How the X Algorithm Actually Works in 2026

The system is built to maximize long-term user retention. The three heaviest signals are engagement velocity, author reputation (Tweepcred), and SimClusters community fit.

Example: A creator with 8k followers posted a sharp contrarian take on AI agents. It received 312 likes and 48 high-quality replies in the first 47 minutes. The post reached 41,000 people in the first hour because of strong early velocity and good graph signals.

2. What Gets You Suppressed

Hard filters (Visibility Library) remove blocked, muted, NSFW, or spam content. Soft downranking happens from low engagement velocity, poor reputation, repetitive behavior, or sudden changes in posting patterns.

Example: An account increased posting from 3x/day to 11x/day with repetitive AI-generated threads. Reach dropped 87% within 9 days as the model flagged low-quality repetitive behavior.

3. Five Unexpected Signals

1. Reply thread depth from high-reputation accounts

Explainer: When respected accounts reply to your post with thoughtful, multi-tweet responses, it sends a very strong signal to the algorithm.

So what: Prioritize earning high-quality replies from good accounts over chasing raw like counts. One strong reply thread from a respected account can be worth more than 200 average likes.

2. Dwell time (how long people actually read)

Explainer: The algorithm measures how much time users spend reading your content, not just whether they liked or reposted it.

So what: Longer, more substantive posts that keep people reading for 15–30+ seconds often outperform short, high-like-velocity posts. Depth can beat speed.

3. Cross-community performance (SimClusters)

Explainer: SimClusters groups users into interest communities. Posts that perform well across multiple different communities get an extra boost.

So what: Content that bridges niches or appeals to overlapping communities tends to get wider distribution than content that only resonates in one narrow group.

4. Consistency of voice and cadence

Explainer: The algorithm builds a mental model of your account based on your historical posting patterns, tone, and frequency.

So what: Sudden big changes in volume, topic, or tone can cause temporary downranking while the model re-learns your account. Consistency helps maintain distribution.

5. Who engages with you (graph strength matters more than volume)

Explainer: The algorithm looks at the reputation and graph connections of the people engaging with you, not just the raw number of engagements.

So what: 10 high-quality engagements from well-connected accounts in your niche are often more valuable than 200 random likes from low-reputation accounts.

4. AI Slop Detection — Characteristics with Examples

Characteristics of AI Slop (with examples)

1. Generic phrasing that appears across many posts
Example: “In today’s rapidly evolving AI landscape, it’s crucial to stay ahead of the curve and embrace innovation.” This sentence (or close variations) appears in hundreds of low-effort posts.

2. Lack of specific, lived experience or original data
Example: “AI agents will change everything.” No numbers, no specific failure cases, no real experiments mentioned.

3. Repetitive rhythm and pacing
Example: Every sentence follows the same structure: “It’s important to…”, “One key thing is…”, “Another aspect to consider is…”. The pattern becomes obvious after 3–4 sentences.

4. Overly polished but hollow language
Example: “This represents a paradigm shift in how we conceptualize human-AI collaboration.” Sounds impressive but contains no actual insight.

5. Low engagement depth (few meaningful replies)
Example: A 12-tweet thread gets 180 likes but only 7 replies, most of which are “Great thread!” or “Thanks for sharing.”

Characteristics of High-Quality Human Content (with examples)

1. Specific details only someone who did the work would know
Example: “When we tested the agent on 47 real customer support tickets, it failed on 19 of them because it kept trying to use chain-of-thought on conversations longer than 8,400 tokens.”

2. Personal voice, quirks, and point of view
Example: “I’ve been skeptical of fully autonomous agents for a while, but last week one of them actually surprised me. Here’s what it got right (and the three things it still completely botched).”

3. Original data, screenshots, or experiments
Example: A creator posts actual screenshots of their agent’s reasoning trace + the exact prompt they used and the failure output.

4. Imperfect but authentic writing
Example: “This might be a hot take, but I think we’re over-indexing on agents right now. I’ve been building with them for 4 months and I keep running into the same wall…”

5. Strong reply velocity and conversation depth
Example: A thread gets 180 likes but also 47 replies, many of which are thoughtful disagreements or people sharing their own data. The conversation continues for days.

5. 12-Month Growth Strategy — Weekly Cadence

Recommended Weekly Mix

Correct vs Incorrect Examples

Correct: You post a 14-tweet thread breaking down a real experiment. You spend the next 3 days replying thoughtfully to every good reply. Velocity stays strong for 5 days.

Incorrect: You post 4 short AI-generated “tips” in one day, then go silent for 48 hours. The algorithm sees low engagement velocity and repetitive structure.

Correct: You reply to a larger account with a specific counter-example + data. That reply gets quote-tweeted, bringing in new followers who already trust you.

Incorrect: You reply “Great point!” or generic agreement to 40 accounts. No new distribution.

6. How to Use AI Correctly

Correct Use of AI

Example: You feed the agent your last 25 posts and ask it to identify patterns in what performed well. You then use those insights to brainstorm 6 new angles for your next thread, which you write yourself in your own voice.

Incorrect Use of AI

Example: You ask ChatGPT to “write a viral thread about AI agents” and post it with only minor edits. The thread gets low engagement and the algorithm begins downranking similar content from your account.

Use AI for research, pattern extraction, and outlining. Always rewrite everything in your own voice before posting. Never post raw AI output at scale.

7. Raw Code Perspective

Codebase Overview

The open-source X recommendation algorithm repository contains approximately 7,700 files and over 154,000 lines of code. It is primarily written in Scala, with supporting components in Java, Python, and Thrift definitions. The codebase is organized into many independent services rather than one monolithic application.

Key components include Home Mixer (orchestration), SimClusters (community detection), Visibility Library (filtering), Real-Graph, UTEG (graph traversal), and multiple candidate generation pipelines.

Key Input Variables to Going Viral

1. Engagement Velocity
Modeled on: Likes, replies, reposts, and time spent reading in the first 30–60 minutes.
Explanation: The algorithm heavily rewards posts that generate rapid, high-quality engagement shortly after being posted. Strong early velocity signals to the model that the content is relevant and worth showing to more people.

2. Author Reputation (Tweepcred)
Modeled on: A PageRank-style score reflecting how trusted and high-quality an account is across the network.
Explanation: Accounts with higher reputation scores get their content distributed more widely and reliably. Reputation is built over time through consistent quality and positive engagement patterns.

3. Community Similarity (SimClusters)
Modeled on: How well the post matches the interest communities the viewer belongs to.
Explanation: The algorithm uses community detection to understand user interests. Posts that align with a viewer’s communities are more likely to be shown, even if the viewer doesn’t follow the author.

4. Graph Features
Modeled on: Connections between the viewer and the author (how many of the viewer’s follows have engaged with the author before).
Explanation: The algorithm looks at the social graph. If people the viewer follows have engaged with the author, the post is considered more relevant and trustworthy.

5. Content Quality & Visibility Signals
Modeled on: Trust & safety scores, spam detection, NSFW flags, and quality heuristics.
Explanation: Even high-velocity posts can be suppressed if they trigger spam, abuse, or low-quality filters. These act as hard gates after initial ranking.

6. Dwell Time & Reply Depth
Modeled on: How long users spend reading the post and the quality/depth of replies it receives.
Explanation: The algorithm values genuine attention and conversation. Posts that keep people reading and spark thoughtful replies receive stronger distribution than those that only get quick likes.

score = (engagement_velocity × 0.35) + (author_reputation × 0.25) + (community_similarity × 0.20) + (graph_features × 0.15) + (content_quality × 0.05)

8. How to Use Nous Research + X Search to Optimise Growth

Modern AI agents (like Hermes/Nous) combined with X search capabilities give creators a significant advantage. Here’s how to use them effectively.

1. Pattern Extraction from Your Own Content

Feed the agent your last 30–50 posts and ask it to identify what performed best and why. This is far more accurate than generic “viral hook” advice.

2. High-Signal Conversation Discovery

Use X search tools to find active, high-quality conversations in your niche. Look for threads with strong reply depth from good accounts rather than just high like counts.

3. Reply Strategy Optimisation

Paste a target thread into the agent and ask it to generate 8–10 sharp, on-brand reply angles. Then rewrite them in your voice. This dramatically increases the quality of your daily engagement work.

4. Competitive & Audience Research

Analyze what accounts in your space are posting and which posts are getting the strongest engagement. Identify gaps and underserved angles.

5. Content Ideation with Context

Use research tools to understand current sentiment and emerging topics in your niche, then combine that with your own voice and data for original content.

Pro move: Keep a running “voice document” of your best-performing posts. Feed it to the agent weekly so it stays calibrated to your actual tone and strengths.

9. Account Click-Through Rate & Subscriber Conversion

Getting impressions is only the first step. What matters is how many people click through to your profile and then convert into subscribers.

Benchmarks

Good Example

A creator posts a sharp, specific thread. The post gets 28,000 impressions. 2,100 people click through to the profile (7.5% CTR). Of those, 168 subscribe (8% conversion). This is considered strong performance.

Bad Example

A generic “10 tips” thread gets 45,000 impressions. Only 720 people click through to the profile (1.6% CTR). Of those, just 11 subscribe (1.5% conversion). The content attracted curiosity but failed to build trust or desire to follow long-term.

What Drives Better CTR and Conversion

1. Strong, specific bio that clearly states value

Good: “I build AI agents that actually work. Sharing real experiments, failures, and what I’ve learned running them in production.”

Bad: “AI • Tech • Future” or “Helping people navigate the AI revolution ✨”

2. Pinned post that showcases your best work

Good: A pinned 18-tweet thread with strong engagement and clear value (e.g. “I tested 12 AI agent frameworks. Here are the 3 that actually survived production.”)

Bad: A pinned post that is just a link to a newsletter or a generic welcome message with no substance.

3. Consistent voice and high-quality replies visible on your profile

Good: When someone visits your profile, they see thoughtful, specific replies to other accounts in your niche. This builds immediate credibility.

Bad: Your replies are mostly short, generic comments (“Great point!”, “This is huge”, “Thanks for sharing”). Visitors see low-effort engagement and leave.

4. Posting content that attracts the right audience (not just high volume)

Good: You consistently post detailed, specific content about AI agent failures in production. The people who click through are already interested in that topic and convert at a high rate.

Bad: You post broad “AI tips” content that attracts a wide but low-intent audience. Many people click through out of curiosity but don’t see enough value to subscribe.

10. How to Create McKinsey-Style Insightful Content on X

McKinsey-style communication (heavily influenced by Barbara Minto’s Pyramid Principle) is one of the highest-signal formats you can use on X. It is structured, data-driven, and action-oriented.

Core Principles

Practical Application on X

Good (McKinsey-style):
“Most AI agents fail in production for one reason: they cannot maintain context beyond 8k tokens.

We tested 47 real customer support workflows. Here’s what broke:
• 19 failed due to context window limits
• 14 failed due to poor tool-use planning
• 9 failed due to hallucinated tool outputs

The fix: Add an explicit ‘context checkpoint’ step every 6k tokens. This single change raised success rate from 38% to 71%.”

Bad (Unstructured):
“AI agents are really interesting but there are still a lot of challenges. Context windows are limited and they sometimes make mistakes with tools. It’s important to test thoroughly and think about how to structure prompts carefully…”

Recommended Structure for Threads

This style performs extremely well because it respects the reader’s time while delivering high density of insight — exactly what the algorithm and high-quality audiences reward.

Appendix A: Using Lists for Strategic Growth

Lists are one of the most underused but powerful tools for focused growth and engagement on X. Here’s how to build and use them strategically.

1. Same-Follower List (Peer List)

2. 5x Follower List (Next-Level List)

3. 10x Follower List (Aspirational List)

Purpose: Observe and occasionally engage with significantly larger accounts.

1. What the Algorithm Actually Optimizes For