Home / Resources / X Algorithm Master Guide

X Algorithm Master Guide

Every major data point in X's recommendation system, what it means, and how to use it for reach, impact, and restriction avoidance. Based on an audit of X's open-source For You architecture (~7,700 files, September 2025-era refresh).

Updated Aug 2026 63 sections ~45 min read Reference · not affiliated with X Corp

01Core algorithm architecture

  • X does not use one single recommendation algorithm. The For You feed is assembled through multiple retrieval systems, feature hydrators, ranking models, filters, rescorers, selectors, and mixers.
  • The broad flow is Candidate Source → Feature Hydration → Filter → Scorer → Selector → Mixer. A post must survive several separate stages before anyone ever sees it.
  • Home Mixer is one of the central For You systems. Much of modern timeline construction runs through this architecture.
  • Product Mixer is the broader recommendation framework. Recommendation is a pipeline rather than one ranking formula.
  • The analyzed repository contains roughly 7,700 files. The actual recommendation system is far more complex than the simplified “likes vs replies” explanations circulated online.
  • The source spans Scala, Java, Python, Rust, and C++. Search, machine learning, graph systems, model serving, and retrieval are separate technical layers.
  • The analyzed source reflects a September 2025-era refresh. It should not be assumed that every implementation detail is unchanged forever.
  • Many production feature-switch values remain private. The public source exposes architecture much better than exact current production weightings.
  • Heavy-ranker weights in the repository default to 0.0. We can see which behaviors X predicts without knowing the exact live value of every prediction.
  • The ranking objective is externally configurable. X can change how much different outcomes matter without rewriting the architecture.
  • Different viewers can receive different objective weights. There is not one identical ranking formula applied to every person.
  • Optimization buckets exist. X can experimentally change ranking behavior for different populations.
  • Daily viewer-level variation exists. One unusually weak post does not prove your account was penalized.
  • Recommendation outcomes are logged back into data systems. What happens after X recommends something becomes training information for future ranking.
How to use this

The biggest practical lesson: never optimize for one isolated “algorithm signal.” A post needs to clear eligibility, get retrieved, score well, avoid penalties, win inventory, and then generate real behavior.

To maximize reach

  • Optimize the complete journey rather than one metric.
  • Treat engagement, topic identity, reputation, relationships, and content quality as one system.
  • Evaluate patterns across dozens of posts rather than reacting to one result.

To minimize restrictions

  • Understand that filters and eligibility rules can override good engagement.
  • Never assume high likes can rescue spam, safety, duplicate-content, or negative-feedback problems.
  • Build strategies around structural mechanics rather than temporary parameter exploits.

02Candidate retrieval

  • The system starts with an enormous universe of possible posts and narrows it before heavy ranking. Most posts never directly compete inside the final expensive ranking stage.
  • The architecture narrows toward roughly 1,500 preliminary candidates from multiple sources. Your first battle is getting selected as a candidate.
  • Candidate sources generally fail open. A failure in one retrieval system does not necessarily destroy the entire feed request.
  • Earlybird supplies major in-network inventory. Followers have a dedicated retrieval pathway.
  • tweet-mixer coordinates substantial out-of-network retrieval. Stranger discovery comes from a separate ecosystem.
  • tweet-mixer contains dozens of retrieval pipelines. There are many routes to out-of-network distribution.
  • SimClusters ANN is one candidate source. Interest-community similarity can make a post retrievable.
  • TwHIN ANN contributes candidates. Learned representations of users and content also determine matching.
  • Deep Retrieval contributes candidates. Semantic similarity matters beyond explicit follow relationships.
  • UTG/UVG graph walks contribute candidates. Engagement relationships can carry posts through social networks.
  • Popular geographic/topic systems exist. Popularity inside contextual markets creates another route.
  • Dedicated video retrieval exists. Video has discovery surfaces unavailable to ordinary text.
  • UTEG can contribute hundreds of posts. “People you follow engaged with this” is a major distribution mechanism.
  • Lists are a separate retrieval lane. List inclusion can produce recommendation opportunities.
  • Communities are a separate retrieval lane. Community posting creates additional candidate inventory.
  • Backfill systems exist. X can fill feeds with secondary sources when stronger sources are insufficient.
  • Exploration/cold-start systems exist. Unknown posts and creators receive testing opportunities.
  • Warm caches can bypass new retrieval. A viewer does not necessarily get a completely rebuilt feed every refresh.
How to use this

Getting ranked well does not matter unless you first become a candidate.

To maximize reach

  • Build access to as many legitimate retrieval lanes as possible: followers, SimClusters, Lists, Communities, video, topics, graph relationships.
  • Think of distribution as a portfolio rather than relying only on followers.
  • Create content recognizable enough to qualify for semantic and topic retrieval.

To minimize restrictions

  • Avoid behavior that gets you removed before ranking: spam signals, poor reputation, duplicate content, safety problems, untrusted URLs.
  • Do not assume visible follower count guarantees candidacy.

03In-network distribution

  • A substantial portion of For You candidates come from followed accounts. Your existing audience is usually your initial ignition source.
  • A follow does not guarantee an impression. Your post still competes with other followed creators.
  • Earlybird uses realtime search over followed accounts. Follower distribution partially behaves like realtime search ranking.
  • A light ranker operates before heavy ranking. Posts can be eliminated before advanced models evaluate them.
  • The light ranker is based on older logistic-regression machinery. Some legacy features remain strategically relevant.
  • Replies receive unusually strong weight in the example light-ranker configuration. Conversation-oriented posts can receive important early advantages.
  • In-network retrieval can be restricted to a ranked subset of followed authors. Technically following someone does not mean their posts remain actively retrieved.
  • The source references a legacy value around 500 followed authors in some retrieval behavior. Weak follower relationships can effectively disappear algorithmically.
  • Real Graph helps determine which followed authors matter most. Strong historical interaction makes follower relationships more valuable.
  • Followers receive structural benefits unavailable to strangers. Converting out-of-network viewers into followers changes future distribution economics.
How to use this

Your followers are not a passive audience list. They are the first-stage distribution engine.

To maximize reach

  • Maintain frequent, meaningful interaction with your core followers.
  • Publish when that group is most likely to react quickly.
  • Create posts designed to generate real replies and favorites immediately.
  • Convert viral strangers into followers while attention is high.

To minimize loss

  • Avoid flooding followers with mediocre posts — ignored impressions can weaken future delivery.
  • Do not allow follower relationships to become completely passive.
  • Treat audience engagement maintenance as part of content production.

04SimClusters

  • SimClusters organizes users into large numbers of latent interest communities. Recommendation communities are much more granular than broad categories.
  • The audit describes roughly 145,000 communities. X can model extremely specific audience niches.
  • Users maintain InterestedIn representations. A person can belong to many overlapping interest clusters.
  • Tweets develop representations in similar community space. X determines which communities are likely to want the post.
  • People who favorite a post strongly influence its community representation. Your likers help route the post.
  • Self-favorites are excluded. Liking your own post provides no meaningful SimClusters advantage.
  • Likes function as routing information. Their importance extends beyond a visible engagement counter.
  • Who likes the post can matter as much as how many. Coherent niche engagement gives X clearer distribution information.
  • Random likes can spread the post’s representation across unrelated interests. Bigger raw numbers can sometimes produce less precise targeting.
  • Hyperactive engagement accounts can be excluded from graph processes. Engagement-pod likes may have less distribution value than expected.
  • Cluster-to-tweet retrieval can behave like a rolling top-400 leaderboard. Virality is partly about entering a small community-specific leaderboard.
  • Small niches are easier to dominate than giant broad categories. Specialization reduces competition.
  • Community scores decay over time. Recent likes matter more.
  • An important favorite/community score uses approximately an 8-hour half-life. Concentrated early velocity is extremely valuable.
  • Cluster success can self-reinforce. Community distribution produces more community likes, which strengthens distribution.
  • Accounts develop KnownFor representations. X learns what your account is reliably about.
  • Topic inconsistency weakens concentration. Random subject switching can dilute the account’s routing identity.
How to use this

SimClusters means your goal should not merely be “get likes.” It should be “get likes from the exact type of people I want more reach from.”

To maximize reach

  • Pick a primary niche and dominate it.
  • Build a core audience whose interests strongly overlap.
  • Create content highly relevant to that group.
  • Concentrate initial engagement in the first few hours.
  • Expand into adjacent niches only after establishing a strong home cluster.

To avoid weak distribution

  • Avoid purchased/random likes and engagement pods.
  • Avoid constantly changing subjects.
  • Avoid trying to appeal to everybody simultaneously.
The ideal early engagement pattern is a dense cluster of relevant people, not the largest possible random number.

05UTEG / UTG / graph propagation

  • UTEG uses social-engagement relationships. Posts can spread because somebody a viewer follows interacted with them.
  • GraphJet provides fast in-memory graph processing. Recent social interaction can immediately influence discovery.
  • Favorites can spread posts through followers’ networks. A relevant like can create secondary exposure.
  • Retweets contribute graph propagation. Redistribution creates additional network routes.
  • Replies contribute graph propagation. Conversation activity can carry content outward.
  • Quotes contribute graph propagation. Quote engagement can open both original and new-post paths.
  • The graph relies on short windows such as 24–48 hours. Propagation is highly time-dependent.
  • Minimum co-occurrence requirements exist. Several overlapping engagers can matter more than one powerful person.
  • The audit references co-occurrence around three in certain mechanisms. Small synchronized groups can unlock pathways single users cannot.
  • Minimum degree requirements exist. X attempts to avoid weak/spam graph relationships.
  • The source references values around 50 engagements in some relevant graph conditions. Active, established users can be stronger propagation nodes.
How to use this

Your best ignition group is not necessarily your biggest followers. It is a small group of active, relevant, overlapping people whose audiences resemble yours.

To maximize reach

  • Build relationships with 20–50 strong niche participants.
  • Have several of them naturally engage shortly after publication.
  • Create content that multiple members of the same social neighborhood find worthy of sharing or discussing.
  • Synchronize engagement instead of spreading it thinly across days.

To minimize restrictions

  • Do not manufacture synchronized activity through pods or automation.
  • Make sure the interaction is genuine and topic-relevant.

06Feature hydration

  • Candidates receive thousands of features. Ranking considers far more than engagement counters.
  • The audit references roughly 6,000 features. Viewer, author, content, graph, and historical information are all available.
  • Dozens of feature hydrators contribute information. Different systems specialize in different behavioral dimensions.
  • SimClusters embeddings are attached. Community relevance becomes a ranking input.
  • TwHIN representations are attached. User/content similarity matters.
  • Graph features are attached. Relationships matter.
  • Realtime engagement aggregates are attached. Current momentum matters.
  • Topic features are attached. Semantic classification matters.
  • Author-quality features are attached. Your history follows every post.
  • Viewer-history representations are attached. Every candidate is judged differently for different users.
How to use this

There is no isolated trick capable of dominating thousands of features. Build strength across several categories simultaneously: good account history, strong niche identity, healthy relationships, clean content, early momentum, high-quality media, and viewer relevance.

  • Avoid accumulating weak account-level signals — new posts inherit author context.
  • Treat every post as a deposit into future feature history.

07Heavy ranking

  • Navi serves heavy ranking models. Posts surviving retrieval receive more sophisticated prediction.
  • Phoenix represents a newer transformer-based system. Raw behavioral sequences are increasingly important.
  • Several ranking systems coexist. X is actively evolving the architecture.
  • Predicted behaviors include: favorite probability, reply probability, retweet probability, reply-engaged-by-author probability, good-click probability, good-profile-click probability, video-quality-view probability, immersive video-quality-view probability, video watch behavior, video watch time, negative feedback, and report probability.
  • Weak and strong negative feedback can be distinguished. Not all pushback is weighted equally.
  • The overall score combines weighted predicted outcomes. X evaluates expected behavior, not raw engagement already received.
  • Negative prediction heads can be threshold-gated. Once negative risk crosses a boundary, punishment can increase sharply.
How to use this

Create content capable of winning on multiple predicted behaviors simultaneously. An ideal post makes someone likely to stop scrolling, read or watch, like, reply, save, share, click your profile — and avoid negative feedback.

To maximize reach

  • Optimize for a portfolio of valuable behaviors, not one visible metric.

To minimize restrictions

  • Avoid strategies that pump one metric while increasing predicted reports, mutes, or “show less.”
A 1,000-reply rage post is not automatically superior to a 300-like reference post with strong dwell, bookmarks, profile conversions, and no negative feedback.

08Replies

  • Replies are unusually valuable in the light ranker. The example reply weight is approximately 9.0.
  • By comparison: favorite ≈ 1.0, retweet ≈ 1.0, click ≈ 0.3, open-link ≈ 0.1, photo expansion ≈ 0.03, certain video-playback behavior ≈ 0.01.
  • Reply-engaged-by-author has its own heavy-ranking head. X explicitly models authors responding to commenters.
  • Standalone replies receive an approximate ×0.75 multiplier. Replies are valuable interaction tools but weaker standalone inventory.
How to use this

The best structure is: original post for reach + replies for acceleration.

To maximize reach

  • Write posts that naturally invite disagreement, additions, questions, or personal perspectives.
  • Reply quickly to meaningful comments.
  • Continue substantive conversations instead of responding with empty emojis or one-word answers.
  • Use replies to strengthen relationships and keep conversation modules active.

Avoid

  • Spammy “reply YES” engagement bait.
  • Meaningless controversy.
  • Using replies as your primary content format.

09Favorites / likes

  • Favorites are directly predicted by ranking.
  • Favorites power SimClusters community routing.
  • Favorite velocity matters.
  • Favorite coherence matters.
  • Favorites-per-post becomes an author-quality signal.
  • High-volume weak-favorite behavior can hurt producer-quality metrics.
  • Engagement-pod likes can provide weak graph value.
How to use this

Likes are most useful when they come from relevant people quickly.

To maximize reach

  • Create posts your target niche genuinely wants to endorse.
  • Prioritize concentrated early likes.
  • Build a focused audience rather than chasing random engagement.
  • Track first-hour like velocity as a key health indicator.

To minimize restrictions

  • Do not buy likes.
  • Do not use engagement pods.
  • Do not publish so frequently that your average favorites per post collapse.

10Retweets & quotes

  • Retweets are positive signals.
  • Plain retweets generally preserve the original scoring identity.
  • Quote tweets create a new post identity. Quotes receive their own text, embedding, and predictions.
  • A strong quote can create a second distribution opportunity.
  • Quotes can be deduplicated against the original. Early quotes can therefore be more useful than late quotes.
  • Engagement from large retweets can contribute to original-author aggregates.
How to use this

For amplification by collaborators, thoughtful quote tweets can be more valuable than simple reposts because they create another independent candidate.

To maximize reach

  • Encourage relevant creators to add their own insight when amplifying.
  • Quote important posts early while the information is still spreading.
  • Make the quote itself valuable enough to stand independently.

Avoid

  • Late, low-value quotes of already saturated viral posts.
  • Endless self-retweets expecting fresh ranking opportunities.

11Bookmarks, shares, DMs & screenshots

  • Bookmarks are tracked.
  • Shares are tracked.
  • DM sends are valuable private behavior.
  • Copy-link actions are tracked.
  • Screenshots appear as explicit user-action events.
  • Private engagement does not have the social cost of visible engagement.
How to use this

Create more content that people want to keep, not merely applaud: cheat sheets, frameworks, lists, benchmarks, visual explanations, tutorials, templates, useful screenshots, reference tables. These posts can generate substantial algorithmic value even when public likes seem modest.

The optimization question is “Would somebody save or send this?” — often more useful than asking only “Would somebody like this?”

12Dwell time

  • Tweet dwell time is measured.
  • Profile dwell is measured.
  • Fullscreen video dwell is measured.
  • Linger duration can be bucketed.
  • Tweet clicks require enough dwell to become “good clicks.” Around two seconds is referenced for good tweet-click behavior.
  • Around ten seconds is referenced for meaningful profile-click behavior.
How to use this

Your job starts before the user clicks anything: make them stop.

To maximize reach

  • Use highly readable opening lines.
  • Use graphics users need to inspect.
  • Produce information-dense visuals.
  • Structure longer posts so people spend time consuming them.
  • Make your bio and pinned content strong enough that profile visitors stay.

Avoid

  • Curiosity bait that creates immediate bounce.
  • Misleading hooks.
  • Weak profiles that waste profile traffic.

13Negative feedback

  • Reports can receive enormous negative weight. The report-bound can reach approximately −20,000, while typical positive-head bounds are around +10,000.
  • “Show less” is explicitly modeled. Viewer-author fatigue can last around 140 days; the multiplier can start around 0.2 and gradually recovers toward 1.0.
  • Negative feedback can weaken social-proof paths.
  • Blocks matter. Mutes matter. Reports affect relationship models.
  • Negative events can exclude posts before ranking.
  • Similarity to content someone disliked can suppress your post.
How to use this

The downside is more asymmetric than most creators realize.

To maximize sustainable reach

  • Produce provocative content your target audience enjoys — not content designed to make unrelated audiences furious.
  • Protect your report rate aggressively.
  • Monitor whether viral posts attract the right audience.

To minimize restrictions

  • Avoid gratuitous rage bait.
  • Avoid misleading claims.
  • Avoid spam-like posting.
  • Avoid repeatedly exposing content to audiences that clearly reject it.
The strategic rule: Good reach = positive engagement minus negative audience cost.

14Safety & hard visibility gates

  • Some restrictions remove posts outright.
  • NSFW content can block recommendation. Account-level NSFW labels can affect all posts; NSFW avatar and banner rules exist.
  • Gore and violence can restrict recommendation.
  • Abuse-policy states can restrict recommendation.
  • DoNotAmplify-type labels exist.
  • Spam labels exist. Duplicate-content labels can become hard restrictions, and engagement-spammer labels exist.
  • Recommendation blacklists exist.
  • Untrusted URLs can be hard-dropped.
  • New accounts can encounter NotGraduated-type restrictions.
How to use this

A hard gate is more serious than a weak ranking score.

To maximize reach

  • Keep account-level assets clean.
  • Use trustworthy links.
  • Publish original content.
  • Maintain healthy account behavior.

To minimize restrictions

  • Review your avatar and header for anything that could trigger sensitive-media classification.
  • Do not systematically repost others’ content.
  • Avoid questionable domains.
  • Treat safety enforcement separately from ordinary engagement optimization.

15Grok classification

  • Grok increasingly classifies posts directly.
  • Quality, slop, and spam are classified.
  • Gore, violence, and NSFW content are classified.
  • Topics are classified: news, sports, music, memes, crypto, science, and more.
  • Political-category features exist. Left/right/neutral predicates appear in the source.
  • Auto-translation functionality exists.
How to use this

Write for humans and classifiers at the same time.

To maximize reach

  • Make the topic unmistakable.
  • Use canonical product, company, and person names.
  • Provide real substance.
  • Make originality obvious.
  • Keep the post understandable without excessive context.

To minimize restrictions

  • Avoid generic AI-generated filler.
  • Avoid content easily classified as spam or slop.
  • Avoid ambiguous sensitive-media material.
As Grok-style systems become more important, machine-readable quality becomes increasingly important.

16Slop

  • SlopFilter exists. Slop rescoring exists.
  • Strong slop classification can reduce out-of-network reach.
  • Treatment can vary by viewer/author state.
  • There may be an invisible slop growth ceiling.
  • New posts are classified fresh.
How to use this

Do not build an account around cheap repetitive virality.

To maximize reach

  • Produce original analysis.
  • Create original visuals.
  • Add unique data.
  • Develop recognizable expertise.

To recover from slop-like treatment

  • Reduce volume.
  • Stop using cloned hooks.
  • Stop posting generic summaries.
  • Increase originality for several weeks.
You usually need a new content pattern — not a new account.

17Text scoring

  • Text is scored before engagement.
  • Length contributes significantly. The supplied composition gives length around 0.5.
  • Entropy contributes around 0.25. Repetitive wording performs worse, and duplicate labeling reinforces the problem.
  • Shoutiness is measured.
  • Offensive text can receive around ×0.2 damping in the relevant scorer. Offensive usernames can contribute.
  • Mixed language can hurt eligibility.
  • Multiple hashtags are penalized.
  • Mention count is tracked.
  • No magical viral word list exists.
How to use this

The ideal X writing style is substantial, concise, varied, readable, clean, single-language, non-template, and conversational. Give people enough information to stop and read while keeping the message sharp.

  • Avoid ALL CAPS.
  • Avoid excessive profanity.
  • Avoid repetitive hooks.
  • Avoid hashtag stuffing.
  • Avoid empty one-line bait.

18Content language

  • Viewer language and tweet language are compared.
  • Device-language compatibility is modeled.
  • Language filters can remove posts.
  • Consistent language increases eligibility.
  • Auto-translation can expand reach. Heavy slang can make translation weaker.
How to use this

If broad reach is the objective, write in clear, translatable language: conventional grammar, recognizable entity names, concise sentences.

  • Avoid slang requiring niche cultural context.
  • Avoid excessive mixed-language posts.
  • Avoid wordplay that becomes meaningless in translation.

19Hashtags

  • Multiple hashtags are explicitly damped.
  • A multiple-hashtag/trend feature exists.
  • Zero hashtags is acceptable.
  • One relevant hashtag can still be useful.
How to use this

Treat hashtags as context, not distribution fuel. Default to 0 hashtags; use 1 when it genuinely improves categorization or event context.

  • No hashtag blocks.
  • No repeated trend stuffing.
  • Don’t turn the post into Instagram-style metadata.

20Mentions

  • Mention count is tracked.
  • User-mention aggregates exist.
  • Relevant mentions can improve context.
  • Mention spam increases negative-feedback risk.
How to use this

Mention important community-central accounts only when they genuinely belong in the story. A useful mention can clarify topic, associate your post with a relevant conversation, and improve audience understanding.

  • Do not tag ten influencers hoping for attention.
  • Do not repeatedly mention people who don’t interact with you.

21Native media

  • Media-related boosts and features exist.
  • Native-image flags exist.
  • Video flags exist.
  • Card flags exist.
  • News URL features exist.
  • Trend features exist.
  • Native content creates more on-platform interaction possibilities.
How to use this

Whenever reasonable, give the post a native payload. Instead of “Read this link,” prefer: headline → key insight → chart/screenshot/video → explanation → link as secondary reference. This maximizes dwell and media interaction without depending on external clicks.

22Video

  • Video is heavily instrumented. It has separate candidate pipelines, graph infrastructure, and dedicated ranking infrastructure.
  • Multiple positive video heads exist. Quality-view, immersive quality-view, watch behavior, and watch time are all predicted.
  • Videos under roughly 10 seconds miss important quality-view benefits.
  • Short-form eligibility spans roughly 5–300 seconds. Vertical and square ratios qualify for short-form pathways.
  • Long-form spans roughly 2–240 minutes. Widescreen formats are supported for long-form, and HD minimums are referenced.
  • Completion rate is a feature.
  • 10–45 seconds vertical is a useful strategic sweet spot inferred from these mechanics.
How to use this

Video should be treated as a separate growth engine.

For discovery

  • Create 10–45 second vertical clips.
  • Hook immediately.
  • Remove dead time.
  • Optimize for completion.
  • Make the topic obvious within seconds.

Avoid

  • Meaningless clips under 10 seconds.
  • Padded videos that destroy completion.
  • Poorly formatted uploads.

For depth, use longer video when the subject deserves it.

23Evergreen video

  • Evergreen video retrieval exists.
  • Older followed-account videos can reappear.
  • Evergreen embedding pipelines exist.
  • Video can therefore outlive normal timeline windows.
  • New followers can increase the future value of old videos.
How to use this

Treat high-quality videos as library assets, not disposable tweets. Create evergreen videos answering recurring questions — tutorials, explainers, important concepts. This creates compounding value as your follower graph grows.

25Questions

  • Questions are machine-recognizable content types.
  • Questions naturally produce replies.
  • Questions still need substance.
How to use this

Ask questions after giving the reader something worth responding to. Create questions that invite expertise, disagreement, examples, or predictions.

Strong: “Open-weight models are approaching frontier capability while inference costs collapse. Which enterprise workload moves local first?”
Weak: “Thoughts?”

26Threads

  • Conversation modules bundle related posts. Root and focal replies can appear together.
  • A module size around three appears in one component.
  • Self-replies can become additional candidates.
  • A later strong reply can revive the root.
How to use this

Treat every thread entry as its own potential viral post. Ideal: 1) strong root, 2) strong second insight, 3) strong data/example, 4) strong conclusion or question.

  • Avoid 20-part filler threads.
  • Avoid forcing one idea into unnecessary pieces.

27Time & decay

  • Ordinary For You eligibility falls off around 48 hours.
  • Most out-of-network systems are strongest inside approximately 24 hours.
  • The first few hours matter most. An important community score uses an 8-hour half-life.
  • 30-minute, two-hour, and 24-hour realtime features exist.
  • Age decay begins before hard expiration.
  • X knows viewer-specific active hours. X also models day-of-week behavior.
How to use this

The launch window matters enormously. Optimize the sequence:

  • T−60 min: be active with your niche.
  • T=0: publish.
  • T+0–60: answer replies rapidly.
  • T+1–8h: let momentum build.
  • T+8–24h: continue conversation if it’s working.
  • After 24h: prioritize conversion and follow-up.
  • After 48h: treat the normal For You lifecycle as largely complete.
Do not obsess over universal posting times. Find your audience’s active period.

28Platform load

  • Latency-driven quality factors exist. The factor can range approximately 0.1–1.0.
  • Candidate caps can shrink substantially under load. The audit describes a potential ~600 → ~60 reduction in one context.
  • Audience shoulder periods may therefore be attractive.
How to use this

Test whether posting slightly before or after absolute peak performs better than peak. You want your audience online, but potentially less platform-wide candidate competition. Use your analytics rather than assuming a universal answer.

29Caching

  • Scores can be cached around three minutes.
  • Cached candidates may avoid immediate rescoring.
  • Warm caches can bypass fresh retrieval.
  • Momentum can therefore occur in pulses.
How to use this

Do not attempt minute-by-minute cache hacks. The useful insight is simply: early engagement can compound through repeated feed refreshes. Keep the first hour active and conversational.

30Author diversity / cannibalization

  • Out-of-network posts can receive approximately ×0.75. Replies also receive approximately ×0.75.
  • Same-author decay exists. The second exposure can fall around ~0.5 depending on configuration, and later exposures can fall toward a floor.
  • Past impressions can contribute to decay.
  • Unengaged impressions are not neutral.
How to use this

Do not make your own posts compete with each other. Best practice: 1–2 strong primary posts per day, several hours apart — and wait longer when something is actively exploding.

Think of every impression as spending audience capital.

31Real Graph

  • Real Graph models relationship strength. Dozens of interactions feed it — the audit references roughly 45 features.
  • Favorites, replies, profile views, dwell, DMs, mentions, mutual follows, and address books all matter.
  • Blocks, mutes, and reports hurt.
  • Relationships decay daily. The coefficient referenced is around 0.955/day.
How to use this

Audience maintenance is not optional. Every day: respond to people, interact with niche peers, engage your best followers, participate in relevant conversations.

Your follower graph is a decaying network, not a static asset.

32The algorithmic unfollow

  • People can technically follow you while rarely receiving you.
  • Weak relationships can fall out of useful retrieval.
  • Follower count can significantly overstate reachable audience.
  • Ongoing engagement keeps edges alive.
How to use this

Do not celebrate follower count without measuring follower engagement. A 20,000-follower account with 2,000 active relationships can outperform a 100,000-follower account with 1,000.

Build active followers, not simply followers.

33Tweepcred / reputation

  • PageRank-like reputation influences retrieval.
  • Follower quality matters.
  • Following-to-follower ratios can reduce reputation. Mass follow-for-follow therefore has costs.
  • Account age contributes.
  • Device and account hygiene contribute.
  • Safety status contributes.
  • Verification contributes.
  • Low-reputation filters can eliminate search visibility.
How to use this

Build a clean, reputable social graph.

Do

  • Earn relevant followers.
  • Interact with credible accounts.
  • Keep following behavior natural.
  • Maintain account hygiene.

Avoid

  • Following thousands hoping for reciprocal follows.
  • Spammy growth tactics.
  • Low-quality follower purchasing.

34Verification

  • No universal heavy-ranker checkmark multiplier is proven.
  • Verification remains a model feature.
  • Earlybird contains verified boosts. Blue-verification parameters exist.
  • cr-mixer contains verified partition behavior.
  • Verification bypasses certain reputation and link restrictions.
How to use this

Verification appears useful mainly because it removes friction and improves some retrieval conditions. If X is strategically important to your business or brand, Premium can be worth viewing as distribution infrastructure — but not as a magical substitute for content quality.

35Author quality

  • Follower count is measured.
  • Historical favorites and favorites-per-tweet are measured.
  • Large-account push tiers exist. 100K, 1M, and 3M follower thresholds appear in producer-quality logic.
  • A >3,000 favorites-per-tweet condition appears in one high-tier rule. Very low report rates are required.
  • High-volume weak-engagement behavior can be downboosted. >56 posts/28 days with <100 favorites/post appears in one negative rule.
How to use this

The system rewards quality-adjusted output, not pure posting frequency. Protect favorites/post, report rate, and engagement/post.

If increasing frequency cuts average quality substantially, reduce frequency.

36Viral creator features

  • Viral creator metrics exist.
  • Viral creator predicates exist.
  • Grok creator classifications exist.
  • Success can become an account-level property.
How to use this

A viral post can be more valuable than its immediate impressions, because sustained success changes your author profile. Focus on repeatability: same niche, consistent quality, multiple winners over time. The objective is to become recognized as a creator who reliably generates successful content.

37Lists

  • Lists have dedicated retrieval.
  • List followers can receive posts from listed accounts.
  • Creators rarely optimize List placement.
  • Good List placement compounds.
How to use this

Actively try to get added to reputable curated Lists in your niche — it’s potentially free additional distribution inventory.

  • Publish valuable niche work.
  • Build relationships with List owners.
  • Create your own valuable Lists.
  • Become a recognized category expert.

38Communities

  • Communities have separate candidate pipelines.
  • Top-K restrictions can operate locally.
  • Competition is smaller than the full network.
  • Engagers are usually topically coherent.
How to use this

Join 2–3 strong, relevant Communities rather than dozens of random ones. Post niche-specific material, discussions, and useful resources. Communities can help both direct distribution and cluster coherence.

39Exploration

  • X deliberately explores unknown posts.
  • Exploration candidates can be inserted after ordinary scoring.
  • Low-view posts can qualify. Max-view-count logic helps preserve unseen candidates.
  • New creators therefore receive auditions.
How to use this

Small creators are not permanently trapped by size. When X gives you an unexpected test burst: make the post immediately understandable, maximize dwell, answer every reply, and have a strong profile ready.

Treat cold-start impressions as conversion tests.

40Low-signal users

  • Viewers are segmented by behavioral state. New/NearZero/VeryLight states exist.
  • Low-signal feeds can be heavily topic-based. Grok topics become especially important.
  • Heavy users receive deeper personalization.
How to use this

For broad growth, content needs two properties: 1) category clarity so X can show it to low-signal and new users, and 2) relationship depth so it performs among power users. Do not optimize only for one audience type.

41Topic legibility

  • Topic classification systems exist.
  • Topic proofs are features.
  • Canonical entities improve machine understanding. Explicit product, company, and model names help.
  • Obscure phrasing can hurt classification.
How to use this

Say what the post is actually about. Prefer “DeepSeek V4,” “NVIDIA Blackwell,” “Federal Reserve,” “OpenAI” over vague “this company,” “this model,” “these guys.”

Machine-readable specificity expands recommendation opportunities.

42Muted keywords

  • Muted keywords can hard-drop content per viewer.
  • Creators cannot see how many people muted a term.
  • Promotional/spam terms can carry hidden costs.
  • Political terms can have large mute audiences.
  • Crypto/NFT words may similarly restrict certain viewers.
How to use this

Use high-friction vocabulary intentionally. A word can simultaneously increase relevance to your target community and eliminate distribution to people who muted it. Do not stuff posts with unnecessary controversial or spam-associated terminology.

43Duplication

  • Duplicate-text labels exist. Duplicate-content hard-drop rules exist.
  • Retweet dedup exists. Quote dedup exists. Media-ID dedup exists.
  • CLIP similarity clustering exists.
  • Repeated visual formats can fatigue.
How to use this

Repost ideas, not copies. Refresh the wording, image, crop, framing, data, hook, and format. Stay semantically relevant while remaining visually and textually original.

44Reposting old content

  • Old normal posts die after the ordinary recommendation window.
  • New reposts get new IDs. They receive fresh age windows, fresh community embeddings, and can receive fresh exploration.
  • Identical media still carries similarity history.
  • Identical text can trigger duplicate concerns.
  • Deleting the old post first usually adds no benefit.
How to use this

Recycling winners is valid. Best practice: wait meaningful time, rewrite the post, refresh the visual, update the idea, and publish as a new edition. Do not simply copy/paste winners every week.

45Delete & repost

  • No universal delete/repost punishment flag was identified.
  • Immediate deletion can be inexpensive. The audit uses roughly 2–3 minutes as a practical low-cost period.
  • Deletion after momentum destroys accumulated value. Likes do not transfer, velocity does not transfer, and original impressions remain spent.
  • Editing is preferable once the post is moving.
How to use this
Major mistake discovered instantly → delete/repost.
Minor typo or issue after engagement starts → usually leave or edit.

Never casually destroy a post with real momentum.

47Push notifications

  • Push has separate recommendation infrastructure. Several candidate sources compete.
  • Open/engage probability is modeled.
  • Around five pushes per day is referenced as a default cap. Around four hours minimum spacing is referenced.
  • Recent Home visitors can be excluded. Around 20 hours appears in one suppression rule.
  • Dormant users therefore become disproportionately important.
How to use this

Push gives you a path to followers who do not regularly scroll. To improve your chances: maintain strong follower relationships, publish posts with obvious immediate value, and build a reputation for content worth opening notifications for.

48Follow recommendations

  • FRS powers account recommendations.
  • Graph walks are used. Strong-tie prediction is used.
  • Address-book data can contribute.
  • Follow recommendation indirectly creates tweet distribution.
How to use this

Optimize the account itself: clear profile picture, clear bio, recognizable niche, strong pinned post, consistent topic, strong engagement after people follow. Winning follow recommendations increases future in-network inventory.

49Quality follows

  • A quality-follow definition exists.
  • Fast engagement after following matters.
  • Immediate unfollows, blocks, mutes, and reports are negative.
  • The first ~24 hours after acquisition are important.
How to use this

When something goes viral, do not disappear. Have a high-quality follow-up ready.

The goal: viral impression → profile visit → follow → second engagement → strong relationship. That turns temporary viral reach into permanent distribution infrastructure.

50Author realtime momentum

  • Author-level engagement realtime aggregates exist. 30-minute, two-hour, and 24-hour-type windows are referenced.
  • One viral post can lift other candidates temporarily.
  • Immediate follow-ups can still cannibalize sessions.
  • Waiting too long loses the halo. The audit infers a possible 6–12-hour sequel window.
How to use this

When a post is exploding: do not immediately post something unrelated, keep interacting, and watch the growth curve. When growth begins flattening but momentum remains elevated, launch a strong related sequel. Do not treat 6–12 hours as a rigid rule — test the principle.

51The mediocre-post death spiral

  • Weak posts consume impressions. Ignored impressions increase fatigue.
  • Next posts start weaker. Lower reach causes lower engagement.
  • Author averages deteriorate. Retrieval becomes harder.
  • Creators respond by increasing volume. That can worsen the loop.
How to use this

If reach is steadily deteriorating: do not automatically post more. Instead:

  • Reduce volume.
  • Pause weak formats.
  • Identify your strongest niche.
  • Return with your best content.
  • Rebuild engagement quality.
Sometimes the most algorithmically aggressive action is to post less.

52Content similarity arbitrage

  • Retrieval rewards similarity to recent interests.
  • Deduplication punishes excessive similarity.
  • Diversity systems also penalize repeated artifacts.
  • The sweet spot is topical adjacency + originality.
How to use this

When a topic becomes hot, do not copy the viral content. Create different data, a different graphic, a different argument, a different application, or a different prediction.

You want to occupy the same embedding neighborhood without being the same artifact.

53Saturation & counterprogramming

  • Diversity systems can punish repetitive same-topic candidates.
  • The strongest post can dominate a similarity cluster.
  • Unique counterprogramming can occupy less-contested inventory.
How to use this

On massive news days, ask: “Can I be the best post about this story?” If no, ask: “What valuable adjacent angle is everybody ignoring?” Being different can be more valuable than being slightly faster.

54Controversy

  • Negative heads use threshold behavior.
  • Risk depends partly on comparison with competing candidates.
  • Controversial posts can perform inside their home community.
  • Cross-community distribution can trigger more negative feedback.
  • Maximum engagement and maximum sustainable reach differ.
How to use this

Controversy is safest when it is relevant to an audience that genuinely wants the discussion. Use sharp opinions — but avoid deliberately enraging irrelevant audiences, claims likely to trigger reports, and mistaking argument volume for healthy distribution.

55Hostile virality

  • Hostile amplification creates real impressions.
  • It can create blocks, mutes, and reports.
  • It can create long-term “show less” fatigue.
  • Future similar posts can inherit suppression.
How to use this

Not all viral reach is good reach. Evaluate: Who is seeing the post? Are they potential future followers? Are they engaging positively? Is your report/mute rate likely rising?

A million hostile impressions can be strategically worse than 100,000 ideal-community impressions.

56Control AI

  • Explicit viewer preference controls exist.
  • Preference instructions can be embedding-matched to posts. A similarity threshold around 0.67 is referenced.
  • This creates personalized boosting and suppression.
How to use this

The ultimate objective is to become content people intentionally want more of. Build recurring content categories — “daily AI infrastructure breakdown,” “weekly model benchmark,” “security audit lessons” — and make the value predictable enough that users actively train their feeds toward your subject.

57Search evergreen

  • Realtime search dominates most queries.
  • Archive fallback exists when recent supply is weak.
  • Niche queries can surface old posts.
  • Low-competition keywords create evergreen opportunities.
How to use this

Create posts answering specific searchable questions — “How many GPUs does Kimi K3 need?”, “MI355X vs MI325X”, “Claude token pricing breakdown”, “How to deploy Qwen locally.” These can earn long-tail search impressions after normal feed distribution dies.

58Account recovery clocks

  • Show-less fatigue: ~140 days.
  • Dislike aggregates: roughly 7–28 days.
  • Author report/dislike rates: around 28 days in some systems.
  • User-author engagement: roughly 14-day windows.
  • Batch aggregates: around a 50-day half-life.
  • Real Graph: approximately 0.955/day decay.
  • Realtime aggregates fade rapidly.
  • Negative-content similarity eventually ages out. Cluster positioning is periodically refreshed, and new Grok classifications happen on new content.
  • A clean month can repair much soft damage.
How to use this

Most reach problems are recoverable. A practical recovery month:

Week 1
  • Cut output.
  • Eliminate problematic formats.
  • Interact with followers.
Weeks 2–4
  • One strong post per day, one niche.
  • Reply heavily.
  • Prioritize saves, video, and reference content.
  • Maintain clean negative rates.
Do not panic after three weak posts.

59Deletion during recovery

  • Deleting an actively harmful post stops new events.
  • It does not erase old reports, graph changes, or historical aggregates.
  • Mass-deleting old posts does not reset ranking history.
How to use this

Delete only to stop ongoing damage, not to perform an imaginary algorithm reset. Use deletion when reports are actively piling up, the post contains a serious error, or safety problems exist. Do not erase hundreds of old posts expecting reach to magically return.

60Recovery behavior

  • Stop generating negative feedback.
  • Reduce frequency. Increase average quality.
  • Focus on existing followers. Reply heavily. Engage outward.
  • Return to one cluster.
  • Use high-intent formats.
  • Reach fresh adjacent audiences.
  • Do not abandon the account prematurely.
How to use this
The recovery formula: quieter + narrower + better + more conversational.

You are attempting to replace recent bad data with recent good data.

61Machine-learning direction

  • Phoenix represents increased sequence-based ranking.
  • Grok represents increased semantic classification.
  • Specific parameter hacks will age poorly.
  • Real engagement, category clarity, originality, and relationships remain durable.
How to use this

Build strategies that still work if every weight changes tomorrow. The future-proof strategy: useful content, clear topic, strong relationships, original artifacts, real engagement, low negative feedback.

Do not build your entire growth system around one leaked number.

62Master strategic expectations

  • Expect followers to provide Stage 1 ignition.
  • Expect the first hour to matter enormously.
  • Expect the first eight hours to carry disproportionate value.
  • Expect 24 hours to be the main out-of-network window.
  • Expect normal For You discovery to collapse after roughly 48 hours.
  • Expect coherent likes to outperform random likes.
  • Expect replies to matter throughout the stack.
  • Expect author replies to be particularly valuable.
  • Expect video to have more scoring surfaces.
  • Expect video to have longer lifetime value.
  • Expect private-signal content to outperform visible metrics.
  • Expect native-value link posts to outperform naked links.
  • Expect hashtag stuffing to hurt.
  • Expect excessive profanity to create friction.
  • Expect hook templates to decay in effectiveness.
  • Expect repeated media to fatigue.
  • Expect excessive posting to cannibalize reach.
  • Expect strong posts to outperform high volume.
  • Expect relationships to matter more than raw follower count.
  • Expect inactive followers to lose value.
  • Expect engagement maintenance to preserve distribution.
  • Expect follower quality to matter.
  • Expect follow-for-follow tactics to carry hidden costs.
  • Expect verification to reduce retrieval friction.
  • Expect Lists and Communities to remain underused.
  • Expect exploration impressions to function as auditions.
  • Expect category clarity to matter especially for new users.
  • Expect personalization to matter more for heavy users.
  • Expect controversial content to face crossover limits.
  • Expect hostile virality to carry long-term costs.
  • Expect paid boosting to amplify existing quality rather than create it.
  • Expect most reach problems to heal over weeks.
  • Expect a clean month to repair substantial soft damage.
  • Expect show-less fatigue to remain longer.
  • Expect machine semantic classification to become more important.
  • Expect real quality to outperform algorithm gaming over time.
How to use this

If you remember nothing else, build around these twelve principles:

  • 1. Niche coherence
  • 2. Fast early engagement
  • 3. Relevant engagers
  • 4. Replies you respond to
  • 5. Native, value-rich content
  • 6. Video as a discovery engine
  • 7. Save/share/dwell value
  • 8. Low posting cannibalization
  • 9. Strong follower relationships
  • 10. Extremely low negative feedback
  • 11. Originality
  • 12. Converting strangers into followers

63Final operating model

  • Your post must first be eligible. Safety, spam, language, URL, account, and viewer-level rules can remove it.
  • Then it must be retrieved. Earlybird, SimClusters, UTEG, TwHIN, Topics, Lists, Communities, and exploration determine whether it enters competition.
  • Then it must score well. X predicts many possible behaviors.
  • Then multipliers alter the result. Author fatigue, out-of-network penalties, reply penalties, slop, negative history, and diversity change the effective score.
  • Then it must win inventory. Other creators, ads, modules, and candidate-supply constraints compete for feed space.
  • Then viewer behavior becomes new data. Likes, replies, dwell, shares, bookmarks, screenshots, profile behavior, mutes, blocks, and reports feed future recommendation.
  • The account is continuously evaluated across time. Your posts continuously update relationship, author, community, quality, and reputation features.
  • Follower count is not the primary asset. Active viewer relationships are.
  • Posts are not the scarce resource. Audience attention and healthy viewer relationships are.
The maximum-reach operating system

Think of X as a two-stage compounding machine.

Stage 1 — Win your existing audience

Publish content your current followers immediately stop on, like, reply to, save, share, watch, and visit your profile from. Respond to their replies. You are generating the evidence required for expansion.

Stage 2 — Win the target community

The initial engagement tells X what community the post belongs to, which social graphs should receive it, and which behavioral profiles resemble the people already engaging. If the new audience responds positively, distribution expands. If they generate reports, mutes, blocks, or “show less,” expansion slows and future relationships become damaged.

Before the post
  • Stay focused on one core niche.
  • Maintain relationships with active community members.
  • Know when your audience is online.
  • Avoid flooding.
The post
  • Strong opening. Clear subject. Substantial value. Original framing.
  • Native media when appropriate. 0–1 hashtags. Minimal unnecessary profanity.
  • A question or conversation opportunity, plus save/share value.
First hour
  • Reply to almost every meaningful response.
  • Let relevant followers generate coherent engagement.
  • Avoid immediately publishing a competing post.
First 8 hours
  • Continue the conversation. Watch favorite/reply velocity.
  • Let SimClusters and graph propagation work.
8–24 hours
  • Keep a winning post alive. Convert profile visitors into followers.
  • Prepare a related sequel; consider supporting content if momentum warrants it.
After the run
  • Study first-hour likes and replies, follower vs non-follower impressions, profile visits, follows, saves/shares, engagement quality, and whether the audience reached was actually relevant.
The deepest strategic rule

Do not optimize for the largest possible number of impressions. Optimize for the largest possible number of positive impressions from people whose future attention you also want.

That creates the compounding loop: relevant audience → positive engagement → stronger community identity → broader relevant distribution → high-quality followers → stronger Real Graph relationships → stronger future in-network distribution → better initial engagement → even greater out-of-network reach.

And it avoids the opposite loop: weak/random content → ignored impressions → author fatigue → lower engagement → weaker account averages → reduced retrieval → more posting → more weak impressions → further reach decline.

The algorithmic objective is not simply to create viral posts. It is to continuously build an account with strong relationships + clear community identity + high expected engagement + low negative-feedback risk + consistently original content. That is the configuration most likely to maximize both reach and impact while minimizing invisible restrictions and long-term distribution damage.

Frequently asked questions

How long does a post stay eligible in X's For You feed?

Ordinary For You eligibility falls off around 48 hours, and most out-of-network discovery systems are strongest inside the first 24. An important community score uses roughly an 8-hour half-life, which is why concentrated early engagement matters far more than a slow trickle. After 48 hours, treat the normal For You lifecycle as largely complete — though video and search can keep earning impressions much longer.

Do hashtags help reach on X?

Mostly no. Multiple hashtags are explicitly damped in X's text scoring, and hashtag stuffing reads as spam. Zero hashtags is perfectly acceptable; one genuinely relevant hashtag can still add context. Treat hashtags as context, not distribution fuel — clear, machine-readable topic language in the post itself does far more for reach.

What matters more on X — likes or replies?

Both are predicted by the ranking models, but replies carry unusually strong weight — roughly 9x a favorite in the example light-ranker configuration — and X separately models whether the author responds to commenters. Likes matter most as routing information: who likes a post helps decide which interest communities it gets distributed to. The strongest pattern is a post that earns fast, topically coherent likes plus real replies you actually answer.

My reach dropped — am I shadowbanned, and can I recover?

One weak post proves nothing — the system has daily viewer-level variation. Sustained drops usually trace to accumulating negative signals: reports, mutes, "show less" feedback, spam or duplicate-content labels, or a collapsing favorites-per-post average. Most of it heals on known clocks — dislike aggregates in roughly 7–28 days, report-rate windows around 28 days, "show less" fatigue up to ~140 days. A recovery month of fewer, better, niche-focused posts with heavy replying repairs most soft damage. Mass-deleting old posts does not reset ranking history.