How X's Algorithm Actually Works
From the source code — every claim traceable to a real file in xai-org/x-algorithm.
The Lessons
How a Post Reaches Your Feed
The big picture: Post Pipeline + Blending Pipeline, two candidate pools
Where Candidates Come From
Thunder, Phoenix retrieval, and SimClusters — the three sourcing systems
How Posts Actually Get Scored
Phoenix predicts probabilities for many actions — not one relevance score
The Three Adjustments Nobody Talks About
Author-diversity decay, out-of-network discount, and new-author boost
What Gets Filtered Before It's Even Scored
Age cutoff, dedup, muted keywords, and the new-user engagement floor
Trust, Safety & Under the Hood
How posts become ALLOW / INTERSTITIAL / DROP, and X's transparency tool
Case Study: The Bidirectional Follow Boost
A real, dated 2026 algorithm change — told in full
What X Won't Show You (And Why That's OK)
The repo deliberately withholds some things to reduce gaming
Weights, Right Now
| Name | Value | Description |
|---|---|---|
| AuthorDiversityDecay | 0.5 | Decay factor for same-author posts in one session |
| AuthorDiversityFloor | 0.25 | Floor for author diversity decay |
| BidirectionalFollowDwellWeightBoost | 0 | Bonus on dwell weight for mutual-follow posts |
| BidirectionalFollowReplyWeightBoost | 15 | Bonus on reply weight for mutual-follow original posts |
| BlockAuthorWeight | -31.2 | Predicted probability you'll block the author |
| ClickWeight | 0.4 | Predicted probability you'll click into the post |
| ColdStartFollowerCap | 1000 | New-author boost applies below this follower count |
| ColdStartImpressionThreshold | 1000 | New authors boosted until this many impressions |
Common Questions
Is the X algorithm really open source?
Yes. xAI published the core ranking pipeline, sourcing systems, filters, and visibility logic at github.com/xai-org/x-algorithm. Some abuse-detection rules and LLM prompts are withheld to reduce gaming, but the ranking weights and pipeline structure are public.
Do the weights multiply raw engagement counts?
No. The repo explicitly warns that weights multiply the predicted probability of each action, not raw counts. A report weight of -234 does not mean one report cancels 468 likes — it scales how much the predicted likelihood of a report affects ranking.
How often does the algorithm change?
The repo is updated regularly. The bidirectional follow boost went from +20 (July 13, 2026) to +15 (July 24, 2026) after real-world feedback. Our changelog page tracks detected changes within hours of the repo updating.