TL;DR
The Heavy Ranker is X's core recommendation engine, a neural network called the Phoenix transformer that scores every post across 19 engagement signals. It calculates a weighted sum of predicted user actions (replies, retweets, likes, dwell time, and more) to produce a final score that determines whether content reaches the For You feed or dies with zero distribution.
When you publish a post on X, it enters a pipeline. First, a candidate retrieval system pulls roughly 1,500 posts that might interest a given user, sourced from accounts they follow (in-network) and accounts matched by machine learning (out-of-network). Then the Heavy Ranker scores every single candidate.
The scoring formula is: Final Score = Sum(weight_i x P(action_i)) across 19 engagement signals. Each signal has a different weight. The algorithm predicts how likely a user is to reply, retweet, like, click the author's profile, spend time reading, share, or take negative actions like muting or blocking.
Posts with the highest final scores appear at the top of the For You feed. Posts that score low never leave the author's followers' timelines, and even there they may get buried.
Not all engagement signals carry equal weight. X's open-source code reveals a clear hierarchy:
Tier 1 (highest weight): Reply, Retweet/Quote, and Favorite. These are the signals the algorithm builds distribution around. Replies carry the most weight because they require the most effort from users. A post that generates 50 genuine replies will outperform a post with 500 likes.
Tier 2 (structure-driven): Dwell time, profile clicks, shares, and link clicks. These signals are influenced by how content is structured. Long-form articles that maintain reading momentum generate high dwell time. Distinctive voice and specificity drive profile clicks.
Tier 3: Follows. A distinct point of view and consistent voice can trigger mid-read follows, but this signal carries less weight than direct engagement.
Negative signals: "Not interested," Block, Mute, and Report. These don't just reduce your score; they actively suppress future distribution. Content that triggers these signals (clickbait that doesn't deliver, obvious AI prose, aggressive selling) gets penalized across future posts, not just the offending one.
Most writing advice for X is reverse-engineered from what performed. Someone sees a viral post, copies the format, and calls it a strategy. The Heavy Ranker reveals the actual causal mechanism: why certain structures outperform, with what exact signal weight.
For example, the conventional advice to "write good hooks" is true but incomplete. The Heavy Ranker tells you that a good hook needs to deliver on its promise (to avoid "Not interested" signals) while generating enough curiosity to earn dwell time (Tier 2). And the ending matters just as much as the hook: it needs to generate replies (Tier 1, highest weight).
Understanding the Heavy Ranker means you can make deliberate structural decisions: end with an arguable statement to drive replies, include one screenshot-worthy line for retweets, and vary your pacing to sustain dwell time.
Write Better Articles structures every article around the Heavy Ranker's signal hierarchy. When you select a goal and niche, the tool selects the psychological trigger and hook pattern that optimize for Tier 1 signals in your specific context. The Algorithm Brief that accompanies every generated article explains exactly which signal the article is designed to maximize and why.
See the Heavy Ranker in action — generate an article and read the "Primary signal" field in your Algorithm Brief. It tells you which engagement signal your article targets and the structural reason behind that choice. Write my article →
The Heavy Ranker is the neural network (called the Phoenix transformer) at the core of X's recommendation algorithm. It scores every post across 19 engagement signals to determine which content gets promoted to the For You feed. Posts with higher predicted engagement receive broader distribution.
The Heavy Ranker calculates a weighted sum: Final Score = Sum(weight_i x P(action_i)) across 19 signals. It predicts the probability that a user will reply, retweet, like, spend time reading, click the author's profile, share, or take negative actions. Each signal type has a different weight, with replies carrying the highest.
You can optimize for the Heavy Ranker, but you can't game it. The algorithm rewards genuine engagement signals like thoughtful replies and sustained reading time. Clickbait that doesn't deliver, engagement bait ('like if you agree'), and spam formatting trigger negative signals that suppress distribution. The most effective approach is structuring content to naturally generate high-weight signals.
Write your own algorithm-optimized article
Write my article →For educational purposes only. AI-generated copy: always review before posting.