Why Your Post Stops Getting Recommendations

1. Your Post Is Competing With New Candidates

One of the simplest reasons reach can stop is competition. Your post may initially rank well as a strong candidate, but new posts continuously enter the candidate pool. As new candidates arrive, ranking changes, and the recommendation system selects the top candidates rather than permanently promoting a post once it performs well. The open-source pipeline uses a TopKScoreSelector after scoring. (GitHub) A post can remain good while becoming less competitive.

2. Early Reach Doesn't Guarantee Continued Reach

Suppose your post initially receives 10,000 impressions. That doesn't establish that it will continue receiving another 10,000. The recommendation process is repeated against changing candidates and user contexts (10:00 AM → High ranking; 11:00 AM → Medium ranking; 12:00 PM → Lower ranking). The open-source code demonstrates the retrieval and ranking pipeline, but it does not publish a fixed “initial test period” such as 15 or 30 minutes. So claims like “Every post gets exactly 20 minutes to prove itself” should not be presented as a documented X rule.

3. The System Predicts User Behavior

Phoenix predicts multiple actions for each candidate (Likes, Replies, Reposts, Quotes, Clicks, Profile clicks, Shares, Dwell, Follows, Not Interested, Block, Mute, Report). (GitHub) The ranking system then combines these predictions into scores. If the predicted value of the post is weaker for additional viewers, its ranking opportunities can decrease.

4. The Audience Changes

Relevance is personalized. A post that is highly relevant to one group (e.g., AI enthusiasts) may yield strong relevance initially, but as it reaches next audiences (general users), relevance can be lower. Phoenix Retrieval uses user engagement history to create user representations and retrieve potentially relevant candidates. (GitHub) Therefore, a post can perform strongly within one audience and then encounter weaker recommendation opportunities as it reaches different users.

5. Users May Have Already Seen the Post

The candidate pipeline contains PreviouslySeenPostsFilter, PreviouslySeenPostsBackupFilter, and PreviouslyServedPostsFilter. These filters exist to prevent already encountered content from repeatedly entering the recommendation process for the same user/session. (GitHub) As more relevant users encounter the post, the pool of new recommendation opportunities can change.

6. The Post Can Become Less Competitive

Imagine ranking scores changing from [Post A: 0.91, Your Post: 0.86, Post C: 0.84] to [Post A: 0.91, Post B: 0.89, Post C: 0.88, Your Post: 0.86]. Nothing necessarily happened to your post; the competition changed. The open-source architecture ranks candidates against one another before selecting the top candidates. (GitHub)

7. Author Diversity Can Reduce Repetition

X's open-source architecture includes an Author Diversity Scorer specifically designed to diversify authors within a feed response. (GitHub) If your posts are performing well, a feed optimized purely for score could become dominated by one creator. Author diversity adjustments can reduce repeated-author representation, meaning even strong posts don't necessarily get unlimited consecutive distribution.

8. Previously Served Content Matters

The pipeline also contains a PreviouslyServedPostsFilter. (GitHub) This is particularly important when interpreting repeated impressions. A recommendation system doesn't need to repeatedly serve the exact same post to the same user. Instead, once shown, other candidates become eligible and the feed continues. Diminishing reach can partly reflect the shrinking pool of new users who can receive the post.

9. Negative Feedback Can Reduce Ranking

Phoenix doesn't only predict positive engagement; it also predicts Not Interested, Block, Mute, and Report, which are explicitly classified as negative feedback in the model. (GitHub) If the system predicts that a particular viewer is more likely to dislike or reject a candidate, that can work against recommendation for that viewer.

10. A Post Can Be Relevant but Still Not Selected

Suppose Phoenix Retrieval identifies your post as potentially relevant. That means candidate—it does not mean guaranteed impression. The pipeline continues through filtering, scoring, ranking and selection. (GitHub) A post can fail at the final competition stage without anything being “wrong” with the post.

11. Out-of-Network Reach Can Naturally Slow

A post may initially receive strong non-follower distribution (followers → initial audience → out-of-network retrieval → new audience → more recommendations), but the process doesn't guarantee unlimited expansion. Once the post becomes less competitive for additional users, or fewer suitable candidates remain, distribution can slow. The open-source system specifically separates out-of-network retrieval from later ranking and selection. (GitHub)

12. Your Post May Be Too Old for Candidate Retrieval

The current candidate pipeline includes an AgeFilter. (GitHub) This means post age is explicitly part of candidate eligibility. The exact production threshold should not be assumed from the existence of the filter alone, but post age can be an eligibility consideration, meaning an older post can naturally have fewer recommendation opportunities.

13. Filtering Can Remove a Candidate

The candidate pipeline contains several filters, including duplicate removal, core-data validation, age filtering, self-post filtering, repost deduplication, subscription eligibility, previously seen, previously served, muted keywords, blocked/muted authors, video-related filtering, and topic filtering. (GitHub) Post-selection processing also performs additional validation. So a post can stop appearing because of an eligibility or visibility condition—not necessarily because its ranking suddenly collapsed.

14. Conversation Deduplication Can Matter

The system includes DedupConversationFilter during post-selection processing. (GitHub) This prevents multiple candidates representing essentially the same conversation from unnecessarily occupying feed slots (e.g., original post, reply, quote, another reply). This is another reason why raw engagement counts don't directly translate into recommendation slots.

15. More Engagement Doesn't Mean Infinite Reach

A common assumption is that if a post keeps getting likes, X must keep recommending it. Not necessarily. Engagement is one part of the prediction and ranking process. Phoenix predicts multiple actions, and the ranking system uses those predictions to score candidates. (GitHub) More likes ≠ guaranteed continued recommendation; the post still has to compete for recommendation opportunities.

16. A Post Can Have Strong Overall Engagement but Weak New-Audience Relevance

An existing audience may love a post, yielding high engagement, but the next group of users may have very different interests. The system evaluates candidates in the context of the viewer, not simply the post's accumulated engagement. Strong past engagement ≠ strong future relevance for everyone, which is one of the biggest reasons viral-looking posts can eventually plateau.

17. Reach Can Plateau Without a Penalty

Rapid early growth followed by a plateau doesn't automatically indicate suppression. A normal explanation can be: Initial strong candidates → Strong ranking → Large early distribution → More competition + Fewer new relevant users + Previously served users → Lower incremental reach. The public architecture supports these mechanisms, although it does not provide a universal formula for how quickly any individual post will plateau. (GitHub)

18. The Recommendation System Is Not a One-Time Decision

Instead of publishing once and having X decide reach permanently, recommendation is a continuous loop of candidate retrieval, ranking, selection, new user/context, retrieval, ranking, selection. X's open-source Phoenix pipeline explicitly separates retrieval from ranking and processes user history as part of the recommendation model. (GitHub)

19. Why Reach Can Stop Suddenly

From the outside, impression growth can plateau sharply. But the underlying change is gradual: Candidate competition increases, new relevant users decrease, previously served users increase, predicted value shifts, and selection probability drops. A visible plateau doesn't tell you which specific mechanism caused it.

22. What Creators Can Actually Control

You cannot control competing candidates, user interests, previous viewing history, ranking weights, author diversity adjustments, previously seen pools, or candidate pools. But you can control:

• Relevance: Make the subject useful to a clearly identifiable audience.
• Value: Give people a reason to read, interact, or share.
• Attention: Make the content worth consuming rather than merely clicking.
• Conversation: Create something people genuinely want to discuss.
• Shareability: Give people something worth passing to others.

23. Don't Chase a “Second Boost”

There is no documented public rule saying “Reply to your post after one hour and X will restart distribution.” Likewise, the open-source code does not establish that self-reposting automatically resets reach, quoting yourself automatically restarts recommendations, editing a post gives it a new recommendation cycle, or replying to every comment guarantees another boost. These may be observations creators make, but they should not be presented as verified algorithm rules.