TikTok Signal Hierarchy

#tiktok #recommendation-systems #machine-learning #user-behavior

TikTok Signal Hierarchy

Not all user interactions are equal. TikTok's algorithm weights signals based on how strongly they indicate genuine user satisfaction. Understanding this hierarchy is key to understanding why the feed works the way it does.

The hierarchy (2026)

Signal	Weight	Type	What it tells the algorithm
Completion rate	Highest (~40-50%)	Implicit	User watched the video -- the clearest satisfaction signal
Rewatch / loop	Very high	Implicit	User watched it multiple times -- unusually strong interest
Shares	High	Explicit	User thought it was worth showing to others
Saves	High	Explicit	User wants to return to this content later
Comments	Medium-high	Explicit	User engaged enough to type a response
Follow from video	High	Explicit	User converted to a long-term relationship
Likes	Medium	Explicit	Low-friction action; easy to give, weak signal
Early swipe (<3s)	Negative	Implicit	Content was misleading or unengaging
"Not interested"	Very negative	Explicit	Strong explicit rejection

Why implicit signals dominate

TikTok prioritizes implicit signals (what you do) over explicit signals (what you say). The reasoning:

Actions are harder to fake. You either watched the video or you didn't. Likes are cheap to give.
Actions reveal true preference. A user might "like" a video to support a creator but scroll past similar content. Watch time reveals what they actually enjoy.
Actions are continuous, not binary. Completion rate is a gradient (20%, 50%, 80%, 100%) while a like is just 0 or 1.

The shift from engagement to satisfaction

Early social media algorithms optimized for engagement -- clicks, likes, comments. This rewarded clickbait and outrage. TikTok's shift to completion rate as the primary signal rewards content that genuinely holds attention, which is why the platform is known for "authentic" unpolished content rather than clickbait.

The like paradox

Likes are the most commonly understood interaction but one of the weakest signals. The technical reasons:

Low friction -- Tapping a heart requires minimal cognitive effort
Social motivation -- Users like content to support creators, not because they want more of it
No information about consumption -- A like tells you the user saw something they approved of, but not whether they watched the whole video
Manipulable -- Like-for-like schemes, bot farms, and engagement pods can inflate likes without genuine interest

The algorithm treats likes as a weak positive signal, not a strong one. A video with high completion rate and few likes will outperform a video with many likes and low completion rate.

Saves and shares: the new gold standard

In 2026, saves and shares have been elevated to the highest-quality interaction signals:

Save: "This is valuable enough that I want to come back to it." Signals utility, educational value, or profound resonance.
Share: "This is valuable enough that I want someone else to see it." The ultimate endorsement -- it brings other users into the app or extends viewing sessions.

Why saves and shares beat likes

A like costs nothing. A save requires intent to return. A share requires putting your social reputation behind the content. The algorithm rewards signals that are expensive to produce.

Negative signals

Negative signals are weighted heavily in the opposite direction:

Negative Signal	Impact
Early swipe (<3 seconds)	Heavy penalty -- signals misleading content
"Not interested" long-press	Strong explicit rejection
Fast-forwarding	User didn't want to watch at normal speed
Muting a creator	Persistent negative signal for that creator
Reporting	Severe penalty pending review

A video that triggers many early swipes gets progressively fewer impressions, even if its completion rate among those who stay is high. The algorithm interprets early drop-off as a mismatch between the video's promise and its content.

How signals flow through the pipeline

Two-Tower Neural Network uses implicit signals (watch history, completion) to build user embeddings
Multi-Stage Ranking Pipeline uses the full signal set to score candidates
Monolith Framework incorporates signals in real-time to update the model
Progressive Amplification Waves use signal velocity to determine distribution expansion

The signal weight shift over time

Year	Primary Signal	Secondary Signals
2020	Likes, comments	Watch time
2021	Watch time	Likes, shares
2023	Completion rate	Shares, saves
2025	Completion + saves	Shares, search intent
2026	Completion + saves + shares	Search intent, topic authority