How the TikTok Algorithm Works: FYP Ranking Signals Explained
The TikTok algorithm is the recommendation system behind the For You page (FYP) — the default feed that serves each user a personalized stream of videos, most of them from accounts the user does not follow. Unlike a chronological or follower-based feed, the FYP is assembled one video at a time by predicting how likely you are to watch, finish, and engage with each candidate clip.
TikTok has published general descriptions of the system, and years of creator analytics have filled in the practical details. The picture that emerges is consistent: every video is evaluated on its own performance, interest signals outweigh social signals, and distribution is earned in stages rather than granted upfront.
The core idea: per-video testing, not per-account reach
When a video is published, TikTok shows it to a small initial pool of users — partly followers, partly non-followers whose interest profiles match the video's content. How that pool responds determines whether the video graduates to a larger pool, and the process repeats in expanding waves. A video that keeps clearing performance thresholds can reach millions of viewers; one that stalls in an early wave quietly stops being distributed.
This batch-testing design explains two things creators notice constantly: an account with 200 followers can land a million-view video, and an account with 500,000 followers can post a video that dies at 2,000 views. Follower count influences the composition of the first test pool, but it is not a meaningful ranking signal beyond that. Each upload starts the audition over.
The ranking signals, roughly in order of weight
TikTok scores candidate videos against each viewer using predicted engagement. Based on TikTok's own statements and large-scale creator data, the signals stack approximately like this:
- Watch time and completion — whether viewers watch the video through, and for how long. This is widely understood to be the heaviest signal.
- Rewatches and loops — watching a video more than once is one of the strongest positive indicators the system can record.
- Shares and saves — sending a video to someone or bookmarking it signals high value, and creators consistently see these correlate with extended distribution.
- Comments — both the count and whether viewers stay on the video while reading or writing them.
- Likes and follows from the video — positive, but weaker than the signals above.
- Negative feedback — taps on "Not interested," hides, and reports actively suppress a video and inform the viewer's future feed.
Content signals: how TikTok knows what a video is about
Before a video can be matched to interested viewers, the system has to classify it. TikTok analyzes the audio track (including automatic speech transcription), on-screen text, the caption, hashtags, the sound used, and increasingly the visual content of the frames themselves. These content signals decide which interest clusters a video is tested against — a clip classified as "home cooking" gets its first audition in front of users who watch cooking content.
Device and account settings — language, country, and device type — are also inputs, but TikTok describes them as low-weight tiebreakers. They shape candidate pools more than rankings.
What the algorithm filters out
Not every video is eligible for full FYP distribution. TikTok maintains a tier between "removed for violating rules" and "fully recommendable": content that is allowed on the platform but ineligible or demoted in the For You feed. This includes borderline-sensitive material, spam-like or unoriginal content (such as reposts with visible watermarks from other platforms), and videos under review. A video in this tier can still be viewed via profile, search, or direct link — it just stops being pushed.
The system also deliberately injects diversity: TikTok states that the feed intersperses videos outside a user's established interests to avoid repetitive loops. This is why niche accounts occasionally see bursts of out-of-niche viewers in their analytics.
Common misconceptions
Three persistent myths are worth correcting. First, there is no credible evidence that the algorithm reads engagement in the first 30 or 60 minutes as a make-or-break window; distribution waves can play out over hours, days, or even weeks, and old videos regularly resurface. Second, hashtags do not control reach — they are one classification input among many, and stuffing them does nothing the audio and visuals haven't already done. Third, the algorithm does not punish accounts for a single underperforming video; each upload is tested independently, even though a long run of weak content can erode how favorably initial test pools respond.