Why Your Post Gets Shown Outside Your Followers

1. Phoenix Retrieval Finds Posts Beyond the Follow Graph

The out-of-network system is powered by Phoenix Retrieval. Its job is to find potentially relevant posts from a much larger corpus rather than limiting candidate generation to accounts the viewer follows. (GitHub) The retrieval system uses a two-tower model:

The implementation describes the user tower as encoding user features and engagement history, while the candidate tower projects post and author embeddings into a shared space. (GitHub)

2. Your Post Doesn't Need to Be Followed to Be Relevant

Imagine someone has never heard of you, but they regularly engage with X algorithm discussions, machine learning, recommendation systems, and AI engineering. If you publish “X's open-source algorithm reveals how out-of-network posts are retrieved,” that post may be a potential candidate for that user even though there is no direct follower relationship. The reason is behavioral relevance, not the follow relationship.

3. X Learns From User Engagement History

The current X repository says the recommendation system uses engagement history—such as what users liked, replied to and shared—to determine what content is relevant to them. (GitHub) The system doesn't need to know whether a person follows the creator; it can instead ask whether the candidate appears relevant to that person's behavioral profile.

4. Similarity Search Finds Potential Matches

Phoenix Retrieval uses normalized user and candidate representations and retrieves candidates using similarity search (specifically dot-product similarity for the retrieval step). (GitHub) A high similarity makes a post a stronger retrieval candidate, but retrieval similarity is an earlier stage, not the final recommendation score.

5. Retrieval and Ranking Are Different

X's open-source architecture has two distinct stages: Retrieval (find potentially relevant posts) and Ranking (evaluate and order the retrieved candidates). Phoenix's documentation describes retrieval as narrowing millions of possible items down to a much smaller candidate set, followed by a more expressive ranking model. (GitHub) Your post first needs to become a candidate, then compete.

6. Being Retrieved Doesn't Guarantee a View

Suppose Phoenix retrieves your post for a user. That does not mean the user will definitely see it. The current pipeline subsequently applies candidate hydration, filtering, scoring, ranking and selection. (GitHub) Retrieved → Eligible? → Scored → Ranked → Selected? → Shown. Retrieval creates an opportunity, not a guaranteed impression.

7. Phoenix Predicts What the User Might Do

After retrieval, Phoenix's ranking model predicts multiple possible user actions (Like, Reply, Repost, Quote, Click, Profile click, Share, Dwell, Follow, Not Interested, Block, Mute, Report, and other engagement/attention signals). (GitHub) These predictions contribute to the candidate's ranking score. (GitHub)

8. The Same Post Can Be Relevant to Many Different People

Imagine your post is about AI agents. Different users might encounter it because they have different but related interests (AI researchers, developers, startup founders, tech investors, AI enthusiasts). The system doesn't need all of these people to follow you; it can potentially discover the post for them through out-of-network retrieval if the content matches their behavioral context.

9. Your Follower Count Isn't the Ceiling

A creator with 1,000 followers can publish a post that undergoes out-of-network retrieval, reaches potentially relevant non-followers, competes in ranking, and yields 10,000+ possible impressions. There is no contradiction here: the first 1,000 people represent the creator's existing follower relationship, while the additional audience comes from recommendation.

10. Out-of-Network Doesn't Mean Random

A non-follower seeing your post does not mean X randomly selected it. The current architecture specifically describes Phoenix Retrieval as ML-based similarity search across a global corpus. (GitHub) Not following you + relevant behavioral context = potential recommendation.

11. Your Post and Author Both Contribute to Retrieval

The Phoenix retrieval implementation constructs candidate representations from post and author embeddings. (GitHub) So the retrieval representation isn't simply the raw text of the post; candidate information + author information feed into the candidate representation for similarity search.

12. Your Post Can Reach People With Similar Interests

If you post about SEO, and User A engages with SEO, User B engages with Google Search, and User C engages with web development, your post could potentially be retrieved for all of them if the model considers it relevant to their behavioral context. This is a recommendation pathway, not a guarantee.

13. Relevance Is Personalized

There isn't one universal answer to “Is my post recommendation-worthy?” The open-source architecture explicitly uses user engagement history in the retrieval process. (GitHub) The same post can yield high relevance for User A, low relevance for User B, and medium relevance for User C.

14. A Post Can Reach Outside Its Topic's Immediate Community

Out-of-network recommendation doesn't necessarily require someone to have previously interacted with your exact account. A person may have demonstrated interest in related content (e.g., your post on X ranking models matched against user history in AI, ML, recommendation systems, and Grok engineering), giving the system enough behavioral context to consider the post relevant.

15. Positive Predicted Actions Help Ranking

Once your post is retrieved, Phoenix predicts how the viewer might respond (High P(Like), Medium P(Reply), High P(Repost), High P(Click), High P(Dwell)). The weighted scorer combines predicted actions into a final score. (GitHub) This gives your post an opportunity to compete against other candidates.

16. Negative Predictions Matter Too

Phoenix also predicts negative actions (Not Interested, Block, Mute, Report), which are treated as negative feedback in the scoring architecture. (GitHub) The system is predicting what the user is likely to do if the post is recommended, encompassing both positive and negative outcomes.

17. Out-of-Network Candidates Still Compete

A non-follower recommendation isn't a free pass. After retrieval, the post competes with other candidates through filtering, Phoenix scoring, weighted scoring, author diversity, out-of-network scoring, and selection before the final feed is assembled. (GitHub)

18. Author Diversity Can Affect the Final Feed

Suppose your posts are repeatedly ranking highly for someone. The system still doesn't want your post filling every slot. The current architecture contains an Author Diversity Scorer that adjusts repeated-author candidates. (GitHub) This means even a strong out-of-network candidate can encounter feed-level adjustments.

19. Your Post Can Be Shown to Non-Followers Without Going Viral

Out-of-network recommendation (a post shown to someone who doesn't follow the author) is different from viral distribution (massive scale). A post can have 10 non-follower impressions without being viral, while 1 million non-follower impressions represent massive recommendation expansion. Out-of-network is simply the mechanism that makes broader reach possible.

20. Why a Post Might Stop Reaching Non-Followers

Out-of-network distribution isn't permanent. A post may stop receiving additional recommendations because other candidates become more competitive, relevance declines, users have already seen it, filtering applies, it isn't selected among top candidates, or feed diversity affects selection. (GitHub)

21. Why Some Posts Get Huge Non-Follower Reach

A simplified viral path: Your Post → Existing followers → Initial signals → Phoenix Retrieval → Relevant non-followers → Positive responses → Strong ranking → More recommendation → More non-followers. This illustrates how recommendation can move a post beyond the creator's existing network.

22. What Creators Should Focus On

You can't directly control Phoenix Retrieval, but you can make your content easier for the right audience to value:

• Make the topic clear: A clear subject helps the intended audience understand what the post is about.
• Offer something useful: Give people a reason to stop, read, click, reply, or share.
• Create original insight: Don't simply repeat what everyone else is saying.
• Make it understandable to strangers: Someone seeing your post for the first time should be able to understand its value without already knowing you.
• Don't optimize around a fake reach formula: The open-source code doesn't provide a fixed “X reposts = Y non-follower impressions.”