I Read X's Algorithm Code So You Don't Have To: What Creators Actually Need to Know

X has done something almost unheard of in social media: they've open-sourced the algorithm that powers your For You feed. After spending hours digging through the code repository, I'm breaking down what actually matters for creators trying to grow their reach on the platform.

The Core Truth: Engagement Predictions Drive Everything

The algorithm operates on a straightforward principle. It predicts how likely you are to engage with a post through likes, replies, reposts, bookmarks, and clicks. These predictions are then weighted differently, and that's where things get interesting for creators. The system assigns dramatically different values to different types of engagement, which means not all interactions are created equal.

What Actually Moves the Needle

Replies are worth 75 times more than likes. This is not an exaggeration buried in the code. When someone replies to your post and you respond back, the algorithm treats this as one of the strongest possible signals that your content is valuable. If you're posting content and ignoring the people who comment, you're actively strangling your own reach. The conversation matters far more than the applause.

Bookmarks carry a 50x multiplier. When someone saves your post for later, the algorithm interprets this as a powerful signal that you've created reference-worthy content. This makes sense when you think about user behavior. People bookmark posts they genuinely want to return to, which indicates real value rather than passive scrolling.

Watch time and dwell time determine visibility. The algorithm tracks whether people stop scrolling when they see your post. If they expand a thread to read more, watch your video, or click to see the full text, these are positive signals. If they scroll right past, your content gets deprioritized. This explains why video content and long-form threads often perform well. They naturally create stopping points that keep people engaged on the platform.

The Link Problem

Here's something that will frustrate many creators: outbound links severely damage your reach. The code reveals that external links function as a visibility killer because they signal an intent to move people off the platform. The algorithm is designed to keep users engaged within X, so any content that directs traffic elsewhere gets penalized.

The practical solution is straightforward. Put your links in your bio or use them in pinned replies rather than in the original post. This preserves your reach while still allowing you to direct interested users to external resources.

How Content Gets Categorized

The system uses something called SimClusters, which groups users and content into topical communities. If you've built an audience around technology and suddenly start posting extensively about cooking, the algorithm won't know where to distribute your content. Your existing audience may not engage because it's off-topic, and you haven't built credibility in the new space yet.

This doesn't mean you can never diversify your content, but it does mean that dramatic topic shifts will impact your distribution. The algorithm has learned to associate your account with specific subjects based on past engagement patterns, and it uses those associations to determine who should see your posts.

The Negative Signals That Tank Your Reach

While blocks and mutes don't appear with specific numerical weights in the public code, the system does predict the likelihood of these negative actions and weights them against you. If your content consistently generates blocks or mutes, the algorithm learns that you're producing something people actively want to avoid. This is different from simply being ignored, which is neutral. Negative actions are actively harmful to your distribution.

The lesson here is that polarizing content can work in your favor if it generates strong engagement from your target audience, but content that annoys or frustrates people will actively harm your reach over time.

The Grok-Powered Ranking System

What makes this algorithm different from previous social media recommendation systems is its reliance on transformer architecture borrowed from xAI's Grok language model. Rather than using hand-crafted rules about what makes content good, the system learns patterns from billions of user interactions.

The transformer evaluates your content alongside your engagement history and predicts multiple outcomes simultaneously. It's asking: will this person like this post, reply to it, share it, bookmark it, or take negative actions against it? The final ranking combines all these predictions with their respective weights.

Two Sources of Content

Your For You feed pulls from two distinct sources. In-network content comes from accounts you follow, while out-of-network content is discovered through machine learning predictions about what might interest you based on your engagement patterns. This dual approach prevents the algorithm from creating pure filter bubbles while still prioritizing content from accounts you've chosen to follow.

The out-of-network discovery system uses what's called a two-tower model. One tower encodes your interests and engagement history into a mathematical representation, while the other tower does the same for candidate posts. The system then finds posts whose representations are similar to yours, which is how you discover content from accounts you don't follow.

Author Signals Matter

The algorithm doesn't just evaluate individual posts in isolation. It considers the historical performance of the author and whether accounts you engage with also engage with this author. If you frequently interact with users who all follow and engage with a particular account, that account's content is more likely to appear in your feed, even if you don't follow them yet.

This creates a credibility system where consistently producing engaging content builds algorithmic trust that carries forward to future posts. Your past performance influences how aggressively the system distributes your new content.

Media and Freshness

Posts with images and videos receive better engagement predictions, which translates to better distribution. This makes intuitive sense because media-rich content naturally stops the scroll and encourages interaction. The algorithm has learned from billions of examples that posts with visual elements perform better on average.

Recency also plays a significant role. The system prioritizes newer content over older posts, which means timing matters. Posting when your audience is most active gives you a better chance of generating early engagement, which then signals to the algorithm that your content is worth showing to more people.

What's Missing From the Public Release

While X has released substantial portions of their algorithm, some critical components are notably absent. The actual trained model weights that determine how strongly different factors are weighted remain private. The full data pipelines that feed the system and the spam detection models are not included. The Phoenix retrieval system that discovers out-of-network content is referenced but not fully detailed.

This means that while we can understand the structure and general principles of the algorithm, we cannot reverse-engineer exact strategies or game the system with precision. The transparency is real but incomplete.

What you need to do to go viral

Do:

Create engaging content - Likes, replies, reposts, shares all boost you
Make longer videos - Videos above minimum duration get VQV boost
Encourage profile visits - Profile clicks are a positive signal
Get people to follow you - Follow actions boost your content
Inspire DM shares - Sharing via DM is a strong signal
Space out your posts - Avoid author diversity penalty
Build your follower base - In-network content has no OON penalty

Don't

Get muted/blocked - These heavily dock your score
Get reported - Reports have negative weight
Trigger "Not Interested" - Strong negative signal
Post controversial content - Safety filters remove it entirely
Spam posts rapidly - Each subsequent post gets discounted
Use muted keywords - Content is filtered for users who muted them

Practical Takeaways for Creators

The code reveals that success on X comes down to genuine engagement rather than vanity metrics. Focus on creating content that sparks conversations, not just passive consumption. Reply to your commenters within the first hour of posting, as early engagement signals are particularly valuable. Create content worth bookmarking by providing genuine value, whether that's insights, resources, or entertainment.

Stay consistent with your core topics so the algorithm knows how to categorize and distribute your content effectively. Use media to stop the scroll and encourage people to spend time with your posts. Time your posts when your audience is active to maximize early engagement.

Most importantly, understand that the algorithm is optimizing for one thing: keeping users engaged on X. Content that facilitates this goal will perform well. Content that drives people away or generates negative reactions will struggle. The system is sophisticated, but the underlying logic is remarkably simple.

Reference