# X (Twitter) Algorithm Reference — For Tweet/Content Generation Compiled from X's open-sourced algorithm. Note: the 2023 repo (github.com/twitter/the-algorithm) is dead. X open-sourced a new algorithm on Jan 20, 2026 at github.com/xai-org/x-algorithm, powered by Grok's transformer architecture (the "Phoenix ranker"), updated every 4 weeks. A May 15, 2026 update added the Grox content-understanding system. Use this as the current source of truth, not the old Scala repo. --- ## Stage 1: Candidate Selection From 500M+ daily tweets, the algorithm picks ~1,500 candidate tweets per feed refresh for a user. - 50% from people you follow (ranked by how often you interact with them) - 50% from people you don't follow (out-of-network, sourced via three systems below) ### Out-of-Network Sourcing (the 50% that matters for growth) **1. Social Graph Traversal** The algo asks: what did people you follow recently engage with, and who likes the same tweets as you. If a high-influence account you follow engages with a tweet from someone you don't follow, that tweet can surface in your feed. Implication: replies/engagement from high-influence niche accounts puts your content into THEIR followers' feeds directly. **2. SimClusters** X groups all users into ~145,000 virtual communities via matrix factorization, recalculated every 3 weeks. Community placement is driven by who you follow, who you engage with, what you like/reply/retweet, and what your own tweets cover. Heavy engagement from one community surfaces your tweet to others in that community even without a follow. Implication: niche consistency is a hard requirement — mixing topics dilutes your SimCluster placement and confuses distribution. **3. TwHIN (Heterogeneous Information Network)** A 1.5B-node graph creating 128-dimensional vector embeddings for users and tweets in shared space, used to calculate similarity between any user-user, tweet-tweet, or user-tweet pair. Implication: consistent niche content + genuine engagement from niche-relevant accounts = tighter embedding = more accurate out-of-network reach. --- ## Stage 2: Heavy Ranker (Scoring Engine) Each of the ~1,500 candidates is scored by a neural network (originally a 48M-parameter MaskNet, replaced by a Grok-powered transformer as of Jan 2026). It predicts 10 engagement probabilities per tweet: 1. Like 2. Retweet 3. Reply 4. Click to expand 5. Click author profile 6. Click a link 7. Watch video 8. Spend 2+ min in thread 9. Report 10. Block Each probability × weight, summed = final score. ### Engagement Weights ``` SIGNAL WEIGHT ───────────────────────────────────────────────────────────── Reply that gets engaged by tweet author +75.0 Reply to tweet +13.5 Profile visit → like or reply to author's tweet +12.0 Click into conversation → reply or like +11.0 Dwell in conversation 2+ minutes +10.0 Retweet +1.0 Like (favorite) +0.5 Video 50%+ watch +0.005 Report tweet -369.0 Block author -74.0 Mute author -74.0 "Show less" / negative feedback -74.0 ``` ### What the weights actually mean - **Reply + author engagement (+75) is 150x a like (+0.5).** X wants conversations, not passive scrolling. A reply chain where the author participates outweighs a thousand silent likes. - **Retweets (+1.0) are only 2x a like.** Most growth advice overrates RTs. A single reply is worth ~27 likes. "Like and RT" strategies optimize the wrong signal. - **Dwell time (+10)** is high-value and underused — content that holds attention in-thread for 2+ minutes is rewarded almost as much as a reply. - **Report (-369x) and block/mute (-74x)** are account killers. One report event outweighs roughly 738 likes' worth of positive signal. --- ## Stage 3: Filter and Serve Blocked/muted accounts removed, in-network/out-of-network balanced, ads mixed in, feed served. --- ## TweepCred (Reputation Gate) A hidden 0–100 reputation score. Below 65 = only 3 of your tweets even enter the candidate pipeline, regardless of content quality. Built from account age, follower ratio, engagement quality, and interaction with high-quality accounts. Functions as an algorithmic credit score that gates everything upstream of content quality. - 30 days minimum account age for full scoring - 0.6 = maximum healthy following/followers ratio --- ## Tweet Lifecycle - **360-minute half-life.** Tweets decay fast. - **First 2 hours decide the tweet's fate.** Early engagement velocity determines total reach — this is why the "what should I post RIGHT NOW" problem matters more than content quality alone. --- ## Premium (Blue) Boost 4x in-network boost, 2x out-of-network. As of last check, adoption is low (~0.26%), making it a disproportionate lever for anyone using it. Premium users are also paid based on engagement from other Premium users, so writing for a Premium-heavy audience matters more than writing for raw volume. --- ## 2026 Shift: Grok/Grox Content Understanding The new system eliminated most hand-coded features and learns purely from engagement patterns — old tricks like hashtag stuffing and "optimal posting time" hacks matter much less now. Grox (content-understanding layer) reads and scores content quality directly: - **AI slop fails twice.** Grox deprioritizes it on content signals, and humans don't reply to generic content, so it never earns the one signal that actually matters (replies). - Musk has stated the algorithm reads every post — 100M+/day — meaning an LLM is evaluating your tweet before a human ever sees it. Feeding it generic LLM output is exactly the pattern it's trained to bury. - **Promptable feeds** are live: users can tell Grok "more startup content, less politics," making topical/niche consistency a direct ranking factor, not just a growth-hack heuristic. --- ## Key Numbers (Cheat Sheet) ``` 150x — Value of author reply-back vs a like 75 — Weight of reply engaged by author 13.5 — Weight of a reply 0.5 — Weight of a like (near-worthless alone) 65 — TweepCred threshold (below = only 3 tweets considered) 360 min — Tweet half-life 50/50 — In-network vs out-of-network feed split 145,000 — Number of SimClusters 1,500 — Candidate tweets per feed refresh 4x/2x — Premium boost (in-network/out-of-network) -369 — Report penalty multiplier -74 — Block/mute penalty multiplier 0.6 — Max healthy following/followers ratio 30 days — Minimum account age for full scoring ``` --- ## Product Implications (for a tweet/reply generator) 1. **Order of build priority: reply engine → trend reader → tweet drafter.** Replies are worth 27–150x a like, and out-of-network reach requires borrowed social proof from engagement, not original posting alone. The "generate tweets" feature is the least algorithmically valuable of the three, despite being the obvious MVP. 2. **Niche discipline is now a hard ranking factor**, not just aesthetic advice — SimCluster placement and promptable-feed targeting both depend on topical consistency. 3. **Any generated content must avoid AI-slop patterns** — Grox scores content quality directly, so generic phrasing gets buried before it can even earn replies. 4. **The algorithm updates every 4 weeks.** A static playbook decays within a month. Any product built on this needs to re-derive its rules from the live repo, not hardcode this doc permanently. 5. **Guardrails that must hold**: no automated follow/unfollow (triggers 3-month shadowban), no auto-posting without human approval (bot detection watches for behavioral patterns), no volume spikes (gradual increases only). --- ## Sources - X algorithm (2026, current): github.com/xai-org/x-algorithm - X algorithm (2023, deprecated — do not build on this): github.com/twitter/the-algorithm - X Engineering Blog: blog.x.com/engineering/en_us/topics/open-source/2023/twitter-recommendation-algorithm - TweepCred source (legacy repo, concept still applies): github.com/twitter/the-algorithm/tree/main/src/scala/com/twitter/graph/batch/job/tweepcred - SimClusters (legacy repo, concept still applies): github.com/twitter/the-algorithm/tree/main/src/scala/com/twitter/simclusters_v2 - Knight First Amendment Institute analysis: knightcolumbia.org/blog/twitter-showed-us-its-algorithm-what-does-it-tell-us