Decoded from Twitter's open-source algorithm. 145,000 community clusters. 6,000 features per tweet. Every signal mapped.
Not all engagement is equal. The algorithm uses weighted predictions for each signal type. Here's the actual hierarchy from the codebase:
| Rank | Signal | Algorithmic Impact |
|---|---|---|
| #1 | Replies + Author Reply-Back PREDICTED_IS_REPLIED_REPLY_ENGAGED_BY_AUTHOR | Highest weight. When someone replies AND you engage back, massive boost. |
| #2 | Bookmarks PREDICTED_IS_BOOKMARKED | Signals "save-worthy" content. Tracked separately from likes. |
| #3 | Shares PREDICTED_IS_SHARED | External distribution intent. High value signal. |
| #4 | Quality Clicks (15s+ dwell) PREDICTED_IS_GOOD_CLICKED_V2 | Click + meaningful time spent = quality engagement proof. |
| #5 | Profile Click → Engagement PREDICTED_IS_PROFILE_CLICKED_AND_PROFILE_ENGAGED | Shows author interest. Builds Real Graph relationships. |
| #6 | Replies PREDICTED_IS_REPLIED | Conversation value. Weighted significantly higher than likes. |
| #7 | Retweets / Quote Tweets PREDICTED_IS_RETWEETED | Distribution signal. Quote tweets add context value. |
| #8 | Video 50%+ Watch PREDICTED_IS_VIDEO_QUALITY_VIEWED | Video-specific quality metric. Completion rate matters. |
| #9 | Likes PREDICTED_IS_FAVORITED | Basic approval. Lowest positive signal in the hierarchy. |
Key Insight: The most underutilized signal is replying to replies on your own tweets. Most accounts optimize for likes (lowest signal) while ignoring the highest-value behavior.
Within the first hour, engage back with replies on your tweets. This triggers REPLY_ENGAGED_BY_AUTHOR — the highest-weighted positive signal.
Tutorials, threads, resources, insights. Bookmarks are tracked separately from likes and signal "I want to return to this."
The algorithm tracks PREDICTED_IS_DWELLED. Longer, substantive content that holds attention (15+ seconds) scores better.
Consistently engage with target accounts. The algorithm builds relationship strength via mutual follows, profile visits, and direct replies.
TweepCred uses PageRank — engagement from high-authority accounts carries more algorithmic weight. Target quality over quantity.
"What's your take?" beats "Like if you agree." Replies are weighted much higher than likes in the ranking model.
SimClusters builds embeddings based on who engages with you. Topic consistency = stronger algorithmic identity matching.
More followers than following. Reputation.scala explicitly penalizes accounts with inverse ratios.
You're leaving the highest-value signal on the table. Every ignored reply is a missed opportunity for algorithmic boost.
Likes are the lowest positive signal. Content that earns only likes underperforms content that earns replies, bookmarks, or shares.
Unfollows are tracked as NEGATIVE_SIGNAL. Pattern detection flags coordinated behavior. Damages Real Graph scores.
"Feedback Fatigue" system down-ranks repetitive authors by up to 80%. Penalty persists for 140 days.
Graph analysis detects coordinated engagement patterns. Artificial engagement undermines authentic Real Graph building.
Models scan for toxicity (pToxicity), harassment (pAbuse), NSFW content. Triggers visibility filtering and downranking.
The algorithm uses semantic understanding via SimClusters, not hashtag counts. Stuffing looks spammy without adding value.
First line determines if content gets ranked. Weak hooks = low initial engagement = algorithmic death spiral.
| Signal | Penalty Weight | Duration | Recovery |
|---|---|---|---|
| Reports | -20,000 (Extreme) | Persistent | Difficult — account-level impact |
| Blocks / Mutes | -1,000 (Severe) | Persistent | Gradual — through positive signals |
| "Not Interested" Clicks | -95% Similar Content | 140 days | Time-based decay |
| Unfollows After Engagement | Real Graph Damage | Decaying | Rebuild through consistent engagement |
| Tweet Unliking | Negative Signal | Tracked | Minor — offset by new engagement |