The definitive guide to maximizing your reach on X (Twitter), based on reverse-engineering the open-source algorithm.
X (formerly Twitter) open-sourced their recommendation algorithm, and xAI ships regular updates — latest: September 18, 2026 (refreshed every ~4 weeks with developer notes). This playbook distills thousands of lines of code into actionable rules that anyone can follow to maximize their reach.
No fluff. No guesswork. Just algorithm-backed strategies.
🆕 New in the Aug–Sep 2026 updates: the real action weights are now public (copy-link share 20.0, reply 5.0, report −234.0…), the production Phoenix model shipped, a visibility-filtering engine (54 rules incl. recommendation-only drops), SimClusters OON source, VMRanker DPP reranking, a new-author cold-start boost, and Under the Hood transparency reporting. See what changed →
| If you want to... | Read this |
|---|---|
| Get the essentials | Golden Rules |
| Understand scoring | Scoring System |
| Optimize before posting | Pre-Post Checklist |
| Avoid penalties | Avoiding Penalties |
| Deep dive | Action Weights Reference |
| See the latest changes | September 2026 Update |
Your Post Score = Σ (weight × P(action))
Where actions include (real weights):
├── POSITIVE: share via copy link (20.0), reply (5.0), quote (5.0),
│ DM share (5.0), follow (4.0), share (2.0), retweet (1.0),
│ like (0.5), click family (0.05–0.4), dwell (0.004–0.05)
├── NEGATIVE: report (−234), mute (−58.8), "not interested" (−43.2),
│ block (−31.2), not dwelled (−0.02)
└── BOOST: reply on a mutual-follow original post → 20.0
Key insight: Phoenix predicts 64 action classes (24 named feed actions + continuous dwell signals); ~26 weighted terms feed the score. Weights multiply predicted probabilities per viewer, not raw counts. Your goal: maximize positive action probability, avoid negative signals.
x-algorithm-playbook/
├── rules/ # Core strategies (start here)
│ ├── 00-golden-rules.md ← The 10 most important rules
│ ├── 01-scoring-system.md ← How your posts are scored
│ ├── 02-content-optimization.md
│ ├── 03-engagement-tactics.md
│ ├── 04-timing-frequency.md
│ ├── 05-avoiding-penalties.md
│ └── 06-growth-strategies.md
│
├── checklists/ # Actionable checklists
│ ├── pre-post-checklist.md
│ ├── weekly-audit-checklist.md
│ └── profile-optimization.md
│
├── case-studies/ # Real examples
│ ├── viral-thread-anatomy.md
│ └── common-mistakes.md
│
└── reference/ # Technical deep-dives
├── action-weights.md
├── filter-system.md
├── september-2026-update.md ← What changed in the latest release
├── may-2026-update.md ← Superseded (weights were still redacted then)
└── algorithm-faq.md
- Get sent, get replies — Copy-link share (20.0) and replies (5.0; 20.0 from mutuals) are the top signals
- Avoid negative actions — Report −234, mute −58.8, not-interested −43.2, block −31.2
- Space your posts — Author diversity: your 2nd post in a feed scores ×0.625 (floor ×0.25)
- In-network first — Out-of-network content takes a verified ×0.75 discount
- Video > Image > Text — Verified 10s minimum; direct video weight is small (0.07)
- Dwell time matters — Continuous dwell term (0.004) rewards longer reads
- Don't trigger filters — 17 pipeline filters + a 54-rule visibility engine can hide you
- Engage authentically — Algorithm tracks your interaction patterns
- Niche down — Semantic IDs + SimClusters reward consistent topics
- Quality > Quantity — One great post beats five mediocre ones
The algorithm uses a Grok-based transformer model (Phoenix) to predict engagement:
┌─────────────────────────────────────────────────────────────────┐
│ YOUR POST ENTERS THE SYSTEM │
└─────────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ CANDIDATE SOURCING │
│ ├── Thunder: in-network posts from accounts you follow │
│ ├── Phoenix Retrieval: ML-discovered out-of-network posts │
│ ├── SimClusters: engagement-cluster candidates (new Aug 2026) │
│ ├── Phoenix Topics · Phoenix MoE (off by default) │
│ └── Who-to-Follow · Ads · Prompts │
│ (grox classifies + semantic IDs tag every post first) │
└─────────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ FILTERING (17 pre-scoring filters can remove your post) │
│ ├── Age filter (>48h — verified) │
│ ├── Muted keywords │
│ ├── Blocked/muted authors │
│ └── Spam/safety + visibility-filtering verdicts │
└─────────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ SCORING (ranking_scorer.rs — verified order) │
│ ├── Phoenix Scorer: predicts 64 action classes + dwell │
│ ├── Weighted sum: Σ (weight × P(action)) — real weights │
│ ├── Cold-start boost: small authors lifted to ~slot 15–16 │
│ ├── Author diversity: ×0.75·0.5^k + 0.25 │
│ ├── OON discount: ×0.75 (×0.5 on topic surfaces) │
│ └── VMRanker: DPP rerank for feed diversity (top 150) │
└─────────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ RANKING + BLENDING: posts interleaved with ads/prompts/WTF │
└─────────────────────────────────────────────────────────────────┘
Since August 2026, the exact production weights are public in
home-mixer/params/param.rs. They multiply predicted probabilities — not raw action counts.
| Action | Real weight | Impact |
|---|---|---|
| Share via copy link | 20.0 | Top positive signal |
| Reply (mutual follow) | 20.0 | 5.0 + 15.0 bidirectional boost |
| Reply · Quote · DM share | 5.0 | Strong engagement |
| Follow Author | 4.0 | High intent signal |
| Share | 2.0 | Moderate |
| Retweet | 1.0 | Good reach |
| Like | 0.5 | Baseline positive |
| Click / link / media | 0.05–0.4 | Weak positive |
| Dwell Time | 0.004–0.05 | Longer = better (continuous) |
| Block | −31.2 | Strong negative |
| "Not Interested" | −43.2 | Harsher than block |
| Mute | −58.8 | Harsher than block |
| Report | −234.0 | Most severe |
Found an insight? Want to add a case study? PRs welcome!
See CONTRIBUTING.md for guidelines.
- X Algorithm Source Code (xAI) — Primary source for this playbook
- X Algorithm (Twitter Archive)
- X Algorithm ML Components
- Original analysis from codebase review
MIT License - Use freely, attribution appreciated.
Star this repo if it helped you grow on X!
⭐ Star on GitHub