TikTok Algorithm

Optimize short-form video for TikTok's recommendation system using evidence, not folklore. This skill separates what is documented (TikTok's own statements + ByteDance's published research) from what is inferred (creator-observed best practice), and never presents guesses as facts.

When to Activate

• Planning or scripting a TikTok / Reels / Shorts video and deciding what to optimize for
• Choosing caption, hashtags, sounds, or posting cadence
• Diagnosing why a video under- or over-performed
• Explaining "how the For You Page works" without hype or myths

Two-Layer Mental Model

Reason about TikTok on two layers. Confusing them is the most common mistake.

Layer 1 — The System (how the model learns)

Source: ByteDance's Monolith recommender, open-sourced 2022 (github.com/bytedance/monolith, arXiv:2209.07663). Monolith is the training/serving framework, not the full "For You" ranking rulebook — treat it as directional, not a spec.

Two published properties matter for creators:

1. Real-time / online training. The model interacts with user feedback in near-real-time and explicitly trades system reliability for freshness to capture emerging trends fast. → Implication: the first hours after posting dominate. Early completion, replays, and shares are what the model learns from first. A weak open is not "slow to catch on" — it teaches the model to stop distributing.

2. Collisionless embedding tables (Cuckoo-hash based), with expirable embeddings and frequency filtering of rare features. → Implication: distinct signals are stored distinctly and persist — every interaction (watch %, rewatch, share, save, comment) feeds the profile. But stale/low-frequency signals decay, so consistency and recency beat a single old hit. Follower count is not a stored ranking feature (see Myths).

Layer 2 — The Ranking Signals (what gets weighted)

Source: TikTok's official newsroom, "How TikTok recommends content." Three factor categories, weighted unequally:

1. User interactions — likes, shares, follows, comments, completions, content you create. (Highest weight.)
2. Video information — captions, sounds, hashtags. (How the system understands/indexes the video.)
3. Device & account settings — language, country, device type. (Lowest weight; not an expressed preference.)

TikTok states finishing a longer video start-to-end is a strong indicator of interest weighted above demographic factors.

Signal Hierarchy (optimize top-down)

Ranked by documented + widely-observed weight. Spend effort where the weight is.

| Tier | Signals | What it rewards | |---|---|---| | Strongest | Watch time, completion rate, rewatches/loops | Content that holds and repeats attention | | Strong | Shares, saves/favorites, comments, follows-from-video | Content worth passing on or keeping | | Moderate | Likes | Passive approval (cheap signal) | | Lowest | Device/account settings | Context, not merit |

Rule of thumb: a share or a rewatch is worth far more than a like. Design for the top of this table first.

Actionable Playbook

Hook (first 2 seconds)

• 71% of viewers decide in ~3s; on TikTok treat the budget as 2s. Lead with the payoff, tension, or number — never a warm-up.
• The hook's job is to buy the next 2 seconds. Front-load the single strongest claim or visual.

Hold completion (the #1 lever)

• Match length to substance. Shorter, fully-watched > longer, half-watched. Aim for a completion-friendly length (often 21–34s for awareness clips; only go longer if every second earns it).
• Cut every 2–3s; remove dead air; burn captions (many watch muted).
• Build a reason to reach the end (payoff, punchline, or list that resolves).

Engineer replays & shares

• Replays: loop the ending into the opening, or pack info too dense to absorb once ("save this").
• Shares: make it useful, surprising, or identity-expressing ("send this to a dev friend").
• Saves: reference-style value (steps, a list, a resource) earns saves.
• Comments: plant one genuine question; pin a comment to seed replies. No engagement-bait if it's dishonest to the content.

Video information (indexing layer)

• Captions: include real keywords the target searches — TikTok's AI reads captions to classify the video.
• Sounds: a trending/relevant sound can aid discovery; relevance over chasing every trend.
• Hashtags: maximum 5. TikTok rolled out a 5-hashtag limit (Aug 2025). Mix a couple of broad-reach tags with a few targeted niche tags. Every tag must earn its place — no stuffing.

Timing & consistency

• Because training is real-time, post when your audience is active so early signals land fast.
• Consistency compounds (embeddings favor recent, repeated signals); sporadic posting decays.

Diagnosing a Weak Video

Walk the funnel top-down and fix the earliest break:

1. Low reach / stalled at a few hundred views → weak early signals. Suspect the hook or completion, not "shadowban."
2. Views but low completion % → open is fine, middle sags. Tighten pacing, cut length.
3. Good completion but no shares/saves → entertaining but not passable/useful. Add a payoff worth sending.
4. Good everything, still flat → try a different sound/caption framing; the indexing layer may be mis-classifying the niche.

Myths to Reject (stay honest)

• "Follower count boosts the algorithm." TikTok states follower count and past high-performers are not direct ranking factors. Big accounts get more views via their existing audience, not a model boost.
• "There's a magic post time / hashtag that guarantees FYP." No single tag or time overrides weak retention.
• "Deleting and reposting resets the algorithm." No evidence; you lose the original's accumulated signals.
• "Shadowban explains my low views." Usually weak early-retention signals. Only community-guideline violations restrict distribution; verify in the app rather than assuming.

Honest Limitations

• Monolith is a framework, not the deployed "For You" ranking logic — use it for system intuition, not exact weights.
• Exact ranking weights are not public. This skill states directional weight (what TikTok confirms is stronger/weaker), never invented percentages.
• The platform changes; re-verify specifics (e.g., hashtag caps, length sweet spots) against current sources before relying on them.

Output Format

When asked to optimize content, return:

• Which signals this piece targets (from the hierarchy) and why
• Hook, completion, and share/save mechanics baked into the plan
• Caption + ≤5 hashtags (broad + niche mix)
• What to watch in the first hours, and the diagnostic to run if it stalls

Sources

• TikTok Newsroom — How TikTok recommends content: https://newsroom.tiktok.com/en-us/how-tiktok-recommends-videos-for-you
• Monolith (paper), ByteDance, RecSys 2022: https://arxiv.org/abs/2209.07663
• Monolith (code), ByteDance: https://github.com/bytedance/monolith
• TikTok 5-hashtag limit (Aug 2025): https://www.socialmediatoday.com/news/tiktok-implements-five-hashtag-limit-per-post/757857/

Related Skills

• video / video-marketing — script structure and hooks
• content-engine — platform-native drafting
• crosspost — per-platform adaptation (Reels penalizes TikTok watermarks; Shorts needs searchable titles)