How X Decides Which Posts to Recommend

1. Candidate Generation

Before X can rank a post, it needs to find potential posts to consider. The older open-source recommendation system documents multiple candidate sources, including in-network posts, out-of-network recommendations, interaction graphs, and follow recommendations. The newer system describes Thunder for in-network candidates and Phoenix Retrieval for discovering relevant out-of-network posts. (GitHub) This is one reason a post can reach people who don't follow its author.

2. User Interests Matter

Recommendation systems need information about the person who may see the post. The public X code includes systems for collecting user signals and engagement history, including explicit actions such as Likes and Replies as well as implicit actions such as profile visits and post clicks. (GitHub) The newer Phoenix system similarly uses a user's engagement history to help retrieve and rank content. In simple terms: what you have interacted with helps X understand what you may want to see next.

3. Out-of-Network Recommendations

One of the most important parts of the recommendation system is the ability to discover content outside your existing network. Phoenix Retrieval uses a two-tower architecture where a User Tower represents the user and their engagement history, and a Candidate Tower represents posts. Similarity search is then used to retrieve potentially relevant posts, which are passed into later ranking stages. (GitHub) This creates an important distinction between follower distribution and recommendation distribution: a post doesn't necessarily need to come from someone you follow to become a recommendation candidate.

4. Filtering

Not every candidate reaches the ranking stage. The current open-source pipeline describes filters that can remove duplicates, old posts, posts the viewer has already seen, posts from blocked or muted accounts, muted keywords, and other ineligible content, alongside later visibility and safety filters. (GitHub) This means a post can fail to appear even before its final ranking score becomes important.

5. Scoring

After candidate generation and filtering, X needs to estimate how relevant each candidate may be to the viewer. The current Phoenix ranking system predicts probabilities for multiple possible user actions (Like, Reply, Repost, Quote, Click, Profile click, Video view, Photo expansion, Share, Dwell, Follow author, Not interested, Block, Mute, Report). (GitHub) The published system then combines these predictions into a weighted score, which is why the idea that “X only counts Likes” is an oversimplification.

6. Positive and Negative Signals

Recommendation systems don't only predict actions X wants to encourage. The current open-source documentation also includes negative actions such as Not interested, Block, Mute, and Report, which can negatively affect the predicted value of a candidate for a particular viewer. (GitHub) So recommendation is better understood as “How likely is this user to have a useful interaction with this post?” rather than “How many Likes does this post have?”

7. Ranking

Once candidates have been scored, they are ranked against each other. The current architecture describes Phoenix as a transformer-based ranking system that predicts multiple engagement probabilities for each candidate, combining those predictions to select the highest-scoring candidates. (GitHub) The older open-source architecture similarly includes light and heavy ranking stages.

8. Author Diversity

The newer candidate pipeline also includes an Author Diversity Scorer, whose purpose is to reduce repeated exposure to the same author and help diversify the content selected for a user. (GitHub) This is another reason why simply having the highest predicted engagement doesn't necessarily mean every post from an account will dominate a user's feed.

9. Why Two People Can See Different Posts

Recommendations are personalized. Two people can encounter completely different posts even when they use X at the same time because their following lists, engagement histories, interests, previously viewed posts, muted keywords, blocked accounts, interaction patterns, and predicted responses can differ. The recommendation system uses user context when retrieving and ranking candidates. (GitHub)

10. Why a Post Can Suddenly Get More Reach

A post can initially be shown to a relatively small audience and later become eligible for wider recommendation. There isn't enough public evidence to claim that every post follows one fixed “viral ladder.” However, the architecture makes the general mechanism understandable: a post becomes a candidate, users interact with it, those interactions provide signals, recommendation models estimate relevance, and the post may be surfaced to additional users. The exact production behavior depends on X's models, configuration, eligibility rules, and constantly changing systems. (GitHub)

11. Does More Engagement Always Mean More Reach?

No. Engagement is important, but it isn't one universal number. The current open-source system predicts many different actions and combines those predictions into a final score. Two posts can therefore receive similar Likes but have different recommendation outcomes. Likewise, a post with fewer Likes could potentially receive broader distribution if the system predicts stronger relevance or other useful actions for a particular audience. (GitHub)

12. Why Some Posts Go Viral

There is no single public “viral score.” A post's distribution can involve candidate eligibility, user relevance, engagement predictions, content discovery, ranking, negative feedback, author diversity, filtering, audience behavior, and competition from other candidates. The important takeaway is that virality is an outcome of recommendation and user behavior—not simply a Like counter.

13. Does Follower Count Determine Reach?

No. Follower count influences the initial network a creator can reach, but the recommendation architecture also supports out-of-network discovery. That means recommendation can extend content beyond an author's followers, which is one of the fundamental reasons a smaller account can sometimes receive substantial distribution. (GitHub)

14. What Does “For You” Actually Mean?

The For You feed is designed to personalize content for each user. The current open-source architecture combines in-network and out-of-network candidates before ranking them together. (GitHub) So For You is not simply a chronological feed with a few viral posts inserted into it; it is a recommendation system.

Under The Hood Analyzer's Approach

Under The Hood Analyzer focuses on separating algorithm facts from algorithm myths. When analyzing X recommendations, we classify information as:

• Confirmed: Directly supported by public source code or documentation. (GitHub)
• Observed: Supported by measurable or repeated platform behavior.
• Inferred: A reasonable interpretation based on available evidence.
• Unknown: Something that cannot currently be established from public information.

This approach helps creators understand what the X recommendation system actually reveals without pretending that we know X's private production logic.

The Bigger Picture

The most useful way to think about X recommendations is not “What trick makes my post viral?” Instead, think: “Why would the recommendation system believe this post is relevant to this particular person?” That's the question worth investigating. The public X recommendation code gives us a rare opportunity to examine the machinery behind content discovery, retrieval, ranking, and filtering, and Under The Hood Analyzer turns that technical information into understandable analysis for creators. (GitHub)