X Released The Algorithm Code. We Read It. Here Is Your 2026 Playbook. | BNC

For 20 years, the X algorithm was a black box. Last week, it stopped being one.

xAI published the complete source code for the For You feed ranking system. Not a blog post explaining it. Not a PR statement describing it. The actual Rust and Python codebase, Grok model weights, and the full scoring logic - every line open for anyone to [read on GitHub](https://github.com/xai-org/x-algorithm).

1,600 GitHub stars in six hours. The engineering world noticed. We went through it.

This is what changed, what died, and the exact playbook for founders who want to build on X without paying for reach.

---

### **Key Takeaways (The TL;DR)** * **Structurally Different:** The algorithm no longer runs on hardcoded rules; it runs on Grok-managed patterns. * **The 15 Signals:** Focus on Quotes, Thread-depth, and "Dwell Time" - Bookmarks are now invisible to the ranker. * **Author Diversity Scorer:** Posting more than 2-3 times per day now actively suppresses your total reach. * **Originality Filter:** Recycled templates and generic AI roundups are flagged and downranked automatically.

---

What You Are Actually Looking At The old algorithm was a patchwork of manual rules. It had hardcoded boosts and human decisions about what deserved reach. It was imprecise and gameable.

The new one is structurally different. Four primary components now run the feed:

1. **Thunder:** Retrieves posts from accounts you follow. 2. **Phoenix:** Discovers content from accounts you have never heard of using a "two-tower" neural model. 3. **Home Mixer:** Orchestrates the whole pipeline and enforces diversity rules. 4. **Candidate Pipeline:** Narrows 500 million daily posts to 1,500 candidates in under 200ms.

The ranking layer is **Grok**. The same transformer architecture powering xAI's language model now reads and scores every post against your specific interest graph.

The 15 Signals That Decide Your Reach Phoenix does not score "relevance." It predicts 15 separate probabilities for every post. Your final reach is a weighted sum of these actions.

### The Real Power Signals One real reply is worth more than many likes. Quotes, where someone adds commentary, carry significant information that a silent repost does not.

A post that converts a reader into a follower is directly rewarded by the ranker. This is not a downstream benefit; it is baked into your visibility score.

### What the "Gurus" Missed **Bookmarks are not in the 15 predicted actions.** They do not feed the ranker. While they are a nice vanity metric, the code shows they are invisible to the algorithm's distribution logic.

The Author Diversity Scorer There is a specific component in the codebase called the Author Diversity Scorer. Its job is to attenuate repeated scores from the same author.

If you post 10 times a day, your later posts cannibalize the reach of your earlier ones. Fewer posts, concentrated into higher-quality signals, outperforms high-volume spray-and-pray.

The math is simple: Posting more is now posting less.

Grok Now Reads Sentiment The January 2026 release confirmed that Grok monitors the tone of every post. Positive and constructive messaging gets wider distribution.

Negative and combative tones get reduced visibility even when engagement is high. The model distinguishes between engagement that signals interest versus engagement that signals conflict.

For founders, this is a tailwind. Builder energy and direct advice are exactly what the model rewards. > ### **Founder Strategy Session** > Want to discuss these algorithm shifts with other builders in real time? We have active X chat groups dedicated to founders and operators applying this exact playbook. > **[Join the BNC X Chat Groups here](/xchatgroups)**

What Just Died (Stop Doing These) Stop doing these today. The code confirms they no longer work in 2026:

* **Spam posting:** The Author Diversity Scorer will split your signal budget and lower your ceiling. * **Reply farming:** Replies are now weighted by *who* replies, not how many. * **Text-only posts:** Media receives a higher signal weight because it triggers more action types (photo expand, dwell, video view). * **Recycled viral templates:** The new Grox classifier layer flags low-originality content. It learned what "copied" looks like. * **Engagement bait:** Appendages like "What do you think?" are now flagged as a liability by the content classifier.

What Just Won (The 2026 Play) These are the behaviors the algorithm has been trained to reward:

* **Discovery for Small Accounts:** Phoenix surfaces content based on interest matching, not follower count. 300 followers can reach 80k people. * **Text + Media Combos:** This expands the range of positive signals you can generate in a single post. * **Thread Narrative Arcs:** Grok reads the full thread context. It scores the arc: setup, friction, resolution. * **Long-form (4,000 chars):** Long-form breakdown receives heavier weight because it generates significantly more "Dwell Time."

Formats That Work Right Now Based on the signal structure of the algorithm, these formats are built to generate the widest range of positive predictions. * Tactical playbooks: Hook plus 5 to 8 numbered steps plus a closer that drives to profile or follow. They generate reply, quote, dwell, profile click, and follow signals. One post, multiple action types. * Personal proof posts: "$X to $Y in Z weeks - here is the breakdown." Specific numbers, specific timeline, specific mechanism. The algorithm rewards posts that convert readers into followers. Nothing converts like specific proof. * Contrarian takes with screenshot evidence: One strong opinion, backed by a screenshot or data. The model learned that contrarian content with proof generates high quote engagement. Quote is worth more than repost in the scoring formula. * Threads with arc structure: Setup (the problem), friction (why the obvious answer is wrong), resolution (what actually works). Three parts. Keep it tight. The model reads the whole thread. * Image carousels: 3 to 7 slides, one bold claim per slide. Each slide drives dwell and photo expand signals. Build them to reward the tap. * Short video under 90 seconds: Real work, not promotion. Showing a process, not selling a result. Under 90 seconds. The model predicts video view completion. Do not waste the first 3 seconds. * Long-form 4,000-character breakdowns: Pick one trending topic, go deep. Dwell time is a standalone signal. The longer someone reads, the better your score. This format earns it.

Tone of Voice That Wins * First-person specific over abstract observation: "I built this" over "founders should consider." The algorithm rewards follow conversion. Specific first-person proof converts. Abstract advice does not. * Concrete numbers and names over vague claims: "17 clients in 6 weeks" over "rapid growth." The model has been trained on human engagement patterns. Humans engage more with specifics. * Builder energy over motivational filler: "Here is what I shipped" over "here is what is possible." The new Grok sentiment layer rewards constructive and specific content. Motivational generics get flagged as low signal. * One strong opinion per post: Not three hedged ones. The algorithm rewards posts that drive quote engagement. You cannot quote a hedge. You can quote a position. * Direct address to the reader: "You" over "founders" or "operators." Dwell time increases when the reader feels addressed. Passive third-person creates psychological distance.

The Play Starting Today Stop everything else. Do these six things. 1. Cut to 2 posts per day maximum: The Author Diversity Scorer is real. Two high-quality posts beat eight average ones every time. 2. Always pair text with image, video, or thread: Text alone leaves signal on the table. Even a single relevant screenshot is enough to add the photo expand signal. 3. Respond to every comment in the first 30 minutes: Set an alarm. Show up. This extends your scoring window and drives conversation depth. It is the highest-leverage single action you can take after publishing. 4. Write in first-person specific: Your experience, your numbers, your outcome. Not general advice. Not industry observation. You. 5. One bold opinion per post, backed by proof: The algorithm rewards quote engagement. Quote engagement requires a position people want to respond to. State yours. 6. Let out-of-network discovery do the heavy lifting: Phoenix is built to find your best content and surface it to people who have never heard of you. Trust the retrieval system. Post for quality, not volume.

One More Thing: The Algorithm Updates Every 4 Weeks X committed to releasing updated code with developer notes every 4 weeks.

This is the first time in the history of the platform that creators have had a publicly maintained window into what the algorithm is actually doing. Every update will tell you what changed.

Cycles repeat approximately every 6 months. The playbook above is built for the current release. When the next one drops, we will read it and update you.

Where to Apply This The tactics above work, but they produce spikes, not systems. To build sustainable revenue, you need a strategic framework.

We have compiled a complete **Social Media & Positioning Strategy Brief**. It covers revenue conversion architecture and content pillar design.

[Access the full strategy document here](/strategy)

If you want to work through these algorithm changes with other founders in real time, join our **X chat groups**: [businessnetworking.club/xchatgroups](/xchatgroups)

> ### **Next-Step Intelligence** > Applying these tactics alone is a start. Applying them with a curated group of peers is how you build a moat. Join our exclusive X chat groups for founders. > **[Get Access to X Chat Groups](/xchatgroups)**

Algorithm code is public now. The advantage goes to whoever acts on it first.

--- **Business Networking Club** | businessnetworking.club | *Join founders, entrepreneurs and operators done with working alone.*