The X algorithm is documented and unsealed. It filters roughly 500 million daily posts down to a personalized 1,500 per user, ranks them with a weighted engagement-probability score, and serves the top results. The published weights are specific. Replies count 13.5x. Author-engaged replies count 75x. Reposts count 20x. Likes and bookmarks count around 30x. Verification adds a 4x follower multiplier and 2x general multiplier. Reports cost 369x. Negative feedback costs 74x. Hashtag-heavy posts get a 1.7x deboost. The decay half-life is 360 minutes.
Standard advice reads these weights as a checklist to game. Get verified, reply to your own comments to trigger the 75x author-reply weight, avoid hashtags, post during peak hours, push for the 2-minute engagement window. All technically correct. All voice-blind. The algorithm amplifies whatever's underneath. If what's underneath is a voice that gets people back to your profile, the weights produce compounding reach. If what's underneath is category-default content, the weights produce marginal incremental impressions on posts that don't convert. The math is the same; the underlying asset isn't.
What the weights actually reward, voice-first
The replies multiplier is the largest non-penalty weight in the published table. 13.5x for replies, 75x if the author engages back. The implication for voice-first creators is that voice-rich replies on bigger accounts are the single highest-leverage move on the platform. A 100-word substantive reply to a 1M-follower account that gets author-engaged is worth more in the ranking score than 5 of your own original posts at average distribution. The voice-first reply strategy is the playbook for what that actually looks like in practice: 5 to 10 voice-rich replies a day, not 30.
The likes/bookmarks weight rewards content that prompts saving for later. Bookmarks are the strongest privacy-respecting signal that your content has reference value, and the algorithm reads them at ~30x. Voice-rich evergreen posts pick up bookmarks for weeks; viral hot takes don't. Twitter bookmarks as voice-research infrastructure covers the inverse half of this (what to bookmark from others). What matters for distribution is the inbound: bookmarks-per-impression is a cleaner voice-fit signal than likes-per-impression.
The penalties are bigger than the boosts. Voice protects you here.
Read the published weights again with the negatives in mind. Reports cost 369x. Negative feedback costs 74x. Offensive-language detection costs 10x. Unknown language costs 20x. The single largest factor in the ranking is what triggers a report, and reports are largely a function of writing register. Drama-bait, contrarian-for-its-own-sake, and category-default outrage take ride the same algorithm but accumulate negative-feedback signal as they scale. Voice-first creators in their actual register rarely trigger reports because they're not performing on top of their voice; the natural-register content sits inside the writer's category-norms and doesn't read as inflammatory to in-category readers.
The hashtag deboost (1.7x) and the offensive-language deboost (10x) are both surface-level pattern matches. Voice-first creators almost never hit these because the patterns that trigger them are pattern-match content (hashtag spam, ALL-CAPS rage-tweeting) that the writing voice would naturally avoid. The penalties are bigger threats to template-driven accounts than to voice-first accounts.
The 360-minute decay and why voice-rich evergreen is the cheat code
Every post loses ~50% of its ranking weight every 6 hours. After 24 hours, the post is functionally dead in the timeline algorithm. The exception is content that keeps picking up engagement (replies, bookmarks, quote-tweets) past the initial window. Voice-rich evergreen posts get re-surfaced for weeks via inbound interactions that the algorithm reads as continued relevance. Hot-take posts get a 6-hour spike and then disappear. The decay function structurally rewards content that has more than 24 hours of useful life, and voice-rich evergreen is what produces it.
The implication for posting cadence: 3 voice-rich posts a week that get re-surfaced over weeks outperform 21 hot-take posts a week that each die in 6 hours. The ranking math says so. What compounds versus what looks like it does covers the broader version of this argument.
Verification (4x followers, 2x general) as a small modifier, not the lever
Premium gives a 4x multiplier to follower distribution and 2x to general. Real but modest. The 10 to 15% visibility lift that creators measure in practice is the rough output of this weight after it's combined with everything else. Premium amplifies what's already in your style. It doesn't manufacture audience. The voice-first take on X Premium covers when the math works and when it doesn't.
The 2-minute engagement window and long-form
The algorithm rewards content that holds attention for ~2 minutes (the published threshold for the 'time spent' factor). Long-form posts and threads can clear this threshold. So can voice-rich shorter posts that prompt re-reads or replies that pull the reader back to the post. The 'just write longer posts' advice is shape-correct and voice-blind: a long category-default post bores readers off in 20 seconds; a short voice-rich post can hold attention for 90 seconds while readers re-read the closing.
Voice-rich content tends to clear the 2-minute window naturally because the writer's specific framings and specific observations don't compress well; the reader has to slow down to actually take in what's being said. Category-default content compresses to skimming because the reader has seen the same shape 50 times already.
What the algorithm doesn't measure (and what voice protects against)
- Audience quality. The weights treat a like from a bot the same as a like from a niche peer. Voice-rich content tends to attract niche peers because it doesn't perform like engagement-farming content; the inbound is structurally higher-quality.
- Re-circulation. The published weights don't directly include 'this person comes back to read your timeline.' But the indirect signals (replies-per-impression, profile-clicks-per-impression) capture some of it. Voice produces re-circulation; templates don't.
- Off-platform compounds. The algorithm doesn't see your newsletter signups, your DMs that become consulting calls, your products that sell. Voice-first creators optimize for these; the algorithm scores their content the same as everyone else's, but the off-platform math is what actually pays.
- Voice consistency over time. A timeline that reads as one writer's voice over 12 months compounds via readers attaching to the writer, not via any single post's algorithmic boost. The algorithm sees each post independently; the audience doesn't.
Did the 2026 Phoenix rewrite change these weights?
Fair question, because the weights above come from the 2023 open-source release and X's ranking stack has been rebuilt since. The architecture changed more than the conclusion did. The 2026 rewrite moved the ranker toward a transformer-based model that scores many engagement actions jointly rather than summing a fixed table of hand-set multipliers, and it leans harder on out-of-network retrieval (surfacing creators you do not follow based on embedding similarity rather than keyword match). What did not change is the direction of the incentives: replies and author-engaged replies still dominate the positive signal, negative feedback and reports still cost far more than any boost adds, and recency still decays fast. The deep-technical reading of the rewrite is in the X-Algorithm series at the May 2026 X algorithm: why voice wins when the ranker becomes a transformer and Phoenix's 19 engagement heads. The voice-first reading actually gets stronger under the rewrite, because a transformer that scores engagement jointly and retrieves on embedding similarity rewards a coherent, recognizable voice more than a fixed-weight table ever did. Treat the published weights here as the still-useful mental model; the rewrite changed the machinery, not the lesson.
Does the X algorithm punish AI-generated content?
There is no weight in the published table for "this looks AI-generated," and the 2026 ranker does not run a detector that demotes posts for being machine-drafted. But the algorithm punishes the things AI-shaped content tends to produce, which arrives at the same outcome by a different route. Voice-flat content earns fewer of the bookmarks and author-engaged replies that drive the largest positive weights, because nobody saves or genuinely responds to a post that reads like a hundred others. It clears the dwell-time threshold less often, because a generic post compresses to a two-second skim. And at scale it accumulates the negative-feedback and not-interested signals that cost far more than any boost adds, because a feed full of interchangeable AI posts is exactly what the audience mutes. So the honest answer is that the algorithm is indifferent to how a post was written and highly sensitive to how the audience reacts to it, and AI-shaped content reliably produces the reactions the weights penalize. Writing in a recognizable voice is not a way to trick the ranker; it is a way to earn the engagement the ranker already rewards.
What actually moves distribution for a small account?
The published weights read differently depending on how big your account already is, and the standard advice ignores this. For a large account, original posts reach a real audience on their own, so the marginal move is posting cadence and bookmark-worthy evergreen. For a small account, original posts reach almost no one regardless of quality, because For You distribution is gated on early engagement the account cannot yet generate. The highest-leverage move for a small account is therefore replies, not posts: a voice-rich reply on a large account borrows that account's distribution, and the 13.5x reply weight (rising sharply if the author engages back) does work an original post at zero base distribution cannot. The practical sequence is replies-first until the account has enough recognized-voice followers to give original posts a base, then a shift toward original evergreen once the base exists. Trying to grow a small account on original posts alone is the most common way to conclude, wrongly, that the algorithm is rigged. It is not rigged; it is gated on a signal small accounts have to earn through replies first. And if you genuinely suspect a block rather than slow growth, our free shadowban checker runs the one visibility test that is actually testable, and says so honestly when it cannot.
Can you game the algorithm without hurting your voice?
The tactics in the standard advice (get verified, reply to your own posts to trigger the author-reply weight, post in the peak window, avoid hashtags) are real and mostly harmless in moderation. The line is whether the tactic changes what you write or only when and how you publish it. Getting verified, timing a post, and skipping hashtags are publishing-layer choices that do not touch the content, so they are free to use. Writing drama-bait to farm the reply weight, manufacturing outrage to trigger quote-tweets, or running reply-bot automation to hit reply volume are content-layer choices that buy a short-term ranking spike at the cost of the negative-feedback signal and the audience trust that compound against you over months. The discipline is to take every publishing-layer optimization and refuse every content-layer one, because the content-layer "optimizations" are exactly the ones that quietly convert your account into the category-default outrage feed the audience is learning to mute. Optimize how you ship; never let the weights edit what you say.
Should you rework your strategy every time the algorithm updates?
No, and this is the quiet advantage of the voice-first reading. Every algorithm update spawns a wave of growth threads decoding the new weights and prescribing new tactics, and creators who optimize at the tactic layer have to relearn the platform each time. The voice-first approach optimizes the inputs the algorithm has rewarded across every version of itself (genuine replies, content people save and return to, an absence of the signals that trigger reports), so it survives the updates without a rewrite. The 2023 fixed-weight table and the 2026 transformer rank posts differently under the hood, but both reward a creator the audience recognizes and engages with, because every version of the ranker is ultimately a proxy for audience reaction. The practical rule: read the update notes to understand the machinery, but do not let a weight change pull you off the one strategy that has worked across all of them. Optimize for the audience the algorithm is trying to model, not for the current model of it.
The voice-first algorithm playbook
- Replies as the primary distribution surface. 5 to 10 voice-rich replies a day on bigger accounts. The 13.5x to 75x weight does most of the work. Skip reply-bot automation; the patterns get caught and the negative feedback compounds. The case against reply-bot automation at scale covers why.
- Voice-rich evergreen, 2 to 4 a week. Each post designed to be readable in 6 months. The 360-minute decay rewards content that keeps getting picked up across the days that follow.
- Bookmarks-per-impression as the cleanest distribution proxy. Track it. If it's flat, the audience isn't finding reference value; usually a voice-flatness symptom.
- Verification at audience-tier where the 4x multiplier produces measurable lift (typically 3K+ followers). Below that, the lift is below noise.
- Avoid the penalty triggers structurally, not tactically. Don't write the content that earns reports; the voice register that earns reports is the same register that suppresses long-term audience attachment.
- Don't use the algorithm as a learning signal for voice. The weights reward engagement; engagement is downstream of voice. Use the algorithm as a publishing constraint, not as a voice editor.
Where Auden fits
Auden trains on a creator's full profile (100 to 200 posts, replies, threads, and images across 10 signals of voice) and produces drafts that match the writer's register. The algorithm rewards what voice produces. The role of Auden in this picture is not to game the weights; it's to keep voice consistent across the cadence the weights reward. A timeline that reads as one writer's voice over 90 days picks up the algorithm's small boosts on every post; a timeline that drifts off-voice picks them up half the time and loses the readers who came for a specific writer.
The algorithm is mechanical. Voice is the asset it amplifies. The voice-first playbook treats the weights as a small modifier on a much larger system: who you're recognized as, who comes back, and what they do off-platform. The published weights confirm the playbook; they don't replace it. For the writer-facing playbook that operationalizes this mechanical-amplification-of-voice reading into a refusal of the four shortcuts the standard growth guides still recommend (bought followers, engagement pods, AI-template hook patterns, sycophantic reply-spraying) plus the five disciplines that produce real organic growth on the 90-to-180-day timeline the weights and the audience-side filtering combined actually permit, see how to grow on X in 2026 without buying followers or running engagement pods.