X Algorithm Playbook

Logo

The definitive guide to maximizing your reach on X (Twitter), based on reverse-engineering the open-source algorithm.

View the Project on GitHub tang-vu/x-algorithm-playbook

Algorithm FAQ

Frequently asked questions about the X algorithm, answered with code references.

General Questions

Q: Is the X algorithm really open source?

A: Yes, partially. X released the recommendation algorithm code in April 2023. The core logic for scoring, filtering, and ranking is public. However, specific parameter values (weights, thresholds) are redacted for security.

Source: github.com/twitter/the-algorithm

Q: How often does the algorithm change?

A: The open-sourced version is a snapshot. X continues to iterate internally. The fundamental mechanics (scoring, filtering, Two-Tower retrieval) are likely stable, but specific weights and parameters change.

Q: Does paying for X Premium help reach?

A: Based on the public code, there’s no explicit “premium boost.” However, premium features (longer posts, edit button) may indirectly help engagement. The algorithm focuses on engagement signals, not subscription status.

Scoring Questions

Q: Which action matters most?

A: Based on code analysis, Reply has the highest positive weight (~2× other actions), followed by Quote Tweet and Follow.

Source: weighted_scorer.rs

Q: How bad is getting blocked?

A: Very bad. Block appears to have approximately -10× the weight of a like. One block can undo the positive signal from 10 likes.

Source: weighted_scorer.rs - BLOCK_AUTHOR_WEIGHT

Q: Does “view” count matter?

A: Views/impressions are not directly weighted. What matters is engagement (actions taken). However, dwell time (how long someone looks at your post) IS tracked and weighted.

Q: Are all retweets equal?

A: The algorithm looks up the original tweet ID for retweets when scoring. Your retweet’s score is influenced by the original content’s predicted engagement.

Source: phoenix_scorer.rs

Reach Questions

Q: Why did my reach suddenly drop?

A: Possible causes:

  1. Getting blocked/muted by multiple users (author socialgraph filter)
  2. Content topic shift (embedding mismatch)
  3. Algorithm changes (happens regularly)
  4. Posting frequency (author diversity penalty)
  5. Lower engagement rate (signals declining interest)

Q: How do I appear on non-followers’ feeds?

A: Through the Two-Tower retrieval system (Phoenix). Your content embedding is matched with user embeddings via similarity search. Being consistent in your niche creates a cleaner embedding that matches better.

Source: phoenix/recsys_retrieval_model.py

Q: Does posting time matter?

A: Yes. The Age Filter removes posts older than a threshold. Posts need early engagement to be distributed widely. Posting when your audience is online is crucial.

Source: age_filter.rs

Q: Do hashtags help?

A: Not directly mentioned in scoring code. Hashtags may help with:

  1. Content embedding (topic signal)
  2. Search discovery
  3. User muted keyword filters (could hurt if hashtag is muted)

Use relevant hashtags sparingly.

Content Questions

Q: Are threads better than single tweets?

A: Threads have advantages:

  1. Count as ONE “author slot” (no diversity penalty within thread)
  2. Higher dwell time (people read multiple tweets)
  3. More engagement opportunities per thread

Source: Author diversity logic in author_diversity_scorer.rs

Q: Does video get boosted?

A: Video has a dedicated score (Video Quality View), but ONLY if the video exceeds a minimum duration threshold. Short clips don’t get this bonus.

Source: weighted_scorer.rs - VQV_WEIGHT, MIN_VIDEO_DURATION_MS

Q: Do images help?

A: Photo Expand is tracked as a positive action, but it’s low-weight. Images help more by:

  1. Stopping scroll (attention)
  2. Increasing dwell time
  3. Making content shareable

A: The code doesn’t show explicit link penalties. However:

Technical Questions

Q: What is “candidate isolation”?

A: A key ML architecture decision. When the transformer scores candidates, each post can only “see” the user context—not other candidate posts. This ensures your score doesn’t depend on what other posts are in the batch.

Source: phoenix/grok.py - make_recsys_attn_mask()

Q: What is the Two-Tower model?

A: A retrieval system with two neural networks:

  1. User Tower: Encodes user + history into embedding
  2. Candidate Tower: Encodes posts into embeddings
  3. Matching: Dot product similarity finds relevant posts

Source: phoenix/recsys_retrieval_model.py

Q: How does the author diversity penalty work?

A: When multiple posts from the same author appear:

Post 1: 100% score
Post 2: ~76% score  
Post 3: ~59% score
...converges to ~20% floor

Formula: multiplier = (1 - floor) × decay^position + floor

Source: author_diversity_scorer.rs

Q: What is OON penalty?

A: Out-of-network (OON) content (from accounts you don’t follow) is multiplied by OON_WEIGHT_FACTOR (less than 1.0). This gives in-network content an advantage.

Source: oon_scorer.rs

Myth Busting

Myth: Posting more = more reach

Fact: Author diversity penalty means your 2nd, 3rd posts get progressively lower scores. Quality > quantity.

Myth: Likes are the most important metric

Fact: Replies have ~2× the weight of likes in the scoring formula.

Myth: The algorithm is completely random

Fact: It’s a deterministic ML system based on predicted engagement. Understanding it helps.

Myth: Buying followers helps reach

Fact: Followers who don’t engage provide no positive signals. Low engagement rate may hurt your distribution.

Myth: You can’t recover from low reach

Fact: The algorithm is forward-looking. Consistent high-quality content will improve your signals over time.

Still Have Questions?

Open an issue on this repo with the question label.

Back to: Rules Checklists