X Open-Sourced Its Algorithm: What the Phoenix Source Code Actually Says

Key Takeaways

• X open-sourced its recommendation algorithm in January 2026. The new system, called Phoenix, replaces every hand-engineered feature with a Grok-based transformer model that reads the actual content of every post.
• A single reply chain where someone responds and you respond back is weighted 150x more than a like. Conversations are the most powerful signal you can control.
• External links get a 30-50% reach penalty. Free accounts posting links have seen zero median engagement since March 2025. Links go in the replies, not the main post.
• Every account has a hidden TweepCred reputation score from 0-100. Below 65, only 3 of your tweets are even eligible for distribution. Most new or low-engagement accounts sit below this threshold without knowing it.

X Published Its Entire Recommendation Algorithm

On January 20, 2026, X published the source code for its recommendation algorithm on GitHub. The system is called Phoenix, and it’s built on the same transformer architecture that powers Grok. It replaces the legacy recommendation engine entirely.

This isn’t a whitepaper or a blog post about how the algorithm supposedly works. It’s the actual production code. TechCrunch reported that X committed to updating the public repository every 4 weeks with developer notes explaining what changed.

The architecture has four components: Home Mixer (orchestration), Thunder (in-memory post storage), Phoenix (the Grok-based ranking system), and Candidate Pipeline (infrastructure for processing posts through the ranking stages).

How Phoenix Decides What You See

Phoenix predicts the probability of 15 different engagement actions for every post, for every user. The positive actions get positive weights. The negative actions carry negative weights that actively suppress content.

The 15 signals: like, reply, repost, quote tweet, click, profile click, video view, photo expand, share, dwell time, follow author, “not interested,” block, mute, and report.

Getting blocked, muted, or reported doesn’t just lose you one follower. It tanks your post’s score for everyone. The negative signals carry real penalty weight.

The X Algorithm Engagement Weights

The engagement weights from the source code show a clear hierarchy. These are the confirmed values:

Two reply conversations where someone responds and you reply back. That’s worth 150 likes. Two of those conversations outweigh 150 likes. The algorithm has made its priorities clear: genuine conversation beats passive engagement by a massive margin.

How the “For You” Feed Gets Built

The system processes roughly 500 million tweets daily and makes about 5 billion ranking decisions. For each user, it narrows the pool to approximately 1,500 candidate posts through a three-stage pipeline.

Stage 1: Candidate sourcing. About 50% of your For You feed comes from accounts you follow (ranked by a Real Graph model that predicts engagement likelihood). The other 50% comes from accounts you don’t follow, surfaced primarily through SimClusters, a system of 145,000 topic clusters that organizes content by subject and social signals.

That second 50% is where small accounts break through. Phoenix uses a two-tower retrieval system that matches user interests to post content using embeddings. A 500-follower account posting focused content about a specific topic can appear in the For You feed of someone who has never heard of them, if the content matches that person’s demonstrated interests.

Stage 2: Neural network ranking. Phoenix’s Grok transformer model scores each candidate post across all 15 engagement dimensions. The final score combines the predicted probability of each action multiplied by its weight.

Stage 3: Filtering and diversity. Diversity mechanisms prevent any single account from dominating your feed. Blocks, mutes, and NSFW filters are applied. The feed is blended with ads, Spaces, trends, and recommendations.

Grok Now Reads the Actual Content of Your Posts

The old system relied on engagement patterns and social graph signals to determine relevance. Phoenix actually processes the semantic meaning of each post through Grok. It understands context, distinguishes between topics that share keywords, and categorizes content by subject matter.

This makes topic consistency critical. If your account posts about SEO one day, watches the next, poker the next, then crypto, you’re scattering your signal across multiple SimClusters. The algorithm can’t match you to a coherent audience because you don’t represent a coherent topic.

Grok also monitors tone. Positive, constructive messaging gets wider distribution. Negative, combative content gets reduced visibility even if engagement is high. Rage bait still generates replies (a high-weight signal), but the sentiment penalty can offset that advantage. The algorithm is trying to reward genuine discussion over angry engagement farming.

Link Suppression Is Worse Than You Think

External links in the main body of a post get a 30-50% reach reduction. For free (non-Premium) accounts, posting links has resulted in zero median engagement since March 2025.

This is one of the most aggressive content penalties on any major platform. If you’re posting links to your blog, product pages, or content in the main tweet, you’re cutting your reach by at least a third before anyone even sees it.

The workaround that most people use: post the content as a text-only tweet, then add the link in the first reply. The algorithmic penalty applies to the main post, not the replies. Some users also pin a reply with the link or mention it verbally (“link in replies”).

TweepCred: The Hidden Score That Gates Your Reach

Every X account has a TweepCred score from 0 to 100. It’s calculated using a PageRank-style approach factoring in account age, follower-to-following ratio, engagement quality, and who interacts with you.

The critical number is 65. Below that threshold, only 3 of your tweets are even considered for distribution. Above it, all tweets are eligible. This means a brand-new account with a poor ratio and no engagement history might not have its content evaluated at all, regardless of quality.

Premium subscribers get a +4 to +16 point boost to their TweepCred score. For accounts near the 65 threshold, that boost can be the difference between distribution and invisibility.

X doesn’t show you your TweepCred score directly, but you can estimate it based on the known factors from the source code. Use our free TweepCred calculator to see where your account stands and which factors are helping or hurting your distribution.

Content Format Performance

X diverges from every other major platform on content format preferences. Text-only posts generate 30% more engagement than video. Video gets 5.4% more than images. Images get 12% more than links. Links perform the worst by a wide margin.

This is the opposite of Instagram, TikTok, and LinkedIn, where visual and video content dominates. On X, writing wins. The character sweet spots from the data: 71-100 characters for quick engagement, 240-259 characters for maximum likes.

Optimal posting frequency: 2-3 quality posts daily with 30-60 minutes of spacing between them. High volume with low per-tweet engagement is penalized. The algorithm rewards quality over quantity.

The Premium Advantage in Numbers

The data from the source code confirms a significant Premium advantage:

At $8/month, Premium is one of the cheapest algorithmic advantages available on any platform. The 4x in-network boost alone justifies the cost for anyone using X as a business development or content distribution channel.

But the multipliers are applied to existing engagement potential. If your content isn’t generating engagement on its own, 4x of nothing is still nothing. Premium amplifies good content. It doesn’t save bad content.

What This Means for Building an Audience on X

The old playbook of “post content, get likes, grow followers” is algorithmically dead. Here’s what the source code tells us actually works:

Write for replies, not likes. One genuine conversation where both parties engage is worth 150 likes. Two of those conversations outweigh everything a viral tweet’s like count delivers. Ask questions. Make statements that invite disagreement or elaboration. Respond to everyone who replies.

Reply to people who reply to you. The author-response signal (+75 weight) is the single most powerful thing you can control. Every reply you leave on your own post’s thread is a 75x engagement signal. Not replying to your replies is leaving the strongest algorithmic lever unused.

Stay on topic. Grok is reading your content and placing you in SimClusters. Consistency in subject matter means better matching to the right audience. If you write about search and AI, keep writing about search and AI.

No links in the main post. Links go in the first reply. This is non-negotiable given the 30-50% penalty.

Post 2-3 times daily with spacing. The first 30-60 minutes of engagement velocity after posting determines distribution. Timing matters. Post when your audience is active, not when it’s convenient for you.

Get Premium if you’re serious. The 4x in-network and 2x out-of-network boosts, combined with the TweepCred score bump, make it the highest-ROI paid tool for content distribution on any social platform right now.

Frequently Asked Questions

Where can I read the actual source code?

The full algorithm is published at github.com/xai-org/x-algorithm. X committed to updating it every 4 weeks with developer notes. The code is Rust-based and covers all four components: Home Mixer, Thunder, Phoenix, and Candidate Pipeline.

How do I check my TweepCred score?

X doesn’t expose TweepCred scores to users directly. The score is calculated internally using account age, follower-to-following ratio, engagement quality, and interaction patterns. You can’t see your exact number, but you can infer it from your distribution patterns. If your posts consistently get near-zero impressions regardless of content quality, your TweepCred is likely below 65.

Does this algorithm apply to replies and quote tweets too?

The ranking weights apply to all post types in the For You feed. Replies within threads are scored differently since they’re displayed contextually. Quote tweets receive full algorithmic evaluation as standalone posts. The engagement weights (reply +13.5, repost +1, etc.) apply to the engagement your post receives, not the type of post you create.

Is rage bait still effective on X?

Partially. Combative content still generates replies (the second-highest weighted signal at +13.5). But Grok’s sentiment analysis now applies a distribution penalty to negative and combative tone. The penalty can offset the reply engagement benefit. Constructive disagreement performs better than hostile confrontation because it gets the reply weight without the sentiment penalty.