$ man content-wiki/x-algorithm

The X recommendation system is an open-source pipeline that decides what appears in your For You feed. It was released publicly as the xai-org/x-algorithm repo and has over 15,000 stars on GitHub. The system has four major components: Home Mixer is the orchestration layer that coordinates everything. Thunder handles in-network retrieval — pulling tweets from people you follow. Phoenix is the Grok-based transformer model that scores and ranks candidates. And the Candidate Pipeline filters, deduplicates, and assembles the final feed. Understanding this architecture is the difference between guessing what works on X and knowing what the system rewards.

PATTERN

Every tweet that enters your feed goes through a multi-action prediction model. The system predicts the probability of twelve different engagement actions: P(favorite) — will you like it. P(reply) — will you reply. P(repost) — will you retweet. P(quote) — will you quote tweet. P(click) — will you click on it. P(share) — will you share externally. P(dwell) — will you spend time reading it. P(follow_author) — will you follow the author after seeing this. P(not_interested) — will you hit not interested. P(block_author) — will you block the author. P(mute_author) — will you mute the author. P(report) — will you report the tweet. The final score is a weighted sum: Final Score = sum(weight_i * P(action_i)). Positive actions (like, reply, repost, quote, share, dwell, follow) push the score up. Negative actions (not_interested, block, mute, report) push the score down. The weights are not equal — replies and quote tweets carry significantly more weight than likes.

PATTERN

This is where it gets actionable for creators. The algorithm assigns different weights to different engagement types. Replies are weighted heavily — a tweet that sparks conversation scores higher than one that gets passive likes. Quote tweets with added commentary are also high-weight because they represent deeper engagement. Reposts carry moderate weight. Likes are the lowest positive signal — they indicate interest but not deep engagement. On the negative side, any tweet that triggers mute, block, not_interested, or report signals gets penalized hard. A single block or report can outweigh dozens of likes. This means controversial content that gets engagement but also triggers negative signals can actually score lower than a modest post with clean positive engagement. The algorithm punishes polarizing content more than people realize.

PATTERN

Phoenix uses a two-tower neural network architecture. One tower represents the user — your interests, your engagement history, the types of content you interact with. The other tower represents candidate tweets — the topic, the author, the format, the engagement pattern. Both towers produce embedding vectors, and the system computes similarity between them. High similarity = the tweet is likely relevant to you. Key design decisions: Phoenix uses no hand-engineered features. Everything is learned from engagement data. It also uses candidate isolation — each tweet is scored independently, not relative to the other tweets in the batch. And it predicts multiple actions simultaneously rather than just one. This means the model captures nuanced engagement patterns that a simpler system would miss.

PRO TIP

Based on the algorithm architecture, here is what actually moves the needle: (1) Optimize for replies and quote tweets over likes. Write tweets that invite conversation, not passive agreement. Ask questions. State takes that people want to respond to. (2) Dwell time matters. Longer tweets that people actually read score higher than short tweets they scroll past. But only if the content earns the dwell — padding with filler hurts. (3) Avoid triggering negative signals. Rage-bait might get engagement, but if it also gets mutes and blocks, the net score drops. Clean engagement beats polarizing engagement. (4) Follow signals are gold. If your tweet makes someone follow you, that is one of the strongest positive signals. Build tweets that showcase expertise worth following. (5) Consistency builds your user tower. The more consistently you post about specific topics, the stronger your signal in the user-tower matching. Niche down. (6) In-network vs out-of-network: Thunder pulls from your followers first. Building a strong follower base that engages with your content means your tweets start with a higher baseline before Phoenix even scores them for the wider feed.

ANTI-PATTERN

Don't try to game the algorithm with engagement pods, reply chains with your own accounts, or coordinated liking. The system detects inorganic engagement patterns. Tweets that get a burst of likes from the same network within seconds of posting get flagged differently than tweets that accumulate engagement organically over time. The model is trained on massive datasets and can distinguish between genuine engagement and coordinated activity. The sustainable approach: write content that earns real engagement. The algorithm rewards authentic interaction because that is what keeps users on the platform. Align with that incentive, don't fight it.