorithm can resurface it.

   Example

   A video posted 3 weeks ago suddenly gets shared in a group chat, generating a burst of views and engagement. The algorithm detects this new activity and pushes the video back into active FYP distribution.

9. 09

   Device and Network Signals

   Technical Signals

   TikTok considers technical signals: device type, network connection, language settings, and location. These help personalize content delivery — higher-resolution videos may be prioritized on Wi-Fi, while shorter videos might be preferred on cellular data.

   Example

   The algorithm might slightly favor different content lengths based on a user's typical watch session length, which correlates with whether they're on Wi-Fi (longer sessions) or mobile data (shorter sessions).

10. 10

    Diversity Injection

    Feed Design

    To prevent filter bubbles, TikTok intentionally injects diverse content into your feed — videos from outside your normal interest categories. This serves two purposes: it discovers new interests you didn't know you had, and it tests content with audiences outside its primary classification.

    Example

    You primarily watch cooking content, but the algorithm occasionally shows you a travel video. If you engage with it, the algorithm learns you're also interested in travel and starts mixing travel content into your regular feed.

11. 11

    Creator Authority Score

    Account Signals

    Each account develops an implicit authority score within specific content categories. Creators who consistently produce high-engagement content in one niche build stronger authority, leading to better initial distribution for their new videos in that category.

    Example

    A creator who has posted 100 high-performing fitness videos gets better initial distribution on their 101st fitness video than a creator posting their first fitness video — even if the content quality is identical.

12. 12

    Regional Distribution Weighting

    Distribution System

    Your account's registered region heavily influences initial distribution. Content is first tested with users in your region before potentially expanding globally. This is why account region is so critical — a US-region account gets tested on the largest, most active TikTok audience.

    Example

    A US-region account's video is tested with 500 US users first. A Vietnamese-region account's video is tested with 500 Vietnamese users. The US test pool is more diverse and engaged, giving the US account's video more expansion potential.