X (Twitter) officially open-sources the recommendation algorithm. This article provides a plain-language analysis of how Grok AI scores content through 14 interaction indicators, and summarizes six practical strategies for editors to break out of echo chambers and increase post exposure.
(Background: X releases the original code of the algorithm “Phoenix”! Feeding it with Grok, Musk chooses to keep model weights confidential)
(Additional context: The ultimate guide to earning money by writing articles on X: Musk launches X Articles with doubled revenue, audience targeting, fact-based content, reducing fluff, promoting subscriptions…)

Table of Contents

• Core concept of the algorithm: two pools
• How does the algorithm score posts?
  • 14 predicted behaviors of the algorithm
• Final score formula
• Six processing stages of the algorithm
• What content gets filtered out?
• Two key design principles
  • Fully AI-driven, no “manual tuning”
  • Author diversity mechanism
• Practical strategies for editors
  • Content strategy 1: Focus on “high-value interactions” rather than “surface metrics”
  • Content strategy 2: Increase “dwell time”
  • Content strategy 3: Avoid “negative signals”
  • Content strategy 4: Strive for “off-platform exposure”
  • Content strategy 5: Seize “timeliness”
  • Content strategy 6: Create content that “promotes following”
• Common misconceptions debunked
• Summary of the core logic of the algorithm

Today (20th), Musk fulfilled his promise by publicly uploading the X platform’s recommendation algorithm code “Phoenix” to GitHub. In this article, we will use the most straightforward explanation of X’s algorithm to help platform editors understand: How does X decide which posts can be seen by more people?

Core concept of the algorithm: two pools

First, understand that the “For You” feed on X is composed of content from two pools:

Pool 1: In-Network (Inside your network)

• Definition: Posts from accounts you follow
• Technical name: Thunder system
• Features: Real-time, low latency, almost immediately pushed after posting

Pool 2: Out-of-Network (Outside your network)

• Definition: Posts you do not follow but the algorithm thinks you will like
• Technical name: Phoenix system
• Features: Uses AI to mine relevant content from the global post library

💡 Key point: Your content will not only be seen by your followers but may also be recommended to “strangers who might like your content”! This is the key to breaking out of echo chambers.

How does the algorithm score posts?

This is the most critical part. X uses a Grok-based AI model (yes, that Grok from xAI) to predict the likelihood of each post receiving interactions.

Predicted behaviors of the algorithm

Final score formula

Final Score = Σ (Weight × Predicted Probability), in plain language:

• Higher predicted values for positive behaviors (likes, retweets, replies, shares) → higher score
• Higher predicted values for negative behaviors (block, mute, report) → significantly lower score

( Six processing stages of the algorithm

User opens X
↓
① Query user data (your interaction history, follow list)
↓
② Candidate content sources (from Thunder + Phoenix)
↓
③ Content enrichment** (adding post metadata, author info)
↓
④ Pre-filtering** (removing unsuitable content)
↓
⑤ AI scoring** (Grok model prediction + weight calculation)
↓
⑥ Sorting & selection** (showing highest-scoring content at the top)
↓
Presented to the user

) What content gets filtered out?

🚫 Pre-filtering (content removed before scoring)

🚫 Post-filtering (content filtered after scoring)

Two key design principles

Fully AI-driven, no “manual tuning”

We have completely eliminated all manually designed features and most heuristic algorithms from the system.

X claims to remove all manually crafted feature rules, relying entirely on Grok AI to learn your preferences from your interaction history.

What does this mean?

• No “best posting time” rules
• No “hashtag count” rules
• Everything is dynamically learned by AI based on actual interaction data

Author diversity mechanism

The algorithm has a built-in “Author Diversity Scorer” that reduces the weight of repeatedly appearing the same author.

Implication:

• Even if your post goes viral, it won’t dominate the entire feed
• Provides exposure opportunities for other creators
• Posting “small and frequent” might be more effective than “posting many at once”

Practical strategies for editors

Based on the above algorithm analysis, here are concrete actionable strategies:

Content Strategy 1: Focus on “high-value interactions” rather than “surface metrics”

Content Strategy 2: Increase “dwell time”

The algorithm tracks how long users stay on your posts.

How to do it?

• Write longer, in-depth content (with an engaging opening)
• Use threads to extend reading time
• Keep videos of moderate length so viewers watch till the end

Content Strategy 3: Avoid “negative signals”

Being blocked, muted, or reported will significantly lower your score.

How to avoid?

• Don’t spam (author diversity mechanism will penalize you)
• Avoid highly controversial or hateful content
• Don’t use misleading titles that disappoint (clickbait without substance)
• Don’t @ mention irrelevant users excessively

Content Strategy 4: Strive for “off-platform exposure”

This is key to breaking out of echo chambers. Phoenix will recommend your content to new audiences based on “similar user preferences.”

How to do it?

• Build expertise in specific fields (AI learns your content type)
• Encourage your current followers to actively interact (interaction data trains AI)
• Use clear topic tags to facilitate AI classification

Content Strategy 5: Seize “timeliness”

Thunder system is real-time; new posts are immediately pushed to followers.

How to do it?

• Observe when your audience is most active
• Post quickly during real-time events (AI learns the value of timely content)
• The first few hours after posting are most critical for interactions

Content Strategy 6: Create “follow-promoting” content

###follow_author### is a high-weight indicator!

How to do it?

• Series content: encourage followers to stay tuned for the next part
• Show unique viewpoints: make people feel “this person is worth following”
• Maintain a consistent posting rhythm: let followers know what to expect

( Common misconceptions debunked

❌ Misconception 1: “Posting at the best time”

Truth: The algorithm has no fixed “best time” rule; it dynamically determines timing based on your audience’s behavior.

❌ Misconception 2: “More hashtags are better”

Truth: The algorithm documentation does not mention hashtag weight; the key is whether the content can generate interactions.

❌ Misconception 3: “Buying followers/interactions can fool the algorithm”

Truth: AI predicts whether a user will have a positive interaction; fake followers won’t generate genuine engagement and may produce negative signals.

❌ Misconception 4: “Likes are the most important”

Truth: Retweets, quotes, replies, and new followers carry higher weight than likes.

) Summary of the core logic of the algorithm

The essence of X’s algorithm:
“Predict what content will generate positive interactions from users, then recommend that content.”

The editor’s core task:
“Create content that elicits genuine, high-value interactions.”

So, instead of thinking “how to trick the algorithm,” focus on “how to create content people genuinely want to interact with,” because the algorithm is designed to predict and reward authentic engagement.

![]###https://img-cdn.gateio.im/social/moments-a20a126e06-2b9392132a-8b7abd-e2c905###

(##