Marketing & Sales · Framework
Phoenix: The X Algorithm Breakdown
A playbook explaining the Phoenix recommendation model and ranking factors on X.
This framework breaks down how X's open-sourced Phoenix recommendation model scores and ranks posts. It outlines the 19 predicted positive engagement signals, 4 deboost actions, author diversity penalties, and specific best practices for optimizing reach. It is useful for understanding the exact scoring mechanics behind social distribution on the platform.
The framework
The X algo runs on Phoenix, a Grok-based model that predicts, for every post, the probability of each of 19 engagement actions: likes, replies, reposts, quotes, profile click, share, share via DM, share via copy link, click on photo, watch video beyond a few seconds, dwell on post, click on quote, follow the author (article also lists this as the 19-action set). More of these = boost; dwell time and video watch time are massive.
4 deboost actions: 'not interested' click, block, mute, report — attracting these hides you; mute bombing makes content disappear.
Author diversity penalty: reach decays with each subsequent post shown to the same feed that day — spamming posts/replies uses up your daily reach. Space out posts.
Phoenix builds interest off the last 128 posts you engaged with — don't engage with content you don't want more of.
Final Score = Σ (weight_i × P(action_i)) + NEGATIVE_SCORES_OFFSET — positive engagements minus negative ones. In-network content has no OON penalty.
Do: create engaging content (likes, replies, reposts, shares boost), make longer videos (VQV boost above minimum duration), encourage profile visits, get follows, inspire DM shares (strong signal), space out posts, build follower base.
Don't: get muted/blocked or reported, trigger 'Not Interested', post controversial content (safety filters remove entirely), spam posts rapidly (each discounted), use muted keywords.
Bottom line: quality over quantity — no more GMs; post content people value and it gets seen.
What it's for
- Optimize post formats to maximize dwell time, video watch duration, and DM shares
- Schedule and space out posts to avoid author diversity reach penalties
- Audit posting practices to prevent deboosting triggers like mutes, blocks, and reports
Key points
- 19 engagement signals predicted by the model, including dwell time, video duration, and DM shares
- 4 direct deboosting actions: block, mute, report, and 'not interested'
- Author diversity penalty mechanics showing reach decay across multiple posts in a single day
- User interest modeling based on the last 128 engaged posts
- Actionable dos and don'ts for content creation and posting frequency
Pricing
Free
Details
- Type
- Framework
- Category
- Marketing & Sales
- Topic
- Social Media Strategy
- Open source
- No
- Platforms
- Web
- For
- Content creators, social media marketers, and founders posting on X
- Tags
- x, twitter, algorithm, growth, social media, content strategy
Why it was saved
- @AlexFinn: Concrete, actionable model of what signals boost or hide content on X — the 19 predicted engagement actions, the 4 deboost actions, the author diversity penalty, and do/don't lists for going viral.
Related
Practical rules for optimizing reach and viewer retention on Instagram Reels.
A structured 30-post monthly formula for Instagram Reels reach and conversion.
A reference list of 40 short-form video frameworks for content planning.
A methodology for testing Reels formats and delegating content production.
Standard operating procedure for researching, producing, and publishing viral short-form videos.
A structured five-step blueprint to scale an Instagram account to 10,000 followers.
Updated