How to Increase Your Chances of Recommendation on X

1. Make Your Topic Relevant to a Real Audience

The first principle is relevance. Phoenix uses a user's engagement history to understand what content is relevant to them. Its retrieval system encodes the user's history into a representation and compares it with candidate representations to retrieve potentially relevant posts. (GitHub)

This means a post doesn't need to appeal to everyone. It needs to be highly relevant to some users. Instead of “I want everyone to like this,” think: “Who is most likely to care about this?” Then make the post clearly relevant to that audience.

2. Give the System a Clear Content Signal

If your post is about a specific subject, make that subject clear. For example, compare “This changes everything” with “X's open-source recommendation model predicts multiple user actions.” The second post gives both humans and recommendation systems a much clearer understanding of the topic. The current X repository also includes a separate content-understanding component for tasks such as post-category classification and spam/policy detection. (GitHub) So clarity is useful beyond human readers.

3. Create Content People Are Likely to Engage With

Phoenix predicts multiple possible actions (Like, Reply, Repost, Quote, Click, Profile click, Share, Dwell, Follow, Video-related actions, Negative feedback). (GitHub) This means creators shouldn't optimize around only likes. A strong post can give users several reasons to interact (Interesting idea → Read → Dwell → Like → Reply → Share / Repost). These are different behaviors that the recommendation system can model.

4. Encourage Genuine Conversation

Replies are explicitly modeled by Phoenix. That doesn't mean asking a pointless question at the end of every post. Instead, create something that naturally gives people a reason to respond (e.g., “The open-source X algorithm predicts likes, replies, reposts and clicks. Which signal do you think matters most?”). That's more useful than “Agree?” The goal should be real conversation, not artificial engagement.

5. Make Your Posts Worth Sharing

Reposts, quotes and different forms of sharing are explicitly represented in the current model. (GitHub) Ask: “Would someone want another person to see this?” Content that provides useful information, a surprising finding, a strong insight, original analysis, humor, or a useful resource can naturally create sharing opportunities. But avoid turning this into “Reposts are worth X points.” The open-source architecture predicts these actions and uses them in ranking; it does not provide a simple creator-facing reach multiplier. (GitHub)

6. Give People a Reason to Stop and Read

The current Phoenix model includes dwell and continuous dwell-time-related predictions. (GitHub) This means attention is part of the modeled behavior. A strong opening can therefore matter because it gives the viewer a reason to continue reading (e.g., “X's open-source algorithm doesn't actually rank posts by likes” is more curiosity-driven than “Here's some information about X's algorithm”). The objective isn't to trick someone into stopping; it's to make the content worth their attention.

7. Don't Optimize for Likes Alone

A common creator strategy is asking “How do I get more likes?” But the current architecture predicts many actions simultaneously. The Phoenix model outputs probabilities for multiple actions, rather than one universal engagement score. (GitHub) Think of multiple positive actions plus negative feedback rather than Likes = Reach. This is a much better mental model.

8. Create Content for the Right Viewer

Because Phoenix uses engagement history, the same post can be evaluated differently for different users. (GitHub) A post about AI coding tools has high potential relevance for a developer and low potential relevance for a sports-only user. You can't force the second user to care. Instead, create content that is strongly relevant to the audience most likely to appreciate it (Specific relevance > generic appeal).

9. Build a Consistent Topic Association

If your account repeatedly publishes useful content around a recognizable subject, you give your audience a clearer reason to engage with your posts. More importantly, the open-source retrieval architecture learns from engagement sequences rather than relying on a simple manually assigned interest list. (GitHub) Your content can therefore build a recognizable pattern of interaction.

10. Avoid Unwanted Negative Feedback

Positive actions aren't the entire story. Phoenix also predicts Not Interested, Block author, Mute author, and Report as negative feedback actions. (GitHub) Creators should consider whether people will find the content useful, but also whether the recommendation will annoy or alienate the intended audience. Avoid misleading bait, repetitive low-value posting, irrelevant content, and manipulative engagement requests.

11. Don't Assume More Posts Automatically Mean More Reach

The current open-source architecture does not establish a rule saying more posts equal proportionally more recommendation. Every post becomes another candidate, but candidates still have to compete through the recommendation pipeline. Posting more can therefore create more opportunities and more competition among your own content. There is no public formula showing that posting a specific number of times per day guarantees more reach.

12. Don't Rely on Your Followers

The For You feed combines in-network content with out-of-network content retrieved through Phoenix's ML-based retrieval system. (GitHub) That means creators should think beyond “How do I get my followers to see this?” and instead think “Who else would find this relevant?” That's where recommendation-driven growth can happen.

13. Make Posts Discoverable Outside Your Network

Phoenix Retrieval uses a two-tower system (User Tower → User representation ↕ Similarity ↕ Candidate representation → Candidate posts). The retrieval stage uses approximate nearest-neighbor search to find potentially relevant candidates from a much larger corpus. (GitHub) This means content with a clear subject and audience can potentially find users who don't already follow the author.

14. Don't Obsess Over an Exact Engagement Ratio

There is no reliable public formula such as 10 likes + 2 replies + 1 repost = Guaranteed recommendation. The current architecture predicts probabilities for different actions and uses them in ranking. (GitHub) So your analytics should be used to understand patterns, not to create fake thresholds.

15. Make Each Post Stand on Its Own

Because recommendation can introduce your content to people who don't follow you, a post shouldn't always depend on “As I said yesterday…” A new viewer should be able to understand what you're talking about, why it matters, and what the takeaway is. This increases the chance that an unfamiliar viewer can understand the candidate without needing your entire account history.

16. Use Original Information

If you want to create recommendation-worthy content, give users a reason to choose your post over hundreds of similar candidates (original research, data analysis, screenshots, experiments, explanations, comparisons, new observations, useful summaries). This isn't an explicit “originality score” documented by the open-source algorithm, but original information can create the kind of relevance and engagement opportunities that the recommendation system models.

17. Don't Confuse Engagement Bait With Recommendation Strategy

The goal isn't making people interact at any cost. For example, “LIKE THIS OR YOU'LL NEVER SEE MY POSTS AGAIN” may create an action, but it can also create a poor user experience. A better strategy is to create a natural reason to respond. The distinction is manufactured engagement versus genuine interest.

18. Keep the Audience Fit Strong

A post can be excellent but aimed at the wrong audience (e.g., posting a detailed quantum physics calculation to an audience of mostly SEO professionals). The post could be excellent, but it may not match the audience's behavioral interests. Recommendation works around user-specific relevance, so audience fit matters.

19. Think in Terms of Recommendation Opportunities

Don't measure success only as “How many impressions did I get?” Instead, think through the pipeline: How many potential users could find this relevant? How many candidate opportunities survived filtering? How strong were predicted actions? How competitive was the post? How many were selected? This is much closer to the architecture X has published. (GitHub)

20. Understand That Ranking Is Competitive

The open-source pipeline retrieves candidates and then ranks them. In the Phoenix demo, candidates are retrieved first and then ranked by predicted engagement probabilities. (GitHub) Therefore, improving your post doesn't remove competition. Your post can improve while the rest of the candidate pool improves even more. This is why recommendation performance can change from one day to another.

21. Don't Chase a “Viral Formula”

The open-source code doesn't give creators a guaranteed formula. Instead, it provides a framework: Relevant content → Candidate retrieval → Predicted user response → Ranking → Selection. Use that framework to improve your content strategy, and don't turn it into “Post at 8:03 PM with exactly three hashtags and ask for two replies.”