X, formerly known as Twitter, no longer relies on a simple chronological timeline to decide what users see in the For You feed. In 2026, its recommendation system is a large machine learning system that looks at a user’s activity, retrieves potentially relevant posts from both followed and non-followed accounts, predicts how the user may interact with those posts, and then ranks the results before presenting the final feed.

The important change in 2026 is that much more of the current recommendation architecture is publicly visible. X and xAI have released the code for the newer X For You recommendation algorithm, including its retrieval and ranking components. The current public repository describes a system that combines posts from accounts a user follows with posts discovered from across X, then uses a Grok-based transformer model called Phoenix to score candidates.

This makes it possible to understand the modern X recommendation system much more clearly than simply repeating common social media advice about likes, hashtags, or posting times.

What Does the X Algorithm Actually Do?

The word “algorithm” can make X seem like one single formula that decides whether a post will become popular. In reality, the recommendation system is a collection of services, data systems, machine learning models, filters, and ranking processes working together.

The current public X For You system is built around a pipeline. It first understands information about the user, finds posts that could be relevant, enriches those posts with additional information, removes unsuitable or already-seen content, predicts different ways the user might interact with each remaining post, and finally ranks the candidates to construct the feed.

This means that two users can open X at almost exactly the same time and see very different posts. The reason is not necessarily that one post has a universally higher score than another. The system is trying to estimate which content is most relevant to that particular user.

The Two Main Sources of Posts

One of the most important parts of the current system is that the For You feed does not only depend on accounts a person follows.

The publicly released architecture describes two major candidate sources. The first is in-network content, which consists primarily of recent posts from accounts the user follows. The second is out-of-network content, which comes from posts discovered through machine learning retrieval across the broader X content pool. The current architecture identifies these components as Thunder for in-network retrieval and Phoenix Retrieval for out-of-network discovery.

This is a major reason why a relatively small account can reach people who do not follow it. A post does not have to remain inside the audience of the author's followers. When the recommendation system identifies a post as potentially relevant to users outside that network, it can become a candidate for those users' feeds.

The system therefore creates an environment where following relationships still matter, but they are no longer the entire distribution mechanism.

What Happens Before a Post Is Ranked?

Before the ranking model decides where a post belongs, X needs to create a pool of possible posts.

For followed accounts, the system can retrieve recent posts through the in-network pipeline. For content outside the user's network, the retrieval system searches for posts that appear relevant based on the user's interests and previous interactions.

The candidate posts are then enriched with additional information. According to the current public repository, this can include the post's text and media, information about the author, verification status, video duration, subscription status, engagement information, language, media type, mutual-follow information, and other contextual signals.

At this stage, X is not yet deciding the final order of the feed. It is preparing the information that later stages will use to make that decision.

Filtering Comes Before Final Ranking

Not every retrieved post is allowed to reach the final ranking stage.

The current public algorithm includes pre-scoring filters designed to remove unsuitable candidates. The repository describes filtering for duplicates, excessively old posts, the viewer's own posts, content from blocked or muted accounts, muted keywords, previously seen or recently served posts, and content that is otherwise ineligible.

This matters because an excellent post can still fail to appear in a user's For You feed when it is filtered for another reason.

The recommendation system is therefore not simply asking, “Which post has the highest engagement?” It is also asking whether the post should be eligible for this specific viewer in the first place.

How X Predicts What You Will Do

The ranking stage is where the current system becomes especially interesting.

The public 2026 architecture describes Phoenix as a Grok-based transformer model used to score candidate posts. Rather than relying on one simple popularity number, the system predicts the probability of multiple user actions. Those predictions are then combined into a final ranking score.

This distinction is important.

A like is one type of interaction. A reply is another. A repost is another. Opening a post, clicking into content, watching a video, or spending time with a piece of content can represent different forms of user interest.

The model can therefore treat several possible reactions as separate signals instead of reducing every interaction to a single generic engagement number.

The publicly released architecture says the system uses the user's engagement history, including actions such as likes, replies, and shares, to determine relevance. It also explicitly states that the newer system has moved away from traditional hand-engineered relevance features and places much more of the work inside the transformer model.

Why Engagement Still Matters

The fact that X uses a large machine learning model does not mean engagement has become irrelevant.

Engagement remains important because engagement is evidence about how people respond to content. When users repeatedly interact with certain subjects, accounts, formats, or communities, those actions provide information about their interests.

However, the important distinction is that the system is not simply counting likes and sending the most-liked posts to everyone.

A post with fewer total likes may be highly relevant to one particular audience and therefore receive strong recommendation opportunities for that audience. Conversely, a post with enormous overall engagement may be less relevant to someone whose interests are completely different.

This is why a post's total engagement should not be interpreted as its only source of distribution.

The Importance of User History

The For You feed is personalized around the person viewing it.

The current system hydrates the recommendation process with information about the user's recent engagement history and account context. The public repository also references followed topics, starter packs, mutual-follow information, previously served content, and other user-specific signals used during recommendation processing.

Over time, these interactions create a picture of what a user is likely to find interesting.

For example, someone who regularly reads technology discussions, watches artificial intelligence videos, follows developers, and replies to programming posts can gradually receive more content from those areas. Another user interacting heavily with finance, football, gaming, or news may build a completely different recommendation environment.

The algorithm is therefore continuously learning from the relationship between the user and the content.

Why Replies Can Be Important

Replies are more than a visible number underneath a post.

A reply represents an active interaction. It indicates that the viewer did more than simply passively see the post. Because the current recommendation architecture predicts multiple engagement actions, interactions such as replies can form part of the model's understanding of whether a post is likely to be useful or interesting to a particular viewer.

However, this does not mean that every reply automatically boosts a post in the same way.

The system evaluates probabilities and context rather than following a simplistic rule such as “one reply equals a fixed number of points.” The effect of an interaction depends on the broader recommendation model and the viewer for whom the prediction is being made.

Why Reposts and Shares Matter

Reposting is a stronger form of active distribution than simply viewing a post.

When users repost something, they effectively tell the platform that the content was important enough to pass along to their own audience. Such actions can provide valuable information to the recommendation system.

The important point is that the system is not merely looking at whether a post has been reposted. It is trying to understand the likelihood of future interactions from the people who may see that post.

This is a fundamental difference between a modern recommendation system and an old-style popularity ranking.

What About Likes?

Likes remain a useful signal because they are explicit expressions of user preference.

When a person consistently likes particular subjects, the recommendation system can learn from that history. X itself has long explained recommendations in terms of signals generated by users' actions on the platform.

Nevertheless, a like is only one type of signal among many. The current public architecture is designed around multi-action prediction rather than a single like-based popularity score.

This means that creating content solely to collect likes is not necessarily equivalent to creating content that the recommendation model considers broadly valuable.

Why Time Spent on Content Can Matter

Modern recommendation systems increasingly care about what users actually do with content rather than only counting visible engagement.

The current X architecture includes models that predict several possible user actions. X's older publicly released recommendation architecture also exposed explicit and implicit user signals, including interactions such as profile visits and post clicks.

This helps explain why a post can sometimes continue receiving distribution even without an unusually large number of public likes.

A user can open a post, read it, click an attached link, watch a video, visit the author, or interact in another meaningful way. Different actions provide different information about whether the content was relevant.

Why Small Accounts Can Be Recommended

A common misunderstanding is that only large accounts have a realistic chance of reaching large audiences.

That is not how an out-of-network recommendation system has to work.

Because X can retrieve posts outside the user's existing network and score them against the user's interests, the system has a mechanism for discovering content from accounts a person does not follow. The current public architecture explicitly describes a global out-of-network retrieval system followed by machine learning ranking.

As a result, follower count is not equivalent to maximum possible reach.

An account with a small audience can publish a highly relevant post, enter recommendation candidate pools, and potentially be shown to users outside its existing follower network.

How Content Relevance Is Understood

Older recommendation systems often depended heavily on manually designed features. The current X repository describes a different approach.

The Phoenix transformer is designed to understand user engagement sequences and content relevance through machine learning. The repository specifically states that the newer system eliminated hand-engineered relevance features and uses the Grok-based transformer to perform the major relevance work.

That means the model can consider relationships between the user, previous interactions, and candidate content without relying entirely on a fixed human-written scoring formula.

This is one of the most important developments in understanding X's algorithm in 2026.

Does Posting Time Still Matter?

Posting time can matter indirectly because newer posts have an opportunity to enter the candidate pool while they are fresh.

However, it is not accurate to describe X in 2026 as a simple chronological system where posting at a particular minute guarantees reach.

The current recommendation process retrieves candidates and then evaluates them. The architecture also contains filtering for posts that are too old, showing that freshness remains relevant while not being the only ranking factor.

For this reason, posting when the intended audience is active can still be sensible, but content relevance and predicted engagement are far more important than following an arbitrary “best time” rule.

Do Hashtags Control the Algorithm?

Hashtags can help communicate the subject of a post, but they should not be treated as a secret mechanism that forces the For You feed to distribute content.

The current recommendation architecture is substantially more sophisticated than a hashtag-matching system. It uses machine learning retrieval and ranking, user history, content information, engagement signals, and filtering.

A relevant hashtag can improve discoverability in contexts where hashtags are useful, but simply filling a post with many unrelated hashtags does not turn it into a high-quality recommendation candidate.

Does Verified Status Guarantee Reach?

No.

The current public architecture includes verification status as one piece of candidate information, but verification is not described as a universal rule that automatically places content at the top of the For You feed.

A verified account still has to produce content that the recommendation system considers relevant to the viewer.

The same principle applies to account size. Large accounts have major advantages in terms of existing audiences and accumulated interaction, but the recommendation system is still designed around the expected relevance of content to individual users.

Why Some Posts Suddenly Stop Getting Views

A drop in impressions does not necessarily mean that an account has been permanently restricted.

Recommendation systems operate continuously. Candidate selection, user interests, competition from other content, freshness, filtering, and predicted engagement all affect what gets shown.

A post may receive an initial audience and then stop expanding because the system's predictions change after observing how people respond to it. Another post may continue receiving distribution because it remains relevant to additional users.

The important thing is that a post's early performance is not necessarily a permanent verdict on the account.

X Does Not Use One Single Feed Score

One of the most misleading explanations of social media algorithms is the idea that every post receives one fixed score such as “likes multiplied by five plus replies multiplied by ten.”

The current X architecture does not work that simply.

The system retrieves candidates, filters them, passes them through machine learning scoring, predicts multiple interactions, combines those predictions into a ranking score, and then constructs the feed.

This means that a post can have different effective value for different users.

The recommendation system is essentially trying to answer a question similar to this: Which of these available posts is this particular person most likely to find relevant and engage with right now?

Why X Open-Sourced Its Algorithm

X has made significant portions of its recommendation architecture available publicly through GitHub.

The original X recommendation repository includes services for areas such as the For You timeline, recommendation infrastructure, user signals, post representations, candidate retrieval, and ranking.

The newer xAI repository goes further for the current For You architecture by publishing the Phoenix-based retrieval and ranking system and associated pipeline components. The May 15, 2026 update added an end-to-end inference pipeline, content-understanding components, additional candidate sources, and systems for areas including ads and user-context hydration.

This makes 2026 particularly important for anyone trying to understand how X recommendations work.

The 2026 Architecture in Simple Terms

The entire process can be understood without knowing the underlying programming language.

First, X identifies the user and gathers relevant context about that user's recent activity and relationships.

Next, the platform retrieves possible posts from accounts the user follows and discovers additional posts from outside the user's network.

The system then enriches those posts with information about their content, author, media, engagement, and other properties.

Posts that should not be considered are filtered out.

The remaining candidates are passed to the machine learning ranking system, where Phoenix predicts different ways in which the viewer may interact with each candidate.

Those predictions are combined into a final ranking, and the highest-ranked eligible content is arranged into the For You feed.

That process happens repeatedly as the platform generates recommendations for users.

What This Means for Creators in 2026

For creators, the biggest lesson is that X is increasingly focused on relevance rather than simple follower distribution.

A creator should therefore think about whether a post is understandable, useful, interesting, entertaining, informative, or otherwise valuable to a clearly identifiable audience.

Strong ideas can help because users are more likely to read, reply, repost, click, or otherwise interact with content they find genuinely relevant.

This does not mean that every post must become controversial or optimized for engagement at any cost. Artificial engagement can also create poor-quality signals and does not guarantee long-term recommendation success.

The more useful way to think about the system is that every post is a candidate competing for attention among other potentially relevant content.

What the Algorithm Does Not Mean

Understanding the recommendation system also means avoiding common myths.

There is no reliable evidence in the current public architecture for a simple universal formula in which a fixed number of likes automatically produces a fixed number of impressions.

There is also no single universal posting time that guarantees distribution, no number of hashtags that guarantees virality, and no follower count at which an account automatically becomes eligible for massive reach.

The actual system is personalized, probabilistic, and continuously changing.

Most importantly, the public repository provides substantial technical information, but it should not be treated as a promise that every internal production detail is permanently identical to the publicly released code. Recommendation systems evolve, and X's public documentation itself reflects updates over time.

Final Thoughts

The X algorithm in 2026 is best understood as a recommendation system rather than a popularity counter.

It combines content from accounts a user follows with content discovered outside that network. It gathers information about the user and candidate posts, filters out unsuitable content, and uses a Grok-based transformer ranking system known as Phoenix to predict multiple forms of engagement and determine which posts should appear higher in the For You feed.

For creators, this means that follower count is only one part of the larger distribution environment. A post can potentially travel beyond its existing audience when the recommendation system finds it relevant to people who have never followed the account.

The most useful way to approach X in 2026 is therefore not to search for one secret “viral formula.” The platform is evaluating relationships between people, content, interests, and predicted actions at a much larger scale.

Once that is understood, the behavior of the For You feed becomes easier to explain. X is not simply asking which post is the most popular. It is continuously asking which available post is most likely to be meaningful to the particular person looking at the screen.