What Makes a Post Recommendation-Worthy?

1. The Post Must First Be Discoverable

Before X can recommend a post, it needs to enter the candidate pool. The current architecture has two major sources: Thunder (posts from accounts the user follows) and Phoenix Retrieval (posts discovered from a broader global corpus). (GitHub)

A post can therefore be recommendation-worthy even if the creator has relatively few followers.

2. The Post Must Be Eligible

Being potentially relevant isn't enough. The current pipeline applies pre-scoring filters that can remove candidates for reasons including duplicate posts, posts that are too old, the viewer's own posts, blocked or muted authors, muted keywords, previously seen posts, recently served posts, and ineligible subscription content. (GitHub) Recommendation-worthy ≠ automatically eligible. A post can have strong potential but still be excluded from a particular user's recommendation opportunity.

3. Relevance Is Personal

X doesn't appear to ask “Is this post good for everyone?” The architecture uses the viewer's engagement history to determine what content is relevant to them. The repository says the system uses engagement sequences such as things the user liked, replied to, or shared. (GitHub) User A with AI history viewing an AI post yields high potential relevance, while User B with sports history viewing the same AI post yields lower potential relevance. The post hasn't changed; the viewer has.

4. Phoenix Retrieval Finds Potential Matches

Phoenix Retrieval uses a two-tower architecture. The user tower encodes user features and engagement history into a user representation, and the candidate tower represents posts. The system then uses similarity search to retrieve potentially relevant candidates. (GitHub) The current example pipeline retrieves the top candidates using dot-product similarity before sending them to ranking.

5. Retrieval Is Not the Final Ranking

This distinction is crucial. Retrieval asks: Could this post be relevant? Ranking asks: How strong is this candidate for this viewer compared with the other candidates? Phoenix explicitly separates these two stages. Retrieval narrows a very large candidate pool, while ranking evaluates the smaller candidate set more deeply. (GitHub) A post can be retrieved but still fail to rank highly enough.

6. The System Predicts User Behavior

This is one of the biggest differences between the X architecture and a simple engagement counter. Phoenix predicts multiple actions for each candidate (Likes, Replies, Repost, Quotes, Clicks, Profile clicks, Video views, Photo expansions, Shares, DM shares, Copy-link shares, Dwell, Follows, Not Interested, Blocks, Mutes, Reports, Dwell time). (GitHub) So instead of Post = 500 likes, the model produces probabilities for each action.

7. Multiple Positive Actions Can Make a Candidate Strong

A post may have several predicted positive behaviors (e.g., High P(Like), High P(Reply), High P(Repost), High P(Dwell)). These predictions contribute to the candidate's ranking. The repository describes the final score as a weighted combination of predicted engagements: Final Score = Σ(weight × P(action)). (GitHub) This is why a recommendation-worthy post shouldn't be reduced to getting lots of likes.

8. Attention Matters Too

The current action set includes dwell and continuous dwell time predictions. (GitHub) That means explicit button clicks aren't the only modeled behavior. A user seeing a post, stopping scrolling, reading/watching, and dwelling provides useful behavioral information without necessarily receiving a like or repost.

9. Negative Feedback Matters

Recommendation isn't simply about predicting positive actions. Phoenix also predicts Not Interested, Block author, Mute author, and Report. (GitHub) The current architecture treats these as negative feedback signals in scoring, combining positive predictions with negative predictions to form the overall score. This means a recommendation-worthy post isn't merely one that can generate engagement; it should also avoid creating strong signals that the viewer doesn't want the recommendation.

10. The Post Has to Beat Other Candidates

Even if a post is relevant, it still has competition. The current pipeline scores candidates and selects the top candidates. (GitHub) Your post can be good and relevant but still lose the recommendation opportunity against stronger candidates. So recommendation-worthy doesn't mean guaranteed to be selected.

11. Author Diversity Also Matters

The current pipeline includes an Author Diversity Scorer that attenuates repeated-author scores. (GitHub) This prevents a feed from simply becoming filled with posts from Creator A. So even multiple strong posts from the same creator may not all appear together.

12. Out-of-Network Content Has Its Own Scoring

X also applies an out-of-network scorer when a post comes from someone the viewer doesn't follow. (GitHub) That creates an important path for creators (Creator → Post → Out-of-network retrieval → Relevant user → Ranking → New audience), showing how a post can travel beyond the existing follower base.

13. A Recommendation-Worthy Post Doesn't Need Huge Existing Engagement

A common misconception is that X only recommends posts that already have lots of engagement. The architecture doesn't support such a simple rule because Phoenix is predicting future user actions rather than merely sorting posts by existing engagement totals. (GitHub) A post with relatively little existing engagement can still be a strong candidate for a particular user if predicted engagement is high.

14. Relevance Beats Universal Popularity

Consider Post A with 100,000 likes and low relevance to User X, versus Post B with 500 likes and high relevance to User X. Post B can potentially be the better recommendation for User X because the system is designed around personalized predicted engagement, not a global popularity leaderboard. (GitHub)

15. Content Clarity Can Help the Right Audience Find It

The open-source system includes content understanding components such as post-category classification and policy-related processing. (GitHub) But this should not be turned into unsupported claims like “X gives posts with keyword X a ranking boost.” The safer conclusion is that clear, understandable content makes it easier for the right audience to understand and engage.

16. Originality Isn't a Public Score

There is no documented Originality Score in the public architecture. So don't claim original posts receive an automatic originality boost. Instead, original content can be useful because it creates stronger reasons for users to read, like, reply, repost, quote, share, follow, and spend time—behaviors the system actually models. (GitHub)

17. A Post Can Be Good but Not Recommendation-Worthy

A post can be good but not retrieved for enough relevant users, filtered for a particular viewer, weakly predicted for that viewer, outranked by competing candidates, affected by author diversity, or not selected. So content quality and recommendation-worthiness are related, but they aren't the same thing.

18. The Recommendation-Worthy Formula (Conceptual)

This is a conceptual framework, not an official X formula. The actual open-source system contains multiple pipeline stages and configurable components. (GitHub)

19. What Creators Can Actually Control

Creators cannot control retrieval, other candidates, viewer interests, ranking weights, selection, or diversity adjustments. But creators can influence the content itself:

• Make the topic clear: Help the intended audience immediately understand what the post is about.
• Provide real value: Give users a reason to read or interact.
• Encourage genuine discussion: Create ideas worth replying to rather than manufacturing replies.
• Make it worth sharing: Give people something they want others to see.
• Earn attention: Don't just optimize for clicks; make the content worth consuming.
• Avoid unnecessary negative reactions: Don't build a strategy around bait, spam, or repetitive low-value content.

20. What Doesn't Make a Post Automatically Recommendation-Worthy

The open-source code does not establish that any of these automatically guarantee recommendation:

• ❌ A large follower count
• ❌ A specific number of likes or reposts
• ❌ A specific engagement rate
• ❌ Posting at a particular minute or using specific hashtags
• ❌ Asking people to like the post, reposting/quoting your own post
• ❌ Having a verified account

21. What the Open Source Actually Shows

The current X repository gives us a much clearer model:

1. Retrieve: Find potentially relevant candidates from in-network and out-of-network sources. (GitHub)
2. Filter: Remove candidates that aren't eligible for the viewer. (GitHub)
3. Predict: Phoenix predicts multiple possible user actions. (GitHub)
4. Score: Combine predicted actions into a ranking score. (GitHub)
5. Adjust: Apply author diversity and out-of-network adjustments. (GitHub)
6. Select: Choose the top candidates. (GitHub)
7. Validate: Apply post-selection processing and visibility checks. (GitHub)