How the X Algorithm Works After the Grok Update
X has published its recommendation code twice: in 2023, and again in January 2026 with a Grok-based ranker. Here's what we know about how the For You feed works now.
Key takeaways
- X shows two feeds: For You (algorithmic) and Following (chronological).
- X open-sourced much of its ranking code in March 2023, revealing how heavily conversation signals were weighted.
- In January 2026 X published a new recommendation system built on a Grok-based transformer (Phoenix), replacing the legacy ranker.
- Replies, conversations and dwell time remain central, and relevance to each user's interests matters more with an AI ranker.
X is unusual among social platforms: it has published its recommendation code publicly twice. That gives us a clearer view of how the For You feed works than we have for most platforms, with the caveat that live systems can differ from published code.
Two feeds
- For You: algorithmic recommendations from accounts you follow and accounts you don’t
- Following: posts from accounts you follow, in chronological order
March 2023: the first open-source release
On March 31, 2023, Twitter published much of its recommendation code on GitHub. It showed a pipeline that:
- Collected candidates from in-network (accounts you follow) and out-of-network sources
- Ranked them with a neural network predicting engagement types (likes, replies, reposts, profile clicks, dwell time)
- Applied filters and heuristics for balance, safety and diversity
The published weights showed conversation signals (especially replies that the author engages with) counting far more than likes.
January 2026: the Grok-based ranker
In January 2026 X published a new recommendation system on GitHub under xAI, built around a Grok-based transformer called Phoenix that learns relevance from engagement patterns rather than hand-coded features, replacing the legacy ranking model. We mark this as reported: check X’s own repository and announcements for the latest details. See the X update history.
What this means in practice
An AI ranker trained on engagement patterns is better at matching posts to each person’s interests. That rewards:
- Topical consistency: the system learns who finds your posts relevant
- Conversation: replies and back-and-forth discussion
- Dwell time: posts people stop and read
- Native media that holds attention
Factors marketers commonly report
- Posts with external links often get lower reach (test links in replies)
- Premium accounts get some visibility benefits
- Engagement in the first hour matters for wider distribution
These are observations from marketers rather than confirmed rules, so test them on your own account.
What brands should do
- Stay focused on a few topics your audience cares about
- Reply and converse, don’t just broadcast
- Use native images, video and threads
- Measure profile visits and conversions, not just impressions
Practical tips: X growth tips for brands. Full review: X account audit.