Why Some X Posts Go Viral
A post going viral on X isn't controlled by one simple metric, such as likes, reposts, or follower count. The open-source X recommendation architecture shows a multi-stage system: posts are discovered as candidates, evaluated for eligibility, scored using predicted user actions, ranked against other candidates, and then selected for users' For You feeds. (X Algorithm)
The important insight is: A post can go viral when the recommendation system repeatedly finds it relevant and valuable for new audiences.
The Viral Reach Loop
A simplified version of the process looks like this:
This creates the possibility of a distribution loop. A post doesn't need a huge audience at the beginning if the recommendation system continues finding new users for whom the content is relevant.
1. It Starts With Candidate Retrieval
Before X can recommend a post, the post has to become a candidate. The current open-source architecture has different sources for finding candidates, including in-network and out-of-network content. (X Algorithm repository)
This creates two broad paths:
In-network candidates
Out-of-network retrieval
The second path is particularly important for viral reach. A post can potentially be discovered by people who have never followed the creator.
2. Out-of-Network Recommendations Create New Reach
X's Phoenix Retrieval system is designed to discover potentially relevant posts beyond a user's existing network. The open-source Phoenix architecture uses user and candidate representations and approximate nearest-neighbor retrieval to find potentially relevant candidates. (Phoenix documentation)
This is one of the fundamental mechanisms that makes large-scale discovery possible. Without out-of-network recommendations, a post would be much more dependent on the creator's existing follower base.
3. Relevance Determines Who Might See It
Once a candidate is available, the system needs to determine whether it is potentially relevant to a particular viewer. Phoenix uses user context and engagement history when evaluating candidates. The current architecture describes the model as using engagement sequences to predict future engagement. (Phoenix architecture)
Another user with a completely different history may receive a different prediction. So: Virality is not necessarily universal relevance. A post can be highly relevant to one audience and irrelevant to another.
4. X Predicts What the Viewer Might Do
This is one of the most important parts of the current architecture. Phoenix doesn't simply ask: “How many likes does this post have?” It predicts multiple possible actions for the viewer.
The current action set includes predictions for like/favorite, reply, repost, quote, click, profile click, share, dwell, follow, video-related engagement, and negative feedback. (Phoenix runner)
These values are predictions—not actual numbers shown to creators.
5. Multiple Positive Signals Can Work Together
A post may have several potentially valuable predicted actions (e.g., High P(Like), High P(Reply), High P(Repost), High P(Dwell)). The ranking system combines predicted actions using weights. The current X architecture describes this as a weighted combination of predicted engagement probabilities. (X Algorithm)
This is much more sophisticated than Likes × 1 + Reposts × 10 + Replies × 5.
6. Virality Isn't the Same as Having Lots of Likes
This distinction is critical. A post can already have 10,000 likes, but that doesn't automatically mean it will be recommended to everyone. The recommendation system is evaluating candidates for particular users. The important question becomes: How likely is this user to engage with this candidate? This explains why a smaller post can sometimes outperform a much larger post for a particular audience.
7. Reposts Can Help Expand the Audience
Reposts are explicitly modeled as an action in Phoenix. (Phoenix runner) A repost can expose the content to another network of users. But avoid saying: “A repost automatically gives your post more reach.” The open-source architecture does not establish a fixed reach multiplier for reposts. Instead, repost behavior is one of several predicted actions contributing to candidate ranking.
8. Replies Can Indicate Conversation
Replies are also separately predicted. A post that creates genuine conversation may therefore generate several useful signals (Post → Reply → Conversation → Additional users encounter content → More potential interactions). However, this does not mean that X has a documented rule: “More replies = automatic viral boost.” The public architecture supports the idea that replies are modeled as a positive action—not that there is a simple reply-to-impression conversion rate.
9. Dwell Adds an Attention Dimension
The current Phoenix architecture also models dwell and dwell time. (Phoenix runner) This matters because users don't have to press a button to demonstrate interest. A user can see a post, stop scrolling, read, and spend time. That behavior can provide information about the usefulness or relevance of the content. So a post can generate recommendation value without accumulating huge numbers of visible likes.
10. Negative Feedback Can Work Against Recommendation
Virality isn't just about positive engagement. The current Phoenix action set also includes negative predictions such as Not Interested, Block author, Mute author, and Report. (Phoenix runner) The ranking architecture applies negative weights to these signals. (Ranking scorer) Therefore, engagement isn't automatically positive: a post can generate interaction while also producing negative recommendation signals.
11. The Post Must Compete With Other Candidates
Even if a post has strong predicted engagement, it doesn't exist in isolation. X has many candidate posts competing for recommendation opportunities. The current architecture includes scoring, author diversity, out-of-network scoring and selection stages. (X Algorithm repository) The highest-ranked candidates have a better chance of selection, but actual feed construction includes additional constraints and adjustments.
12. Author Diversity Can Affect Distribution
Imagine five highly ranked candidates all come from the same creator. A useful feed may not want Creator A repeated five times consecutively. The current architecture includes an Author Diversity Scorer that adjusts repeated-author candidates. (X Algorithm repository) This means a strong individual post score doesn't necessarily mean every post from that creator will dominate the feed.
13. A Viral Post Can Reach Multiple Audiences
One reason a post can become very large is that recommendation isn't necessarily limited to one initial audience (e.g., Initial audience → AI-interested users → Developers → Startup audience → Technology audience → General audience). If the post remains relevant to multiple groups, recommendation can continue expanding. This is a conceptual explanation of audience expansion—not a documented X “viral ladder.”
14. Viral Reach Is Better Understood as Repeated Recommendation
A useful mental model is: Virality isn't one giant boost. It can be repeated successful recommendation across many users. Each recommendation opportunity is evaluated in context, producing large-scale distribution without requiring the creator to have an enormous follower count.
15. Why Some Posts Stop Going Viral
A post that initially performs strongly can eventually stop receiving recommendations. That doesn't automatically mean X “killed” it. Possible reasons include new competing candidates, different viewer relevance, previously seen content, candidate filtering, feed selection, author diversity, and changing audience composition. (X Algorithm repository) So: Early viral growth ≠ permanent distribution.
16. Followers Are Not the Ceiling
A creator with 5,000 followers can potentially reach far more than 5,000 people through recommendations because the architecture supports out-of-network retrieval. The system can discover posts for users who don't follow the author when the content appears potentially relevant. (Phoenix documentation)
17. Viral Doesn't Mean “Highest Engagement”
Suppose Post A has 20,000 likes but low predicted relevance for a particular user, while Post B has 2,000 likes and high predicted relevance for that user. Post B could still be the better candidate for that particular viewer. The ranking model is not simply sorting the entire platform by total likes; it is predicting user behavior.
18. There Is No Public “Viral Score”
This is particularly important for creators. The open-source architecture does not give us a creator-facing number such as Viral Score = 92/100, nor does it establish 100 likes → viral or 1 repost → 20 additional impressions. (X Algorithm repository) So any tool claiming to know an exact internal “viral score” should clearly label it as an estimate, not an official X score.
19. What Creators Can Actually Learn From This
The open-source architecture suggests several practical principles:
- Make the audience obvious: The clearer the topic, the easier it is for content to be relevant to a particular audience.
- Create reasons to engage: Likes, replies, reposts, quotes, clicks, and shares are all modeled actions.
- Earn attention: Dwell-related behavior is also modeled.
- Create content worth sharing: Sharing and reposting are distinct recommendation signals.
- Avoid unnecessary negative feedback: Not Interested, mute, block, and report are modeled negative actions.
- Don't obsess over one metric: The system combines multiple predicted behaviors.
20. What We Can Verify vs. What We Cannot
Supported by Open-Source Architecture
- ✓ Out-of-network recommendations exist.
- ✓ Phoenix uses user context and engagement history.
- ✓ Multiple engagement actions are predicted.
- ✓ Positive and negative actions are modeled.
- ✓ Predicted actions contribute to ranking.
- ✓ Candidate selection happens after ranking.
- ✓ Author diversity and other adjustments exist.
Not Established by Public Code
- ✗ “X gives viral posts a manual boost.”
- ✗ “A post is tested for exactly 15 minutes.”
- ✗ “100 likes guarantees recommendation.”
- ✗ “One repost gives exactly X impressions.”
- ✗ “A specific engagement rate triggers virality.”
- ✗ “Every post goes through the same audience tiers.”
The Viral Post Model
That last step is the important one. If the post continues to be relevant to new groups of users, it can continue receiving recommendation opportunities.
Under The Hood Takeaway
Some X posts go viral because the recommendation system can repeatedly discover them as relevant candidates for new users beyond the creator's followers. The current open-source architecture shows a process involving candidate retrieval, user-specific prediction, weighted scoring, ranking, diversity, selection and filtering. (X Algorithm)
So virality is better understood as: Repeated successful recommendation across increasingly large or diverse audiences—not a single viral switch or a fixed engagement threshold. And importantly, the open-source code gives us a framework for understanding how recommendation can happen, but it does not provide a guaranteed formula for predicting exactly when a post will go viral.
Official Open-Source References
- • X / xAI — For You Algorithm repository — current architecture, Phoenix, Thunder, filtering, scoring and candidate pipeline. (GitHub)