X has done what no other major social platform will: publish the actual code that determines what 600 million users see every day. The xai-org/x-algorithm repository went live in January 2026, fulfilling a promise Elon Musk made to update the open-source codebase every four weeks with developer notes. Whether this transparency is genuine or strategic is debatable. What's not debatable is that the code reveals exactly how the platform distributes content. Here are the ten things that actually matter.

1. The Algorithm Is Built on Grok

The ranking engine, called Phoenix, uses transformer architecture ported directly from xAI's Grok-1 language model. This replaces the older heuristic-based system. The model reads posts, watches videos, and predicts engagement probabilities based on semantic relationships between content, creator, and user. X claims to have eliminated every hand-engineered feature from the system. The transformer does the heavy lifting.

2. You're Competing Against 500 Million Posts Per Day

Every time you open the For You feed, the algorithm narrows roughly 500 million daily posts down to approximately 1,500 candidates, then ranks them in under 200 milliseconds. Candidates come from two sources: posts from accounts you follow (handled by a component called Thunder), and out-of-network content discovered by Phoenix's Two-Tower retrieval model. The split is roughly 50/50.

3. Replies Are Worth Far More Than Likes

The engagement weights in the code tell a clear story. A repost is worth approximately 20 times a like. A reply is worth roughly 13.5 times a like. A bookmark carries about 10 times the weight. A reply that generates a response from the author? That's cited at 150 times the value of a like. The algorithm is optimizing for conversation depth. Passive likes barely register.

Advertisement

4. Negative Signals Are Devastating

The asymmetry here is striking. According to one analysis of the code, a single block carries a weight of negative three compared to a like's positive half-point. Another estimate puts negative actions like blocks, mutes, and "not interested" clicks at approximately negative 74 times weight. One bad reaction can undo dozens of positive engagements. If your content triggers mutes or reports, your distribution collapses.

5. Premium Is a Structural Advantage, Not a Feature

A Buffer analysis of 18.8 million posts found that Premium accounts receive approximately ten times more reach than free accounts. Non-subscribers now see median reach below 100 impressions per post, while Premium accounts consistently hit 600 or more. Premium+ users often exceed 1,550 impressions. The code confirms that verification status functions as a ranking signal. Premium accounts receive a documented 4x visibility boost for in-network content and 2x for out-of-network reach. This is pay-to-play distribution, built into the feed architecture.

6. Links Get Buried

External links are algorithmically suppressed. X wants to keep users on-platform, and posts containing links receive near-zero distribution from non-Premium accounts. Even Premium users see links underperform compared to text and native video. The common workaround is posting the main content without a link, then dropping the URL in a reply.

7. The First Hour Decides Everything

Time decay is steep. A post loses roughly half its potential visibility score every six hours. The algorithm closely monitors engagement in the first 30 to 60 minutes. A post that gets 10 replies in 15 minutes dramatically outperforms one that accumulates the same engagement over 24 hours. Timing matters. You need your most engaged followers online when you publish.

8. Sentiment Now Shapes Distribution

Grok monitors the tone of every post. Positive and constructive messaging receives wider distribution. Combative or negative tones lead to reduced visibility even when engagement is high. The AI integration here is significant: the system is actively shaping the emotional tenor of the platform by rewarding certain kinds of speech.

Advertisement

9. Candidate Isolation Keeps Scoring Stable

One clever architectural choice: during ranking, candidate posts cannot "attend to" each other. Each post is scored based on user context alone, not on what else happens to be in the batch. This keeps scores consistent and cacheable, meaning a post won't randomly get a different score just because it was processed alongside something controversial. It also means timing strategies designed to avoid competition from high-performing posts don't work.

10. The Exact Weights Aren't Public

Here's the catch. The code reveals the structure, the pipeline, the transformer architecture. But the actual trained model weights and precise coefficient values are not included. We know replies matter more than likes. We don't know if the ratio is exactly 13.5x or something else. Thresholds, age cutoffs, and some service components are also excluded. This is transparency with limits.

What This Means

X has committed to updating the repository every four weeks. The AI-first approach reflects a broader industry shift toward learned systems over hand-tuned rules. Whether researchers and regulators find this level of disclosure sufficient remains to be seen. For users trying to build an audience, the message from the code is straightforward: spark conversations, post when your followers are active, avoid links in your main posts, and consider paying the $8. The algorithm rewards engagement depth over surface metrics.

The repository is available at github.com/xai-org/x-algorithm.