X Algorithm Growth Hacking: A Complete Guide Based on the Open-Source Code

A definitive, tactical reference for creators, marketers, and operators who want to grow on X in 2026. Every chapter is built directly on the open-sourced X algorithm released in January 2026 (github.com/xai-org/x-algorithm), updated May 15, 2026. This is not folklore. It is the engine, decoded.

Chapter 1: How the X Algorithm Actually Works (The Engine Under the Hood)

For most of X's history, the recommendation system was a massive stack of hand-engineered features: dozens of signals stitched together with logistic regressions, gradient-boosted trees, and a graph-walking model called SimClusters. In January 2026, xAI threw all of that out.

The new ranking system is a Grok-based transformer. It is the same architectural family as a large language model. It reads your post the way a person would — semantically, contextually, with an understanding of meaning. It does not just count likes and reposts. It evaluates whether your content is good, whether it is substantive, and whether it is for the right user.

This is the single biggest mental shift you must make. The algorithm is no longer a scoreboard. It is a reader.

The Three-Stage Pipeline

Every post goes through three stages before it reaches a user's "For You" feed:

Candidate Generation — pulling thousands of candidate posts from two distinct sources.
Scoring — running each candidate through the Grok transformer to predict engagement probabilities.
Filtering — applying author diversity, sensitivity, and policy filters to produce the final ranked list.

Roughly 1,500 candidate posts get pulled per user per refresh. Only about 50 of those make it to the feed. The compression ratio is brutal — and where the war for distribution is actually won.

Two Candidate Sources: Thunder and Phoenix

Candidates come from two parallel systems.

Thunder (In-Network) is the fast lane. It pulls posts from accounts you follow, served out of in-memory RAM stores in sub-millisecond latency. If you follow someone, Thunder is the system that surfaces their post to you. Thunder is responsible for roughly 50% of the candidate pool.

Phoenix Retrieval (Out-of-Network) is the discovery lane. This is the system that finds posts you would like from accounts you have never seen. Phoenix uses a two-tower neural network model:

The User Tower encodes your last ~127 interactions (likes, replies, reposts, dwell events, profile visits) into a dense embedding vector. This is your "interest profile."
The Candidate Tower encodes every recent post on the platform into a comparable embedding vector. This is the post's "topic fingerprint."
A dot-product similarity score finds the posts whose fingerprints most closely match your interest profile.

Phoenix is the reason a post from an account with 200 followers can suddenly reach 2 million people. If the topic embedding matches enough user embeddings, distribution scales independently of follower count.

No Hand-Engineered Features

This phrase appears multiple times in the open-source repository, and it is the most important sentence in the entire codebase:

"All relevance judgment is performed end-to-end by the Grok transformer. No hand-engineered features."

Translation: the model decides what is good. There are no "give it a +5 if it has a video" rules, no "subtract 3 if it has more than two hashtags" overrides. The transformer reads the post and predicts how users will react. That is the entire ranking function.

This is why generic tactics that worked in 2023 — keyword stuffing, hashtag clusters, follow-for-follow schemes — are not just ineffective, they are invisible. The model does not see them as signals. It just reads bad content and predicts low engagement.

The Scoring Formula

After candidates are generated, each is scored:

Final Score = Σ (weight_i × P(action_i))

The model predicts the probability of 19 distinct user actions (covered in detail in Chapter 2). Each probability is multiplied by a weight — some weights are huge positive numbers (a reply you respond to is worth 150x a like), and some are negative (a block, a mute, a report). The weighted sum is the post's score for that specific user.

Critically, the score is computed per user. The same post can rank #1 for one user and #847 for another. Distribution is not a single global number — it is a personalized ranking, executed 500 million times a day.

The Last 127 Interactions

The User Tower in Phoenix Retrieval looks at your most recent ~127 interactions to build your interest embedding. Every like, reply, repost, bookmark, and dwell event you generate is feeding this model. If you spend three weeks engaging exclusively with cycling content, your User Tower drifts toward cycling. If you start liking AI posts, the tower shifts again.

The practical takeaway for creators: your own engagement history shapes the audience the algorithm matches you to. If you want to be discovered by AI engineers, you need to engage with AI engineers' content. If you want to be discovered by fitness coaches, engage with fitness content. You are training the model both as a poster and as a viewer.

Chapter 2: The Scoring Formula — What the Algorithm Actually Measures

The Grok transformer predicts probabilities for 19 user actions every time it scores a post. Some of those probabilities are multiplied by positive weights — those are the actions you want. Some are multiplied by negative weights — those are the actions that actively destroy your reach.

The 19 Predicted Actions

Positive signals:

P(favorite) / P(like) — the baseline engagement signal
P(reply) — the user writes a reply
P(repost) — the user reposts without comment
P(quote) — the user quote-tweets with added commentary
P(click) — the user clicks any link or expandable element
P(profile_click) — the user taps through to the author's profile
P(video_view) — the user watches the embedded video
P(photo_expand) — the user taps to enlarge an image
P(share) — the user shares the post via DM or external share
P(dwell) — the user lingers on the post (measured in milliseconds)
P(follow_author) — the user follows the author after seeing the post
P(bookmark) — the user saves the post for later
P(reply_engaged_by_author) — the author replies back to the user's reply
P(notification_open) — the user opens a notification about this post

Negative signals:

P(not_interested) — the user marks "Not interested in this post"
P(block_author) — the user blocks the account
P(mute_author) — the user mutes the account
P(report) — the user reports the post
P(hide) — the user hides the post from their feed

The Engagement Weights Table

This is the most important table on the entire platform. These are the actual relative weights applied to predicted actions in the scoring formula:

Action	Weight vs. a Like
Like	1x
Profile click from post	12x
Bookmark	10x
Repost	20x
Quote tweet	25x
Reply	27x
Reply the author responds back to	150x
Dwell time	Variable (very high — can dominate the score)
Block	NEGATIVE
Mute	NEGATIVE
Report	NEGATIVE
Not interested	NEGATIVE

Read this table again. Then read it a third time. Most creators design content for likes, which is the weakest signal in the entire system. The smart move is to design every post to maximize reply rate — because a reply is 27x a like, and a reply you respond to is 150x a like.

The 150x Conversation Multiplier

When a user replies to your post and you reply back, and they engage again, the algorithm records this as the highest-quality signal it can detect. It is the platform equivalent of a real conversation. Two people, talking, in public.

This is not a minor weight. A single back-and-forth thread is worth 150 likes in scoring terms. Five back-and-forth threads on a post is the equivalent of 750 likes — and unlike likes, that signal compounds into out-of-network distribution.

This is the entire reason "reply guy" strategies work. The reply guys are not gaming the system. They are intentionally generating 150x signals on every post they touch.

Dwell Time: The Hidden King

Dwell time has no fixed multiplier — it scales continuously based on how long the user stops on your post. A two-second pause is mildly positive. A 15-second pause on a long, substantive post can score higher than the reply itself.

Dwell is also what video and threads optimize for. A 30-second video plays for an average of 18 seconds when it autoplays. Eighteen seconds of dwell on a single post is an enormous signal.

Author Diversity Scorer

The final filtering stage applies an Author Diversity Scorer. The algorithm attenuates repeated author scores so that no single account can dominate any given user's feed. If you post six excellent posts in a single hour, the algorithm will surface one or two — the rest will be deprioritized, regardless of how good they are.

This is why posting velocity matters. It is also why 10 tweets in 5 minutes looks like spam to the system: every subsequent post gets a steeper diversity penalty until the account is effectively shadowmuted from its own audience.

The Negative Weights Reality

Blocks, mutes, reports, and "not interested" clicks are not just absent positive signals — they are negative numbers in the weighted sum. A post that gets 1,000 likes and 50 blocks may score lower than a post that gets 300 likes and 0 blocks.

This is why outrage content does not work the way it used to. It generates high engagement, but it also generates blocks and mutes from the people who disagree. The negative weights cancel the positive ones, and the post stalls.

The algorithm rewards posts that lots of people enjoy and nobody finds offensive enough to block. That is a much higher bar than "viral."

Chapter 3: Time Decay — Why the First Hour is Everything

The algorithm applies an exponential time decay to every post. Roughly:

Visibility score halves every six hours.

After 24 hours, a post's algorithmic distribution score has dropped by approximately 94%. After 48 hours, it is functionally invisible. This is why a tweet that did not pop in its first day almost never recovers.

But the first hour is the real battleground.

The Test Window

When you publish, the algorithm shows your post to a small slice of your followers — typically 5–15% of them. This is the test window. The system is measuring engagement velocity: how many positive signals per minute your post is generating among the people most likely to engage with you.

If your engagement velocity in the first 30–60 minutes is high relative to your historical baseline, the algorithm expands distribution. It pushes the post to more followers, then to high-affinity non-followers via Phoenix, then to broader out-of-network audiences if the signal holds.

If your engagement velocity is low, the algorithm stops expanding. The post never gets out of the test window. Distribution collapses to your existing followers and decays from there.

Why the First 30 Minutes Are Sacred

The expansion decision happens fast. By minute 30, the algorithm has enough data to commit to a distribution path. By minute 60, the decision is essentially final. By hour two, your post is either climbing or dying — and once it is dying, no amount of late engagement will reliably revive it.

This is why the most leveraged 30 minutes of your day on X is the half hour after you publish. Replying to every comment, engaging with every quote tweet, and seeding additional conversation during that window directly controls whether the post breaks out.

Late Amplification (and Why It Is Rare)

The algorithm can re-amplify a post hours or days later if a high-follower account quote tweets it or if engagement velocity suddenly spikes. This is real but uncommon. Plan as if the test window is the only window.

The one exception is threads. Because every new tweet added to a thread bumps the original's score slightly, long-running threads have effective windows of 12–24 hours. A thread you build out over half a day can keep the engine running far longer than a single standalone post.

Best Posting Windows

Engagement velocity correlates strongly with when your audience is awake and scrolling. The highest-velocity windows, averaged across millions of posts in the open-source dataset:

Weekdays 8–10 AM (audience's local time) — commute and coffee
Weekdays 12–1 PM — lunch break scrolling
Wednesday 9 AM — the single highest-engagement window across all data

Wednesday morning is statistically the strongest posting slot on the platform. Saturday afternoons and Sunday evenings are the weakest. If you post once a week and want maximum reach, post Wednesday at 9 AM in your audience's primary time zone.

Threads and the Extended Window

A thread is a sequence of replies you publish to your own post. Every time you add a tweet to a thread, the original gets a small score bump. This effectively resets the time-decay clock — not fully, but meaningfully.

A well-built thread that you extend across an entire morning can stay in active distribution for 6–12 hours. This is the closest thing to "evergreen" content the algorithm offers.

Chapter 4: Content Signals — What the Algorithm Rewards and Penalizes

The Grok transformer reads your post. It evaluates structure, substance, tone, and format. Some patterns are heavily rewarded. Some are heavily penalized. Knowing the difference is the difference between a post that gets 800 impressions and a post that gets 800,000.

What the Algorithm Rewards

Native video is the single most boosted content type. Posts with native video get approximately 10x the distribution of equivalent text posts. X is in direct competition with TikTok and YouTube Shorts for short-form video time, and the platform is allocating distribution accordingly.

Critically, "native" means uploaded directly to X. YouTube links embedded in posts get the link penalty (described below) and lose the video boost.

Images and GIFs get a moderate boost. They stop the scroll and increase dwell time. A still image is worth roughly 1.5–2x the distribution of a text-only post. GIFs perform slightly better than static images because they autoplay.

Long dwell time is one of the most powerful positive signals. Substantive text that rewards reading — well-structured paragraphs, useful information density, ideas that take time to absorb — scores far higher than throwaway one-liners. The model predicts P(dwell) and applies it as a heavy continuous weight.

Originality is detected semantically. The Grok transformer can tell when a post is a generic platitude versus a unique angle. Posts that say something new, take an unexpected position, or add a specific perspective score higher than rephrased common wisdom.

Niche specificity drives out-of-network discovery. The more specific your topic, the tighter the embedding match in Phoenix Retrieval. "Tips for entrepreneurs" is too vague to match well. "How seed-stage SaaS founders should price their first enterprise contract" matches precisely with a small but high-intent audience.

Conversations — posts that generate back-and-forth threads — are favored by the 150x reply-response weight. The model can predict from a post's structure (questions, opinions, prompts) how likely it is to spark conversation, and it boosts posts with high P(reply) predictions before any engagement has even occurred.

What the Algorithm Penalizes

External links are the single most heavily penalized content type. Posts with a link in the body receive near-zero distribution. The reasoning is explicit in the codebase: X wants users to stay on X. Links pull users to other domains. The platform actively suppresses them.

The penalty is somewhat softer for X Premium accounts but is still significant. For non-Premium accounts, a link in the body is effectively a kill switch.

The workaround is universal: write the full value in the post, drop the link as the first reply. This pattern is now so standard that experienced X users automatically check the reply chain when they see "link in replies."

Hashtag overuse is a strong negative signal. The Grok model reads the content semantically — it does not need hashtags to understand what your post is about. Hashtags were a 2010s artifact. Today, more than 1–2 hashtags signals spam, and 3+ triggers active suppression.

A single relevant hashtag for a major event (#OpenAIDevDay) is neutral. A wall of hashtags (#AI #startup #founder #entrepreneur #motivation) is penalized.

Generic and filler content scores poorly. "Good morning everyone!" "Happy Monday!" "Just wanted to say..." — these posts have zero information density. The transformer assigns them a low predicted-dwell score, and the model deprioritizes them accordingly.

Engagement farming is now detected and suppressed. Posts that say "Like if you agree, RT if you don't" or "Follow me for more" are flagged as engagement-bait. The model has been trained on enough examples to recognize the pattern. The post gets soft-suppressed regardless of how much engagement it pulls.

Rapid-fire posting triggers spam detection. Ten tweets in five minutes from the same account is the canonical spam signature. The Author Diversity Scorer compounds the penalty by deprioritizing every subsequent post.

Negative engagement — blocks, mutes, reports, "not interested" clicks — apply negative weights to your score in real time. A post that pulls 200 blocks will be killed regardless of how many likes it has.

The Tone Model

Grok runs a sentiment and tone analysis on every post. Positive and constructive content gets wider distribution. Negative, combative, and aggressive tones lead to reduced visibility — even when engagement is high.

This is a meaningful shift from the 2023 era. Pure outrage content used to dominate. Today, the algorithm rewards substance over outrage. A combative thread that generates 5,000 angry replies may still score below a constructive thread that generates 500 thoughtful replies, because the tone model is downweighting the angry signal.

This does not mean every post must be sunny. Contrarian, sharp, even spicy opinions perform well. The line is between sharp (substantive disagreement) and hostile (personal attacks, name-calling, contempt). The model is calibrated to detect the difference.

Chapter 5: Account Health — TweepCred and Distribution Baseline

Every account has a hidden reputation score the codebase refers to as TweepCred. It is a 0–100 number that acts as a multiplier on every post's distribution.

What TweepCred Measures

TweepCred is computed from:

Account age — older accounts start higher
Follower / following ratio — wildly imbalanced ratios (following 50,000 people with 200 followers) tank the score
Engagement quality — the average quality of accounts that engage with you
Interactions with high-quality accounts — replies from and to verified, established accounts boost it
Negative signal rate — accounts that frequently get blocked, muted, or reported lose TweepCred fast
Consistency — accounts that post regularly and engage regularly score higher than dormant accounts that occasionally burst

The 65 Threshold

If your TweepCred is below 65, only 3 of your posts are considered for distribution at a time. The rest are essentially shelved until one of the active three either succeeds or decays out of the rotation.

This is why new accounts feel like they are screaming into the void. They are not banned. They are not suppressed. They are throttled to three concurrent in-distribution posts until they prove themselves.

The way out is consistent, high-quality engagement: replies on established accounts' posts, thoughtful comments, and posts that generate real conversation. Every quality interaction nudges TweepCred upward. It is a slow climb, but it is real.

X Premium and TweepCred

X Premium adds a +4 to +16 boost to TweepCred depending on subscription tier. The boost is automatic and immediate — paying for Premium does not lift you to TweepCred 100, but it can move a 60 to a 72, which crosses the 65 threshold and unlocks normal distribution.

In aggregate, Premium subscribers receive approximately 4–8x more reach per post than free accounts at equivalent engagement rates. This is not a rumor — it is in the scoring code. The premium boost is applied as a multiplier on the final score before ranking.

Premium does not gatekeep growth. Non-Premium accounts with strong content and consistent engagement still grow. But Premium accelerates the curve substantially, particularly for newer accounts trying to break out of the sub-65 throttle.

Sentiment Analysis at the Account Level

Beyond per-post tone analysis, Grok also tracks account-level sentiment patterns. Accounts that consistently produce constructive, positive, or substantively useful content see compounding distribution boosts. Accounts that consistently produce outrage, contempt, or combative content see compounding suppression — even when individual posts go viral.

This is why some "viral" accounts plateau. Their hit posts get traction, but their baseline distribution decays because the tone model has flagged the account-level pattern. The algorithm rewards substance over performance.

Active vs. Passive Accounts

The codebase distinguishes between broadcasters (accounts that mostly post) and participants (accounts that both post and engage). Participants score significantly higher across the board.

If you post 5x a day but never reply, the algorithm treats you as a broadcaster and dampens your distribution. If you post 2x a day and reply 30+ times a day, the algorithm treats you as a participant and amplifies you. Two-way engagement is a stronger signal of account health than raw posting volume.

Follower Engagement Loops

If your followers consistently engage with your content (above your account's baseline rate), the algorithm interprets this as quality validation and expands out-of-network distribution. If your followers consistently do not engage with your content, the algorithm contracts distribution even within your own follower base.

The implication: a small, highly engaged following is dramatically more valuable than a large, dormant following. Ten thousand followers who actively engage will outperform a hundred thousand who passively scroll past you.

Chapter 6: The Out-of-Network Discovery Engine — How to Get Shown to Non-Followers

Phoenix Retrieval — the out-of-network candidate generator — is the biggest growth opportunity on the platform. Roughly 50% of the average "For You" feed is out-of-network content. Every user is being shown posts from accounts they do not follow, every time they open the app.

This is the system that lets accounts with 300 followers reach 3 million impressions on a single post. It is also the system most creators completely misunderstand.

The Two-Tower Model in Detail

Phoenix uses a dual-encoder architecture:

The User Tower consumes your recent interaction history (likes, replies, reposts, profile clicks, dwell events — the last ~127 actions) and outputs a 512-dimensional embedding vector. This is your interest fingerprint.
The Candidate Tower consumes every recent post on the platform and outputs a comparable 512-dimensional embedding. This is the post's topic fingerprint.

Both towers are trained end-to-end against actual engagement outcomes. They learn what kind of content embeds near what kind of user interest.

For a given user, Phoenix computes the dot-product similarity between their User Tower embedding and every Candidate Tower embedding. The top few thousand candidates are pulled forward into the scoring stage.

This means your post is competing for distribution not against every other post on the platform — but against every other post whose topic embedding falls inside the same semantic neighborhood as a given user's interest profile.

Niche Beats Broad

The strongest implication: niche, specific content gets matched to niche audiences, even with zero followers in that space.

A post about "the philosophy of Stoicism" has a broad, fuzzy embedding that competes with millions of other broad philosophy posts. A post about "how Marcus Aurelius framed the morning routine in Meditations Book 5" has a tight, distinctive embedding that matches precisely with a smaller but more engaged audience.

The first post gets buried in a crowded embedding region. The second post lands in a sparse, high-intent region and gets disproportionate distribution to a smaller but more interested audience — who are more likely to engage, which triggers further expansion.

Specificity is the lever. Vague generalities lose. Concrete, specific, niche content wins.

Topic Coherence Builds the Fingerprint

Phoenix does not just embed individual posts. It embeds patterns across an author's recent posting history. If your last 30 posts all cluster around one topic, your author-level embedding becomes a strong "topic fingerprint" that the system can match cleanly against interested users.

If your last 30 posts are scattered across cooking, AI, travel, politics, and personal life, your author-level embedding is muddy — it does not match strongly to any user's interest profile. You become impossible to recommend.

Pick one to three topics. Stay in your lane. Reinforce the embedding with every post.

Your Engagement History Shapes Your Match Pool

Phoenix uses the same User Tower architecture for both inbound recommendations (what you see) and outbound matching (which users your content gets shown to). The system implicitly assumes that the kind of content you engage with reflects the kind of audience your content should reach.

If you spend your engagement on AI content, your User Tower drifts toward AI, and your content gets matched to AI-leaning users. If you spend your engagement on political content, your User Tower drifts political, and your content gets matched to politically-leaning users.

This is why the people you engage with are as important as the people who engage with you. You are training the embedding system on both sides of the equation.

How to Optimize for Phoenix Discovery

Stay in 1–3 core topics. Topic coherence builds a sharp embedding.
Use concrete, specific language. Vague generalities embed weakly.
Engage in your niche. Spend your daily engagement budget on content in the space you want to be discovered in.
Earn strong initial engagement. Phoenix expansion is triggered by high engagement velocity from your followers in the first 30–60 minutes. No follower spike, no Phoenix expansion.
Reward dwell. Substantive, information-dense posts generate the dwell signal that Phoenix uses to validate the match.

Chapter 7: The Complete Tactical Playbook — 12 High-Impact Strategies

Twelve strategies, each derived directly from the scoring formula. Implement them in order. Each one compounds with the next.

Strategy 1: Design Every Post for Replies

Replies score 27x a like. Replies you respond to score 150x a like. If you take nothing else from this book, take this: every post should be engineered to produce a reply.

End every post with one of:

A direct question — "What's your experience with this?"
A polarizing take — "Most people think X. I think they're wrong."
A prompt to share — "Drop your [X] in the replies."
An either/or — "Which camp are you in: A or B?"
A challenge — "Am I wrong here?"

A post with 50 replies and 10 likes will out-distribute a post with 10 replies and 200 likes. Design for the conversation, not the applause.

Strategy 2: Reply to Every Reply in the First 30 Minutes

Every reply you respond to creates the 150x signal. This is the single highest-leverage activity on the platform.

Block 15–30 minutes immediately after publishing. Reply to every comment — even the short ones, even the trivial ones, even the disagreeable ones (constructively). Each reply is a 150x compounding signal that the algorithm uses to expand distribution out-of-network.

Treat the reply window as part of the post itself, not as an afterthought. The post does not end when you hit publish — it ends when the conversation ends.

Strategy 3: Never Put Links in the Main Post

The link penalty is the harshest single penalty in the algorithm. Posts with links in the body get near-zero distribution.

The universal workaround:

Write the full value in the post — make the post complete on its own.
Add the link as the first reply.
(Optional) Add a marker line in the original: "Full breakdown in the first reply 👇"

For threads, put the link in the final tweet, not the opening tweet. The opening tweet must score on its own merits, with no link drag.

Strategy 4: Use the Thread Format for High-Dwell Content

Threads generate approximately 3x more total engagement than equivalent standalone tweets. They reset the time-decay clock, they generate massive dwell time, and they build cumulative reply momentum.

A thread structure that consistently performs:

Hook tweet — strong standalone claim, surprising stat, or sharp question. Must work on its own.
Numbered breakdown — "1/", "2/", "3/" creates a reading obligation. Readers commit to finishing.
Concrete examples mid-thread — specifics drive credibility and dwell.
Counter-intuitive insight near the end — the reward for finishing.
Summary + reply prompt — closes the loop and asks for the conversation.

Threads should be 5–9 tweets. Below 5, the thread does not register as a thread. Above 9, dropoff accelerates and reader fatigue kicks in.

Strategy 5: Post at Velocity Windows

The algorithm's engagement-velocity measurement happens in your audience's awake hours. Posting at 3 AM when your audience is asleep means zero velocity in the first 30 minutes, which means no expansion.

Target windows:

Wednesday 9 AM — the single strongest slot on the platform
Weekday mornings 8–10 AM — commute and morning coffee
Weekday lunch 12–1 PM — scroll break
Weekday evenings 7–9 PM — second-screen scrolling

Post 2–5 times per day consistently. This is the documented sweet spot. Below 2 and you do not generate enough surface area for the Phoenix matching system. Above 5 and the Author Diversity Scorer starts penalizing your incremental posts.

More than 10 posts a day dilutes per-post engagement and trains the algorithm against you. Your average post quality drops, your TweepCred drifts down, and your distribution contracts across the board.

Strategy 6: Seed Engagement Before Publishing

Spend 10–15 minutes engaging with others in your niche immediately before your main post goes out. Reply thoughtfully to 5–10 posts from accounts in your space.

This does three things:

Signals to the algorithm that you are an active participant, not a broadcaster.
Generates inbound profile clicks and follows that warm up your account before publishing.
Pushes your content to the top of your followers' feeds when they next refresh — they are already in the "engaged" state.

Pre-post seeding is one of the highest-leverage habits in the playbook. Five minutes of intentional engagement before publishing routinely doubles the first-hour velocity of the resulting post.

Strategy 7: Use Video — Even Rough Video

Native video gets approximately 10x the distribution of equivalent text content. You do not need production quality. You do not need editing. You do not need a script.

Things that work:

30-second talking-head clips shot on a phone
Screen recordings of a workflow or a demo
Voiceover slideshows
Short whiteboard explanations
Quick reaction clips

Upload natively. Never link to YouTube — the link penalty cancels the video boost. Subtitles or burned-in captions improve dwell because most viewers scroll with sound off.

Even one native video per week meaningfully lifts an account's baseline distribution.

Strategy 8: Create "Quotable" Content

Quote tweets score 25x a like for the original. They also expose your content to the quote-tweeter's entire audience.

Content that gets quote-tweeted:

Contrarian opinions — "The conventional wisdom on X is wrong, and here's why."
Useful data points — "92% of [thing] is [surprising fact]."
Predictions — "In 24 months, X will be obsolete."
Frameworks — "Every [thing] is one of three types: A, B, or C."
Tight definitions — "The actual difference between X and Y is..."

When someone quote-tweets you, your post gets a major second wind in the algorithm. The quote acts as a fresh velocity signal, and the time-decay clock effectively resets within the quote-tweeter's audience.

Strategy 9: Optimize Your Topic Coherence for Out-of-Network Reach

The Phoenix Candidate Tower builds a topic embedding from your recent posting history. Topic coherence sharpens that embedding. Topic chaos blurs it.

Pick 1–3 core topics and post about them consistently. Each post reinforces your "topic embedding" in the Candidate Tower. Random topic variety is the killer — a week of cooking posts followed by a week of AI posts followed by a week of fitness posts produces an unrecognizable embedding that matches nobody's interest profile.

The more specific your niche, the tighter the audience match. "Tech" is too broad. "AI agents" is better. "Production deployment patterns for LangGraph agents" is precise, and it will match cleanly with a small, hungry audience.

Strategy 10: Avoid the Negative Weight Triggers

Blocks, mutes, reports, and "not interested" clicks are negative numbers in your score. Outrage content backfires: it pulls engagement, but it also pulls the negative signals, and the math cancels.

Practical guardrails:

Disagree sharply, but never personally
Critique ideas, not identities
Avoid contempt — the tone model detects it directly
Skip cheap dunks on out-groups — high block-rate among the targeted group
Watch your reply tone — combative replies on your own thread feed the tone model

Sharp contrarianism is rewarded. Hostility is penalized. The model is calibrated to detect the difference.

Strategy 11: Use Bookmarks as a Proxy for Content Quality

Bookmarks score 10x a like. They are also the cleanest signal of intrinsic content quality the algorithm tracks — bookmark means "I want to come back to this." It is impossible to fake.

Content types that get bookmarked:

Frameworks — step-by-step models people can apply
Checklists — usable, reference-quality lists
Tactical breakdowns — specific how-tos with examples
Reference data — comparison tables, benchmarks, definitions
"Save this for later" content — anything an aspirational reader will want to revisit

One bookmark-bait post per week dramatically lifts your average engagement weight. These are your highest-scoring posts in the long run because each bookmark = 10 likes in scoring terms.

Strategy 12: Build Conversational Threads with High-Value Accounts

Strategic replies to popular posts in your niche expose you to that account's audience. When a high-follower account replies back, their entire audience sees the exchange. The algorithm interprets engagement with quality accounts as a quality signal for you, which lifts your TweepCred and your distribution.

The rules:

Replies must be substantive — adds a fact, a counter-point, a useful expansion. Empty "great post!" replies do nothing.
Reply early — the first 10 replies on a popular post get the most visibility.
Be the smartest comment in the thread — quality replies pull profile clicks and follows.
Build relationships over time — repeated thoughtful engagement with the same account compounds.

Spending 20 minutes a day on smart replies to 5 popular accounts in your niche is, for many creators, more impactful than another original post.

Chapter 8: Common Myths Debunked

Myth: The algorithm suppresses small accounts.

Reality: It suppresses low-engagement content. Small accounts with high engagement rates get excellent distribution. The TweepCred system throttles new accounts to 3 concurrent in-distribution posts, but it does not blacklist them. Ratio matters, not absolute numbers. A 500-follower account with 8% engagement will out-distribute a 50,000-follower account with 0.2% engagement on a per-post basis.

Myth: Posting more = more reach.

Reality: More posts past 5/day dilutes per-post engagement, which drops your account-level baseline, which contracts your overall distribution. The Author Diversity Scorer also actively penalizes high-frequency posting within a single user's feed. Quality and consistency at 2–5 posts/day beat volume every time.

Myth: Hashtags help discovery.

Reality: The Grok transformer reads your post semantically. It understands what your content is about without hashtags. Hashtags add zero signal to the model, and 3+ hashtags are actively flagged as spam. Discovery happens through Phoenix Retrieval's topic embeddings — driven by your actual content, not metadata.

Myth: Deleting and reposting resets the algorithm.

Reality: Deletion-and-repost is a canonical spam pattern. The system detects it and applies suppression to the second post. If a post underperforms, let it go. Learn from it and write the next one better.

Myth: Scheduling tools hurt you.

Reality: The algorithm evaluates the content and its engagement velocity, not the publishing mechanism. Buffer, Hypefury, Typefully, native X scheduling — all neutral. Use whatever lets you post at peak windows consistently.

Myth: The algorithm changes every week.

Reality: The core ranking factors — engagement velocity, the 27x reply weight, the 150x reply-response weight, 6-hour halving decay, the link penalty — have been stable since the January 2026 Grok update. Tweaks happen to weights, but the structural pillars do not move. Building on the principles in this book is durable.

Myth: Only Premium accounts can grow.

Reality: Premium provides a 4–8x distribution boost via the TweepCred multiplier. It accelerates growth, especially for accounts trying to break out of the sub-65 TweepCred throttle. But non-Premium accounts with high engagement, niche specificity, and consistent posting still grow strongly. Premium is an accelerator, not a gatekeeper.

Myth: Going viral makes your account.

Reality: A single viral post moves followers but does not move TweepCred or topic embedding meaningfully. The accounts that compound on X are the ones with consistent, sustained, high-quality output over 6+ months. Virality is a side effect of doing the right things consistently, not the goal itself.

Myth: Replies don't matter unless they're from big accounts.

Reality: Every reply scores 27x a like, regardless of follower count. Every reply you respond to scores 150x. The signal is the conversation, not the celebrity of the conversant.

Myth: Long posts hurt because of attention spans.

Reality: Long, substantive posts maximize dwell time, which is one of the highest-weighted signals in the entire system. The Grok transformer rewards information density. Long does not mean rambling — it means earning the reader's attention with substance. Long quality wins.

Chapter 9: The GhostWriter Integration — Writing X Posts That Beat the Algorithm

Everything in the previous chapters compresses into a single, repeatable writing process. This is how to write every X post with algorithm awareness built in.

The Anatomy of an Algorithm-Optimized Post

A high-scoring post has five components, in order:

1. The Hook (first line, first 280 chars). Must work standalone — the algorithm previews and tests on the opening alone. Use a strong claim, a surprising stat, a sharp question, or a relatable pain point. Avoid throat-clearing ("I've been thinking lately..."). Open with the punch.

2. The Body. Specific, substantive value. Dense information. No filler. The body earns the dwell signal. Every sentence should justify its place — if a line does not add information or sharpen the angle, cut it.

3. The Reply Trigger. End with a question, an opinion invitation, a polarization prompt, or a call to share experience. Replies are the 27x signal. Engineer for them.

4. No Link in Body. If there is a link, it goes in the first reply. The body should be complete on its own — the link is supplemental, not required.

5. Length Calibration. Long enough to reward reading (dwell time signal). Short enough to retain readers through the reply trigger. The sweet spot is typically 600–1,200 characters for a standalone post. Threads can run much longer because they distribute the dwell across multiple tweets.

Five Post Templates That Score High

Template 1: The Experiment Result

I [did X] for [Y days/months]. Here's what happened: [surprising result]. [Brief explanation of why]. Would you try this?

This template combines: specificity, concrete data, a surprising payoff, and a direct reply trigger. Works for productivity, fitness, business experiments, learning sprints.

Template 2: The Unpopular Opinion

Unpopular opinion: [contrarian take]. Here's why: [reasoning in 2–3 sentences]. [Supporting example or evidence]. Agree or disagree?

Generates quote tweets (25x weight) and divisive replies (27x). The contrarian frame triggers P(quote) predictions before any engagement.

Template 3: The Bookmarkable List

The [N] things nobody tells you about [topic]:

[Item one — specific]

[Item two — specific]

[Item three — specific] ... Which surprised you most?

Bookmarks at 10x weight. The "nobody tells you" frame triggers curiosity. Numbered list creates reading obligation and reference value.

Template 4: The Correction

[Concrete fact or stat]. Most people get this wrong because [reason]. The real explanation: [accurate framing]. What's your take?

Educational, authoritative, conversation-starting. Works because it positions the author as knowledgeable while still inviting dialogue.

Template 5: The Replacement Frame

Stop [common behavior]. Do [alternative] instead. Here's why: [explanation]. [Specific example of the alternative working]. What's working for you?

Imperative voice grabs attention. Replacement frame is inherently practical. Reply trigger invites readers to share their version.

The 5-Minute Pre-Post Checklist

Run this checklist on every post before publishing:

Does the first line work as a standalone hook?
Is the body specific and substantive, with no filler?
Is there a question, opinion invitation, or reply-trigger at the end?
Are links removed from the body (moved to first reply)?
Are there 0–1 hashtags (not 3+)?
Is the tone constructive (not contemptuous or combative)?
Is the post in one of your 1–3 core topic lanes?
Would someone bookmark this to return to later?
Does this invite conversation, not just agreement?

Eight or nine "yes" responses means the post is ready. Five or six means rewrite before publishing. The discipline of running this checklist is what separates accounts that grow from accounts that stall.

What to Do in the First 30 Minutes After Publishing

The 30 minutes after publishing are more important than the 30 minutes spent writing the post. Specifically:

Reply to every comment, fast. Aim for 100% response rate inside the test window.
Quote-engage your own post into a second one if traction is high — extends the window.
Engage with other accounts in your niche during the window to keep your presence active.
Do not delete and repost if it underperforms. Let it go.
Do not edit substantially — edits can reset distribution analysis on some implementations.

Chapter 10: The Playbook in Practice — 30-Day Growth Sprint

A four-week, day-by-day framework for applying algorithm knowledge to grow fast. This is the operational rollout of every principle in this book.

Week 1: Foundation

Goal: Establish posting rhythm, learn what already works, and build the 30-minute reply habit.

Day 1–2: Audit. Pull your last 30 posts. For each one, record: number of replies, reply rate (replies / impressions), bookmarks, and quote tweets. Identify the top 5 posts by reply rate. Look for patterns — topic, format, length, opening line. These patterns are your historical product-market fit.

Day 3–4: Establish rhythm. Begin posting 2x per day at peak windows (8–10 AM and 12–1 PM in your audience's time zone). Reply to every comment within 30 minutes. This is non-negotiable.

Day 5–7: First thread. Build one thread of 5–8 tweets in your strongest topic area. Use the structure from Chapter 7: hook tweet, numbered breakdown, concrete examples, counter-intuitive insight, summary + reply prompt. Measure thread engagement against your standalone post baseline.

End of week 1: You should have 14 posts and 1 thread. Total replies should be at least 3x your previous baseline. If not, audit the reply triggers in your posts — most likely the issue is unclear or absent prompts at the end.

Week 2: Optimization

Goal: Apply specific algorithmic levers and measure the lift.

Day 8–9: Remove all links from post bodies. Test the engagement lift on link-bearing content — write the value in the post, link in the first reply. Compare distribution against a control week.

Day 10–11: First video post. Record one native video — 30 seconds, phone camera, no editing. Upload natively, with captions if possible. Compare its reach against your text equivalents in the same week. Most accounts see 3–10x impression lift on their first native video.

Day 12–14: Niche reply campaign. Identify 5 accounts in your niche with active, engaged audiences. Reply thoughtfully (substantive, value-adding) to one post from each, every day. By end of week 2, you should have 25–35 high-quality replies on others' content. Track inbound profile visits and new followers from this activity.

End of week 2: Total impressions should be up 30–60% over week 1. Out-of-network impression share (impressions / followers) should be measurably higher. New followers from reply activity should be visible.

Week 3: Amplification

Goal: Trigger higher-weight signals — quote tweets, bookmarks, and out-of-network expansion.

Day 15–17: Quotable content campaign. Write three posts engineered for quote tweets: one contrarian opinion, one useful data point, one prediction. Use the templates from Chapter 9. Track quote tweet rate per post — this is your 25x signal.

Day 18–20: Build a bookmarkable reference. Create one post that is intended to be saved and revisited: a checklist, a step-by-step framework, or a reference list. This is your 10x bookmark signal. Aim for at least 3% bookmark rate (bookmarks / impressions) — a strong reference post will hit 5–10%.

Day 21: Reply rate target. On at least one post this week, hit 3–5 high-quality replies in the first hour. This is the velocity signal that triggers out-of-network Phoenix expansion. If you can pre-seed engagement (Strategy 6) and reply-respond fast (Strategy 2), this is achievable on most posts.

End of week 3: You should see at least one post with 5x+ out-of-network expansion. Reply rate should be 2–3x your week 1 baseline. Bookmark count should be measurable on multiple posts.

Week 4: Compounding

Goal: Identify what's working, double down, and lock in the long-term posting machine.

Day 22–24: Topic embedding analysis. Review week 1–3 data. Which topics generated the highest out-of-network reach (impressions / followers)? These are your strongest Phoenix matches. Plan the next month's content cadence around these topics. This is your topic embedding strategy — narrow the lane, sharpen the fingerprint.

Day 25–27: Controversial constructive thread. Build one thread that takes a sharp, substantive contrarian position in your strongest topic. Constructive, not hostile. The goal is to maximize quote tweet rate (25x) while staying inside the tone model's positive zone. This is the highest-ceiling post format the algorithm offers.

Day 28–30: Audit, document, and lock in. Review all 30 days of data. Document: average reply rate, top posts, top topics, top times of day, top formats. Codify your personal playbook. From day 31 forward, this is your standing operating procedure.

Key Metrics to Track Throughout

These are the only metrics that matter. Ignore vanity counts.

Reply rate per post (replies / impressions). This is the truest signal of algorithmic favor. Target 1.5%+ consistently.
Out-of-network impression share (impressions / followers). 2x is good, 5x is strong, 10x is excellent. This measures Phoenix expansion.
Bookmarks per post. Target 0.5–2% bookmark rate. High bookmarks = high intrinsic content quality.
Profile clicks per post. 12x signal. Target 1–3% profile click rate.
Dwell time proxy (longer post + high save rate). No direct dwell metric is exposed, but a 1,000-character post with high engagement implies strong dwell.

Track these weekly in a simple spreadsheet. Three weeks of trend data will tell you which strategies are working for your specific account, and where to lean.

Closing: The One Sentence That Matters

The Grok-based algorithm rewards the same thing humans reward: substantive, specific, conversational content from active participants who add value to the platform. Every weight, every penalty, every signal in this book reduces to that single principle. Build the habit of writing posts you would want to read, in conversations you would want to join — and the algorithm will do the rest.

X Algorithm Growth Hacking: A Complete Guide Based on the Open-Source Code

About

Preview