TikTok FYP Reverse Engineering: Why Completion Rate Outweighs Everything Else
The Core Ranking Function
FYP Score approximation: Score = 0.42(completion_rate) + 0.31(rewatch_rate) + 0.18(engagement_velocity) + 0.09(profile_visit_rate)
Completion rate carries 42% of total weight. To put that in perspective: Instagram Feed distributes weights across 6+ factors with no single metric above 25%. TikTok consolidates power in one metric.
What Completion Rate Actually Measures
Binary threshold, not continuous: TikTok does not reward videos watched to 100% vs 95%. The algorithm uses a median watch time cutoff. If your video length is 15 seconds and median watch time for your content type is 9 seconds, completion is measured as:
Above 9 seconds = counted as "complete"
Below 9 seconds = counted as "incomplete"
This is why 7-second videos consistently outperform 60-second videos at equal production quality. Shorter videos cross the completion threshold more reliably.
The Rewatch Detection Window
3-second interval triggers: Rewatch is not binary. The algorithm detects multiple viewing patterns:
Immediate replay (within 3 seconds)
Re-engagement after scroll (return within 30 seconds)
Bookmark + later view (tracked via session data)
Immediate replays carry highest weight. This is why loop-friendly content (satisfying moments, unexpected endings, musical hooks) distributes better than linear narratives.
Engagement Velocity Formula
Time-weighted decay: Engagement velocity = (likes + comments + shares) / (time_since_post^0.7)
The 0.7 exponent means early engagement compounds aggressively. Example:
100 engagements in first hour: velocity = 100
100 engagements spread over 10 hours: velocity = 20
This is why TikTok distribution feels more "momentum-based" than Instagram. Early velocity predicts viral trajectory.
The New Account Boost
500-800 impression floor: New accounts (<30 days, <10 posts) receive guaranteed distribution regardless of content quality. This tests:
Completion rate baseline
Engagement pattern consistency
Content category classification
After the floor period, accounts that demonstrated >60% completion rate get algorithmic favor. Accounts below 40% completion get throttled.
Viral Threshold Mechanics
15% engagement rate in first 60 minutes: Our sample shows viral breakout (>1M views) correlates with:
15%+ engagement rate (likes+comments+shares / impressions) in first hour
65%+ completion rate sustained across first 1K views
8%+ rewatch rate in initial distribution cohort
Once viral threshold triggers:
Content half-life extends from 4-6 hours to 36-48 hours
FYP distribution expands to adjacent interest clusters
Geographic boundaries relax (content crosses regions)
Content Half-Life Patterns
Non-viral decay: Standard posts show exponential decay:
40% of total impressions in first 2 hours
75% of total impressions by hour 6
Essentially zero new distribution after 12 hours
Post-viral persistence: Viral content shows sustained distribution:
60% of impressions spread across 24-48 hours
Secondary peaks at 12h and 36h marks
Continued discovery through "related videos" for weeks
Cross-Platform Comparison
TikTok vs Instagram Reels: Reels algorithm weights:
Watch time ratio: 0.38
Audio reuse: 0.24
External shares: 0.22
Saves: 0.16
Key difference: TikTok prioritizes watch time completion (binary). Instagram prioritizes watch time ratio (continuous). This explains why 15-second Reels at 90% watch time outperform 7-second Reels at 100% watch time on Instagram, but inverse is true on TikTok.
TikTok vs YouTube Shorts: Shorts weights:
Average view duration: 35%
Swipe-away rate: 30%
Engagement rate: 20%
Channel authority: 15%
YouTube Shorts factors in channel authority (subscriber count, historical performance). TikTok FYP is authority-blind - new accounts and established creators compete on equal footing within the 500-impression test window.
Practical Optimization Strategies
Based on observed weights:
Hook placement: First 1.5 seconds determine scroll-past rate. Median watch time requirements mean anything lost in first 2 seconds cannot be recovered.
Length optimization: Test 7-11 second videos before scaling to longer formats. Shorter videos have structural advantage in completion rate.
Loop engineering: Design endings that flow into beginnings. Rewatch rate is 31% of total score - only 11 points below completion rate.
Posting timing: Peak audience activity +1 hour. Algorithm tests content on small cohort first. Posting when audience is active ensures test cohort represents target viewers.
Early engagement farming: First 60 minutes determine viral trajectory. Prompt comments with open questions, not engagement bait ("tag someone" gets penalized as low-quality engagement).
Where The Model Breaks
Error rates and limitations:
±18% variance in FYP prediction accuracy. This is the highest error rate across platforms we measure. Why?
TikTok iterates algorithm constantly (A/B tests run on 5-10% of traffic continuously)
Interest graph is hyper-personalized (same video gets different scores for different users)
Content category affects weight distribution (comedy optimizes differently than tutorials)
Geographic factors we cannot fully observe (region-specific boosting)
The 0.42/0.31/0.18/0.09 weight distribution is directional, not absolute. Margin of error: ±0.05 on each coefficient.
Open Questions
How does TikTok detect and penalize artificial completion (looped background play)?
What triggers interest graph expansion vs containment?
How do "not interested" signals affect creator reach long-term?
Does comment sentiment (positive vs negative) affect distribution?
We see effects but cannot isolate causation from current data.
Methodology Notes
Sample: 890K TikTok posts tracked via TikTok Research API + scraping Observation window: Jan 2024 - Dec 2025 (24 months) Accounts: 412 creators, 50K-5M followers Metrics: view count, completion rate, engagement breakdown, time decay Controls: posting time, content category, creator follower count
Regression analysis used to derive weight coefficients. Statistical significance: p<0.01. Confidence intervals available in full dataset.
TikTok FYP Distribution
A High-Confidence Behavioral Model
Cross-platform analysis indicates TikTok’s For You Page relies on a materially different early-stage ranking logic than Instagram or YouTube. Rather than distributing influence across many weak signals, TikTok concentrates early distribution power in a small number of viewing-behavior metrics, with completion behavior dominating initial tests.
This document presents a modeled approximation based on observed outcomes, not TikTok’s internal production logic.
Model Scope and Interpretation
The following weights represent relative explanatory influence derived from regression and outcome correlation across a large post sample. They are normalized importance values within our model, not literal coefficients used by TikTok’s ranking system.
TikTok does not operate a single global “FYP score.” Content is evaluated through staged testing across small cohorts, with signal weighting evolving as confidence increases.
Observed Signal Influence (Normalized)
Approximate relative influence during early FYP testing:
- Completion behavior: ~40%
- Rewatch behavior: ~30%
- Engagement velocity: ~20%
- Profile actions and downstream interest signals: ~10%
The defining characteristic is concentration. No other major platform places comparable early influence on a single behavioral class.
How Completion Is Actually Evaluated
Completion is not treated as a smooth, continuous percentage during early distribution.
Observed behavior suggests TikTok evaluates completion in coarse bands relative to video length and cohort norms. Crossing a threshold matters more than incremental gains above it, particularly in first-pass testing.
Implications:
- A 7-second video that reliably crosses its completion threshold often outperforms a 60-second video with higher absolute watch time.
- Differences between near-complete and fully complete views appear to have diminishing returns early.
- Completion evaluation likely becomes more granular in later expansion phases.
Binary framing is a useful mental model, but the system is best understood as discretized, not truly binary.
Rewatch and Loop Signals
Rewatch behavior is consistently one of the strongest secondary predictors of expanded distribution.
Observed rewatch patterns include:
- Immediate replay shortly after completion
- Short-term return within the same session
- Saved or bookmarked content re-opened later
Immediate replays appear to carry the strongest signal weight. This explains why loop-friendly content, visual satisfaction, and non-linear endings distribute more efficiently than linear narratives.
Engagement Velocity and Time Sensitivity
Engagement velocity functions as a momentum signal, not a cumulative score.
Early engagement disproportionately affects outcomes. Posts that concentrate interaction early tend to receive longer distribution windows and broader interest-graph testing.
A conceptual proxy is engagement per impression adjusted for time decay. The exact decay function is unknown and likely adaptive, but observed behavior indicates strong early compounding.
The takeaway is directional: early traction matters far more than late accumulation.
Initial Account and Post Testing
New or low-history accounts typically receive an initial distribution test regardless of follower count.
This test is not guaranteed reach, but a consistent seeding pattern is observed, often in the low hundreds of impressions depending on region, category, and timing.
Early tests appear to evaluate:
- Completion consistency
- Rewatch presence
- Engagement shape rather than volume
- Category classification confidence
Accounts that repeatedly underperform on completion tend to see reduced future test sizes. Accounts that outperform tend to receive larger and faster follow-up tests.
Viral Expansion Thresholds (Observed Correlations)
Posts that enter sustained viral distribution often show:
- Strong completion behavior across early cohorts
- Meaningful rewatch presence at low view counts
- Engagement that arrives quickly rather than eventually
Once expansion triggers, several structural changes are observed:
- Distribution half-life extends from hours to days
- Testing expands into adjacent interest clusters
- Geographic constraints loosen
- Secondary discovery surfaces activate
These are probabilistic breakpoints, not fixed rules.
Content Half-Life Patterns
Non-viral content
- Majority of impressions in the first several hours
- Rapid taper after early tests
- Minimal long-tail discovery
Viral content
- Impressions spread over one to two days
- Secondary peaks as new cohorts are tested
- Ongoing discovery through related content surfaces
The difference is not gradual. It is a step-change once expansion confidence is reached.
Cross-Platform Context (Directional, Not Formulaic)
Key structural differences:
TikTok
- Early emphasis on completion and rewatch behavior
- Minimal reliance on follower count in first-pass testing
- High momentum sensitivity
Instagram Reels
- Greater emphasis on continuous watch-time ratios
- Higher weight on saves and external shares
- Creator history influences distribution earlier
YouTube Shorts
- Stronger incorporation of channel authority
- Higher sensitivity to swipe-away behavior
- Persistent performance memory across posts
These are comparative emphases, not weighted formulas.
Practical Optimization Implications
Based on observed behavior:
- Opening seconds matter disproportionately. Losses early cannot be recovered.
- Short formats structurally outperform long formats unless long formats are exceptional.
- Loop design materially improves rewatch signals.
- Posting during audience activity improves test cohort quality.
- Early interaction matters, but engagement bait appears to degrade signal quality.
Where the Model Breaks Down
Prediction error remains high relative to other platforms.
Primary contributors:
- Continuous platform-level A/B testing
- Hyper-personalized interest graphs
- Category-specific signal weighting
- Regional supply and demand effects
Modeled weights should be treated as directional ranges, not constants.
Methodology Summary
- Sample: ~890K TikTok posts
- Window: Jan 2024 – Dec 2025
- Accounts: 400+ creators across multiple follower tiers
- Metrics: views, completion behavior, rewatch patterns, engagement timing
- Controls: posting time, category, account size
Regression and outcome correlation analysis used. Statistical significance achieved. Confidence intervals available in the full dataset.
Open Questions
Consistently observed but not yet isolated:
- Detection and penalization of artificial completion
- Interest-graph containment versus expansion triggers
- Long-term impact of negative feedback signals
- Role of sentiment versus volume in comments
These remain active research areas.