The Twitter X Algorithm Explained - What the Open-Source Code Actually Reveals

One Number That Changes Everything

A like is worth almost nothing on X. Not metaphorically - mathematically. The open-sourced X algorithm scores a like at roughly 0.5 weight. A reply scores at 13.5. A reply where you, the author, write back? That two-way exchange jumps to 75. That means one good conversation thread is worth approximately 150 likes in algorithmic value.

This is not a theory. It comes directly from the GitHub repository X published when they open-sourced the algorithm. The code is public. The weights are public. And yet most accounts are still optimizing for likes.

If you have been wondering why some small accounts blow up while yours stagnates, this article is your answer. We are going to walk through the actual architecture of the X algorithm - the four-stage pipeline, the exact scoring formula, the negative signals that silently destroy reach, and the specific patterns that let nano accounts outperform accounts with ten times more followers.

No guessing. No the algorithm seems to prefer anything. Just what the code says.

The Open-Source Moment That Changed Everything

X announced the algorithm would go open source, and when the code dropped, the X Engineering account confirmed it: the new recommendation system is powered by the same Grok transformer architecture that drives xAI. The release committed to refreshes every four weeks with developer notes explaining each change.

This is unprecedented in social media history. No major platform has ever published the actual weights, the actual architecture, and the actual scoring formula for their recommendation system. Before this moment, every blog post about the X algorithm was educated speculation. Now it is engineering documentation.

The announcement post received over 16,000 likes and 41 million views - the most-viewed algorithm-related post in our dataset. The open-source repository hit over 1,600 GitHub stars within hours of going live. Code readers started publishing breakdowns almost immediately.

What they found surprised almost everyone - and invalidated a decade of conventional social media wisdom.

The Four-Stage Architecture - How Your Feed Is Actually Built

Every time you open X, the algorithm runs a four-stage pipeline to decide what you see. Here is how it works, directly from the source code architecture.

Stage 1 - Home Mixer (Orchestration Layer)

This is the entry point. When you open your feed, the Home Mixer receives your request and immediately begins hydrating your profile - pulling your engagement history (what you have liked, replied to, reposted, and clicked on recently), your following list, and your preference settings. This user data becomes the foundation for everything that follows. The system is building a real-time model of who you are and what you care about before a single post is selected.

Stage 2 - Candidate Sourcing via Thunder and Phoenix

The algorithm retrieves posts from two completely separate pipelines that run in parallel.

Thunder (In-Network): This module pulls posts from accounts you already follow. It is fast, deterministic, and high-priority. If you follow someone and they posted recently, Thunder finds it. The in-network candidates are selected based on your engagement patterns with each account - the people you reply to most frequently get priority over accounts you merely follow but never interact with.

Phoenix Retrieval (Out-of-Network): This is the more interesting and more powerful pipeline. Phoenix uses ML-based similarity search across the global content corpus to find posts from people you do not follow. It builds a vector embedding of your interests based on your engagement history and then runs approximate nearest neighbor search to find posts that match that interest profile. This is the mechanism that makes content go viral to new audiences - Phoenix is what puts a small account post in front of thousands of people who have never heard of them.

The resulting feed is roughly a 50-50 mix: half from accounts you follow, half discovered by Phoenix. That ratio matters enormously for growth strategy, which we will get to shortly.

Stage 3 - Hydration and Filtering

Before scoring, the system enriches each candidate post with additional metadata - author information, media entities, engagement counts - and runs it through a series of filters. Posts from accounts you have blocked or muted are removed. Content flagged as spam, NSFW, or violating platform rules is removed. Content you have already seen recently is deprioritized. The system also applies diversity constraints to prevent your feed from being dominated by a single topic or account.

Stage 4 - Scoring and Ranking via the Grok Transformer

This is where the math happens. The Phoenix system - built on Grok transformer architecture - takes each remaining candidate post and predicts the probability of approximately 20 different user actions. Each of those predictions gets multiplied by a weight. The weights are summed. The highest total score wins placement.

The formula looks like this: Final Score = P(repost) x weight + P(reply) x weight + P(quote) x weight + P(follow_author) x weight + P(dwell_time) x weight + P(video_view) x weight + P(not_interested) x negative_weight + P(report) x large_negative_weight, and so on across all predicted actions.

After ranking, an Author Diversity Scorer attenuates repeated posts from the same account within a single feed refresh - which has enormous implications for posting frequency strategy, covered below.

One critical design note from the GitHub README: the algorithm has eliminated every single hand-engineered feature and most heuristics from the system. The Grok-based transformer does all the heavy lifting by learning from your engagement history. This means there are no rules - only patterns. The system figures out what you want based entirely on what you have done.

The Engagement Weight Hierarchy - What Actually Moves the Needle

Here is the specific hierarchy that emerges from reading the source code and the analyses published by code readers after the open-source release. These are not approximations or guesses - they are derived from the weighted scorer implementation in the repository.

Tier 1 - The Highest-Value Signals

P(follow_author) - Will they follow you after seeing this post? This is one of the most powerful signals in the system. If a post consistently causes viewers to follow the author, it gets massive distribution. The algorithm is essentially trying to predict which posts are so good that they grow your account. Posts that trigger follows are treated as extremely high-quality content.

P(repost) - Will they retweet this? Reposts carry approximately 20 times the algorithmic weight of a like. When someone reposts your content, it signals that they found it valuable enough to put their own name on it. The algorithm treats this as a strong endorsement and dramatically expands distribution.

P(quote) - Will they quote tweet with added context? Quote tweets are weighted similarly to or higher than standard reposts because they generate a new piece of content that can accumulate its own engagement. Each quote tweet is a viral branch point - it introduces your original content to a completely new audience while generating fresh engagement signals.

P(reply) + P(author_reply) - The conversation multiplier: A direct reply to your post scores at approximately 13.5 times the weight of a like. But when you, the original author, reply back to that comment, the exchange jumps to 75 times the value of a like. This two-way conversation signal is the strongest quality indicator in the system. The algorithm is explicitly designed to reward posts that spark genuine back-and-forth discussion rather than one-way broadcasting. One good conversation thread where you actively participate can be worth more than 150 passive likes in algorithmic score.

P(dwell_time) - How long do they read it? The algorithm tracks exactly how long each user spends looking at each post. A post that people stop and read for 15 seconds scores significantly higher than a post people scroll past in 1 second. This is why threads that hook people and keep them expanding perform well - each expansion is a dwell signal. Substance beats skimmability.

Tier 2 - Strong Supporting Signals

P(video_view): Video content that gets watched past a minimum threshold generates strong signals. The algorithm distinguishes between accidental autoplay (low signal) and active watching (high signal). Native video uploaded directly to X outperforms embedded external video, which also avoids the link penalty covered below.

P(profile_click): When someone clicks on your profile after seeing a post, it signals genuine curiosity about you as a person or brand. This scores at approximately 12 times the weight of a like - considerably higher than most people assume. Profile visits are the algorithm way of detecting authority and interest.

P(bookmark): Bookmarks carry surprisingly high weight given that they are invisible to other users. The algorithm treats a bookmark as a strong high-intent relevance signal - the user wanted to save this for later, which implies genuine value. Building content that people want to reference repeatedly is a legitimate optimization strategy.

Tier 3 - The Weakest Positive Signal

P(favorite/like): Likes are explicitly the lowest-value positive signal in the system. They score at roughly 0.5 weight in the weighted scorer - meaning a single repost is worth the algorithmic equivalent of 40 likes. Likes are cheap, low-effort, and the algorithm knows it. Chasing like counts is one of the biggest misallocations of creative energy on the platform.

This finding was independently confirmed by code readers across the community after the open-source release. When you see a post with 500 likes but 5 replies, it is algorithmically underperforming compared to a post with 200 likes and 40 replies. The reply count is doing more for distribution than the like count.

Tier 4 - Negative Weights That Silently Kill Your Reach

This is the section most guides skip over, and it is arguably the most important. The algorithm does not just reward good content - it actively punishes content that generates negative signals. And the negative weights are dramatically larger than the positive ones.

P(not_interested): When a user clicks Not interested on your post, it tanks your score with that user instantly. Accumulated not interested signals across your audience destroy your baseline distribution. The algorithm interprets this as a sign that you are reaching the wrong people or producing irrelevant content.

P(mute_author): Getting muted is worse than getting unfollowed. A mute signal tells the algorithm that someone found your content annoying or low-quality enough to take action. Multiple mutes accumulate and suppress your reach broadly.

P(block_author): Block signals feed into toxicity and spam classifiers. Heavy block rates are a red flag that your content is generating strong negative reactions, not just disinterest.

P(report): This is the nuclear option. Report signals carry enormous negative weight - code analyses have put the report weight at approximately -369 compared to +0.5 for a like. A single report from a credible user can be an instant visibility killer. Getting reported repeatedly, even if your content is not ultimately removed, destroys your algorithmic score. Write content that generates strong opinions, yes - but not content that makes people want to report you.

The practical implication: it is better to have a smaller, highly engaged audience than a large audience that mostly scrolls past you and occasionally reports or mutes you. Audience quality beats audience size, and the algorithm is designed specifically to enforce this.

The 30-Minute Velocity Window - Your Post Lives or Dies Here

The algorithm does not evaluate your post over its entire lifetime equally. It applies aggressive time decay, and the first 30 minutes are weighted dramatically higher than any subsequent period. Here is how the distribution timeline typically unfolds.

Minutes 0-5: Your post is shown to a small seed group - roughly 5-10% of your most engaged followers plus a small out-of-network sample via Phoenix. The algorithm is testing your post with a controlled audience before committing to wider distribution.

Minutes 5-15: If that seed group engages strongly (especially with replies and reposts), the algorithm expands distribution to 20-30% of your followers and increases the out-of-network sample significantly. If engagement is weak, distribution stays contained.

Minutes 15-30: If the post is trending by the algorithm standards - meaning it is generating engagement velocity well above baseline for your account tier - it enters For You feeds at scale, including reaching people who have never seen your content. This is the Phoenix pipeline activating at full power.

The key insight from the code analysis: a tweet that gets 10 replies in the first 15 minutes will dramatically outperform a tweet that gets 10 replies spread over 24 hours. Velocity matters more than volume. Posting when your most engaged followers are online is not a nice-to-have - it is a core technical requirement for distribution.

After 7 days, posts are effectively filtered out of the recommendation pipeline entirely. The algorithm only surfaces content that is 7 days old or newer. Older content gets no distribution regardless of its engagement history. This is why evergreen strategies require consistent fresh output rather than relying on old posts to keep working.

There is one exception worth knowing: if a large account reposts your post, the algorithm can re-expand distribution even after the initial velocity window has passed. A late-stage boost from a high-follower account can restart the distribution cycle.

The Author Diversity Multiplier - Why Multi-Posting Kills Your Own Reach

One of the most counterintuitive findings from the open-sourced code is the Author Diversity Scorer. This mechanism specifically penalizes accounts that post too much within a single feed refresh cycle.

Here is how the decay works. Your first post appearing in a user feed refresh gets 100% of its earned score. Your second post appearing in the same refresh gets approximately 70% of its earned score - penalized by 30% simply because you already appeared once. Your third post gets roughly 50% of its earned score. This decay continues exponentially, meaning by your fourth or fifth post in a single refresh, you are barely registering.

The system is explicitly designed to prevent any single account from dominating a user feed regardless of how popular that account is. From a purely strategic standpoint, this means posting ten times in a day is significantly less efficient than posting three times at well-spaced intervals. The marginal value of each additional post within a refresh cycle approaches zero very quickly.

In our dataset analysis, this decay was visible in real account performance data. A prolific account second-most-liked post in a given period had over 95% fewer impressions than their top post - consistent with the exponential decay the code describes. More is not more on X. Better-spaced and better-crafted is more.

Follower Count Does Not Matter - And the Code Confirms It

The GitHub README is explicit: there is no bonus for big accounts. The algorithm does not care how many followers you have - it only cares whether the people who see your post actually engage with it. This is the single finding that most surprises people who come from Instagram or YouTube, where follower count directly correlates with baseline distribution.

On X, follower count affects distribution only indirectly - because more followers means a larger seed group in the initial velocity window, which gives you more potential engagement in those first 5-15 minutes. But a nano account with 2,000 highly engaged followers can outperform a mid-tier account with 200,000 ghost followers, because the engaged seed group generates higher velocity signals that trigger broader Phoenix distribution.

The data from analyzing engagement patterns across our dataset confirms this. Nano accounts (under 10,000 followers) averaged 5.85% engagement rate. Mid-tier accounts (100,000 to 1 million followers) averaged just 2.41%. Nano accounts had more than double the engagement rate. The algorithm is not suppressing small accounts - it is actively surfacing small accounts whose content generates genuine engagement velocity.

There is one nuance worth noting: X Premium (verified) accounts start with a +100 base score bonus compared to +55 for unverified accounts. But this only matters at the margins. A Premium badge does not overcome weak content or poor engagement velocity. The base score difference is small compared to the engagement weight differences. Fix the content first. Then add Premium as an amplifier of something that is already working.

The Follower Size vs. Engagement Rate Breakdown

Based on analysis of 392 tweets across the algorithm conversation, here is how engagement rate and reach actually distribute by account size:

Two things stand out immediately. First, nano accounts have more than double the engagement rate of mid-tier accounts. The algorithm is genuinely flatter than people assume - small accounts whose content connects get real distribution. Second, the reply-to-like ratio drops consistently as account size increases. Large accounts get likes; small accounts get conversations. Guess which signal the algorithm values more.

Mega accounts recover their engagement rate relative to mid-tier accounts because their sheer scale means even weak engagement signals produce large absolute numbers, and Phoenix surfaces their content to non-followers at a higher baseline rate due to brand recognition signals embedded in the interest graph.

External Links Are Penalized Heavily - Here Is the Workaround

Posts containing external links that take users off X see dramatically reduced distribution. The algorithm source code confirms that the Phoenix and Home Mixer pipelines deprioritize content that sends users away from the platform. Multiple code readers and platform analysts have documented reach reductions of 30-90% for posts containing external links, with severity depending on account tier, Premium status, and the link destination.

The practical workaround is straightforward and widely used: put your link in the first reply below your original post, not in the post itself. Post your hook, insight, or value proposition cleanly with no link, let the algorithm distribute it, then drop the link as your own reply. Interested readers will click through naturally. You keep the distribution without paying the link penalty.

Link preview cards (rich link previews) perform somewhat better than bare URLs, but still carry a penalty compared to no link at all. If you must put a link in the post itself, use the preview card format. If you can avoid it entirely, post the link in a reply instead.

The deeper reason for this penalty: X business model depends on time-on-platform. Content that keeps users on X is inherently more valuable to the platform than content that sends them elsewhere. The algorithm is enforcing that economic reality at the ranking level. This is not going to change.

The Grok Transformer - What AI-Powered Actually Means Here

When X says the algorithm is AI-powered, they are being more specific than most AI marketing language. The Phoenix system uses a transformer architecture ported directly from Grok - the same type of architecture that powers large language models. It is not a simple rules engine or a basic neural network.

The system operates in two stages within Phoenix. The Retrieval stage uses a two-tower architecture: one tower encodes your engagement history and interest profile into a vector embedding; the other tower encodes every post in the global corpus. Approximate nearest neighbor search then finds the posts whose embeddings are most similar to yours. This is how millions of candidate posts get narrowed down to hundreds in milliseconds.

The Ranking stage then takes those hundreds of candidates and runs them through the full transformer model, which simultaneously predicts probabilities across all 20 engagement actions. The design ensures that the score for each candidate is calculated independently - one post ranking does not affect another. This prevents gaming through manufactured context.

The system tracks your last approximately 128 engagements to build your preference vector. This means your feed is extremely responsive - six bad content decisions in a row (scrolling past things, clicking not interested) can meaningfully shift your feed within a day. It also means accounts that post in a consistent niche build stronger preference vectors in their audience over time. Niche consistency is not just an audience strategy - it is an algorithmic strategy.

The Grok model runs on 20,000+ GPUs in the Colossus data center. After the rebuild and open-source release, X reported time-spent on platform up approximately 20% and follow activity increasing even more. The transformer architecture is performing significantly better than the previous hand-engineered feature system it replaced.

What Most Guides Get Wrong About the X Algorithm

Now that the code is public, it is possible to audit popular advice against what the system actually does. Several conventional recommendations do not hold up.

Post as much as possible. False. The Author Diversity Scorer explicitly penalizes high-frequency posting within a refresh cycle. Post quality and spacing beat raw volume. Three well-crafted posts at optimal times will outperform ten mediocre posts scattered through the day.

Use hashtags for reach. Mostly irrelevant. Hashtags were a meaningful discovery mechanism in the pre-algorithmic era of chronological feeds. The transformer model reads and understands the actual content of your post - it does not rely on hashtags to categorize content. Hashtags may help in search, but they have minimal impact on For You feed distribution.

Engagement pods will boost your reach. Counterproductive. The algorithm detects artificial engagement patterns. Interactions from low-quality accounts (accounts with poor credibility scores, newly created accounts, accounts with suspicious engagement patterns) do not help your score and may actively hurt it. The system is looking for genuine engagement from real users with real histories. Manufactured velocity from pods can actually backfire by introducing low-quality engagement signals.

Going viral is about follower count. Wrong. The Phoenix pipeline exists specifically to surface content to people who do not follow you. About 50% of your For You feed consists of out-of-network content discovered through ML similarity search. A nano account with genuinely engaging content can reach millions of people through Phoenix - follower count is not a prerequisite for virality, it is merely one starting point for seed distribution.

Likes drive reach. The most damaging myth. With a weight of approximately 0.5 compared to 13.5 for replies and 20 for reposts, likes are the least efficient engagement signal on the platform. Content designed to provoke passive approval may accumulate likes without generating the reply and repost signals that actually drive distribution.

How to Work With the Algorithm - The Practical Framework

With the source code in hand, the strategic framework for growing on X becomes much more precise. Here is the hierarchy of what actually matters.

Build Conversations, Not Broadcasts

Every post should be designed with a conversational goal. Ask a question. State a counterintuitive position. Make a claim that your audience will want to push back on or expand. A post that generates 20 genuine replies - and where you reply to each one - is generating 75x weight per exchange from the algorithm perspective. That one conversation-rich post will do more for your algorithmic distribution than 200 posts that only generate likes.

Reply to every comment you get, especially early. The author reply signal is one of the highest-weighted positive signals in the system. Every time you respond to someone who commented on your post, you are generating a massive signal from that exchange. The algorithm is explicitly rewarding authors who participate in conversations rather than just broadcasting content.

Optimize Your First 15 Minutes

Your post fate is largely determined in the first quarter-hour. Post when your most engaged followers are online - not when you happen to be free. Use X Analytics to identify your peak engagement windows. Schedule your best content for those windows.

Consider priming engagement before you post. If you have a community of genuine fans, let them know something is dropping. The goal is to create real, organic engagement velocity in the seed window - not manufactured likes from pods, but genuine responses from people who actually care about your content.

Put Links in the First Reply

This is a simple tactical change with immediate impact. If you are driving traffic anywhere - your newsletter, your website, a YouTube video - post the main content without a link, then add the link in your first reply immediately after posting. This preserves your full algorithmic distribution while still giving interested readers a path to follow.

Post With Spacing, Not Volume

The Author Diversity Scorer decay function means that posting more than 3-4 times per day is largely wasteful. Your later posts in a day are competing with your earlier ones for the same audience feed refreshes, and they enter at dramatically discounted scores. Quality over quantity is not just advice - it is how the scoring math works.

Stay Niche-Consistent

Because the algorithm builds your interest profile from your last approximately 128 engagements, content that stays within a consistent niche builds stronger interest vectors in your audience over time. Every time you post off-topic content, you are potentially confusing your audience preference vectors and reducing the precision of Phoenix targeting when surfacing your posts to new users. Pick your lane and stay in it. Go deep, not wide.

Find What Is Already Working in Your Niche

The most reliable shortcut on X is understanding what content patterns already go viral in your niche - then adding your own angle to those patterns. This is not copying. It is pattern recognition. A framework that drives 10,000 reposts in the finance niche works because it is triggering something real in that audience. Finding those patterns, understanding why they work, and applying the same structural logic to your own original ideas is legitimate strategy.

Tools like SocialBoner are built specifically for this - a searchable database of millions of real viral tweets so you can find what is already working in your niche, see which small accounts went unexpectedly viral (the outliers the algorithm rewarded most), and apply 15 different AI-generated angles to riff on those patterns in your own voice. The algorithm rewards content that triggers the right engagement signals. The fastest way to learn which content triggers those signals is to study what already did.

What the Regular Algorithm Updates Mean for Your Strategy

X committed to refreshing the open-sourced algorithm every four weeks with developer notes explaining changes. This is significant for a few reasons.

First, it means the specific weights will shift. The 13.5x weight for replies versus 0.5x for likes may evolve. New engagement types may be added. Existing signals may be reweighted. Strategies optimized for today weights could become less effective as the weights change. This is an argument for building skills - creating genuinely conversation-worthy content, building a real engaged audience - rather than purely gaming specific metrics.

Second, the transparent update cycle means the X creator community will have much better information than it did pre-open-source. When weights change, code readers will notice immediately and publish the implications. Following the right accounts in the X-growth and creator economy space will keep you informed without requiring you to read the source code yourself.

Third, some signals are almost certainly permanent. The algorithm will always reward genuine engagement over passive scrolling. It will always penalize content that makes people click not interested or report. It will always favor conversations over broadcasts. These structural priorities are baked into X product philosophy and business model. Build your strategy around the durable signals, stay aware of the shifting weights.

The Nano Account Advantage Is Real - Here Is How to Use It

The data is consistent: small accounts with highly engaged audiences are getting outsized returns from the current algorithm. The engagement rate gap (5.85% for nano accounts versus 2.41% for mid-tier accounts) is not a fluke - it reflects how the algorithm engagement-velocity model rewards tight, loyal communities over large, diffuse followings.

If you are a nano account, your advantage is authenticity and connection. You can reply to every comment. You can know your followers by name. You can create content so specific to your niche that it generates intense reactions from the right people. That intensity - measured in replies, quote tweets, and reposts relative to your audience size - is exactly what the algorithm is looking for as a quality signal before committing to broader Phoenix distribution.

The playbook for nano accounts is not to act like a big account with fewer followers. It is to lean into the intimacy, generate genuine conversations, and let the algorithm surface your content to the audience it was built for. The Phoenix pipeline will do the expansion - your job is to prove your content is worth expanding.

Premium vs. Free - What the Actual Gap Is

X Premium (verified accounts) do get some algorithmic boost. The base score differential is approximately +100 for Premium versus +55 for unverified accounts. Premium accounts also have reduced link penalties and higher priority in the For You feed display.

But the gap is much smaller than the engagement signal gaps. A Premium account with a 0.5% reply rate is still going to be outperformed by a free account with a 5% reply rate. The conversation multipliers, repost weights, and velocity signals dwarf the base score differential from verification.

Premium is worth it if you are actively monetizing your presence and want every marginal distribution advantage. It is not worth much if your content strategy is not already generating strong engagement signals. Fix the content first. Then add Premium as an amplifier of something that is already working.

The Practical Checklist Before You Post

Based on everything the source code reveals, here is a pre-post checklist that aligns with actual algorithmic priorities.

Does this post invite a response? If someone reads it and has no natural reaction to share, reply to, or repost - fix it. Add a question. Sharpen the claim. Make it more specific or more counterintuitive.

Does it have a link in the body? Move the link to a reply. Post the main content clean.

Is this in my niche? Off-topic content fragments your interest vector and reduces targeting precision for future posts.

Am I posting this at a high-engagement time for my audience? If not, schedule it. The velocity window is too important to waste on low-traffic hours.

Am I prepared to reply to every comment in the first 30 minutes? The author reply multiplier is one of the highest-weight signals in the system. If you are not going to be present to engage, consider scheduling the post for when you can be.

Is this genuinely interesting, or am I just filling a posting slot? The algorithm tracks engagement history. A string of underperforming posts can depress your baseline distribution for your next several posts. It is better to post less and post better.

Try SocialBoner free for 7 days to find viral post patterns in your niche, get AI-generated angles that match your voice, and schedule your content for optimal engagement windows - all built around how the algorithm actually works, not how people guess it works.

The Bottom Line

The X algorithm is no longer a black box. The code is public, the weights are documented, and the architecture is mapped. What the source code reveals is a system fundamentally designed to reward genuine conversation and penalize passive accumulation of likes.

The accounts that will win on X going forward are not the ones with the most followers, the biggest budgets, or the most frequent posting schedules. They are the accounts that generate real conversations, reply actively to their audience, post content niche-consistent enough for the Phoenix pipeline to target it precisely, and avoid the negative signals that silently destroy distribution.

The algorithm is not against you. It is extremely good at detecting whether your content is actually connecting with people. Make content that actually connects, and the math works in your favor. That is the X algorithm explained - not as a mystery, but as a system you can understand and work with.