The short answer: the Twitter X algorithm 2026 runs on published code, not folklore. X's open-source For You repository has the ranking weights the feed runs on. A like is worth 0.5, a reply 5, and a reply to an original post from someone you mutually follow 20. Sharing a post's link is worth 20, and a report −234. Posts older than 48 hours never enter the ranking at all.
Most "Twitter algorithm explained" posts still quote 2023 numbers, or ones nobody can source. We read the current repository on 1 October 2026 (last updated 18 September). This page sets its weights against the 2023 release and spells out what each one means for your reach.
TL;DR
What the code says | What it means for you |
|---|---|
Every action the model predicts gets a weight; the score is the weighted sum | Ranking is a prediction about each viewer, not a count of your likes |
Reply 5, rising to 20 for original posts from mutual follows | Conversations with people who follow you back are the biggest positive lever |
Share via copy link 20, share via DM 5, follow author 4 | Posts people pass along beat posts people tap |
Report −234, mute −58.8, not interested −47.52, block −31.2 | Annoying the wrong audience costs more than any like earns |
Posts over 48 hours old are filtered before scoring | Your For You window is two days, not a week |
Replies and reposts from accounts a viewer doesn't follow are filtered | A reply to a stranger doesn't reach their audience's For You feed |
The X algorithm is a published pipeline with seven stages#
The repository's README describes one request to the For You feed as a fixed sequence. Nothing about it is a black box anymore, even if some inputs (the trained models, a few safety rules) stay private.
Three details in that diagram change how you should read every growth tip:
In-network and out-of-network posts are ranked by the same model. Thunder supplies recent posts from people the viewer follows. Phoenix retrieval and SimClusters find posts from everyone else. One transformer, Phoenix, scores the lot.
The model reads the viewer, not just the post. Phoenix's main input is "the viewer's recent engagements." The same post gets a different score for every person it might be shown to.
Ranking and visibility are separate systems. Ranking sets the order. A separate service decides afterwards whether each post may be shown at all. That split explains most "my reach collapsed" mysteries, and we'll come back to it below.
Twitter X algorithm 2026 weights: what each action is worth#
The score is Σ (weight × predicted probability of the action). These are the production defaults in home-mixer/params/param.rs. X says a cron job keeps that file in sync with the primary production values.
These are the X algorithm ranking factors you can actually read off the code. Smaller weights cover the attention signals: a click into the post is worth 0.3, opening a link 0.2, video open 0.07, photo expand 0.05 and dwell 0.05. Two signals are set to zero today: a profile click and a "video quality view." That's worth knowing if you've been told profile visits are the metric to chase.
Don't read the weights as exchange rates#
The repository goes out of its way to correct one misreading: "one report cancels out 468 likes" is wrong. The weights multiply the predicted probability that a particular viewer will take each action, not your post's raw counts. X's comment in the code adds that the base rate of a report is "more than 1000x lower than a Like." The big negative weight exists so that a rare prediction can move the ranking at all.
The comment also explains why brigading has less effect than people fear. A report only counts if it happens on a post served in the Home timeline. Because the model is personalized, reports from one group mostly change what similar users see, not the post's ranking for everyone.
What changed since the 2023 open-source release#
Most guides still quote the 2023 heavy-ranker weights from the original Twitter release. Several of them no longer exist in the same form.
Signal | 2023 | 2026 | Practical change |
|---|---|---|---|
Reply | 13.5 | 5 (20 from a mutual on an original post) | Who replies now matters as much as whether they do |
Reply the author engages with | 75 | No separate weight | The famous "75x" figure is gone from the weighted sum |
Profile click | 12 | 0 | Profile visits no longer move ranking directly |
Click into the conversation | 11 | 0.3 | Curiosity clicks count for far less |
Share via copy link | not listed | 20 | New, and tied for the top positive weight |
Report | −369 | −234 | Still the heaviest weight in either direction |
If a guide tells you the author-reply bonus is 75x, it's quoting a 2023 file. Replying to your commenters still helps, because each reply is a new post that can earn its own predicted engagement. It just isn't a named multiplier anymore.
The comparison only goes so far. The 2023 file described a different model with different actions and base rates, so line up the direction of change, not the ratios.
The mutual-follow boost is the biggest change of 2026#
The 20 for a reply comes from one specific rule. In xai-value-model/weights.rs, the reply weight gets a bonus when the candidate is boost-eligible. inputs.rs defines that as not a reply, not a repost, and from an author the viewer mutually follows. The bonus is 15 on top of the base 5.
X documented the rollout in the repo's bidirectional boost change note:
What it means in practice:
Your original posts get a 4x reply weight with people you follow back and who follow you. That's the audience the boost is built for.
The boost skips replies and reposts. Only your root posts qualify, so a reply-only account doesn't benefit.
A dwell-time version was tested and not shipped.
BidirectionalFollowDwellWeightBoostsits at 0.0.
So a mutual-follow network is now an actual ranking input, where before it was just good manners. Following back the people who reply to you thoughtfully, and replying to theirs, compounds in the code itself.
Freshness: your post has 48 hours, and reach outside your followers is discounted#
The pre-scoring filters decide what never gets ranked. These are the ones that shape reach most:
Filter | What it removes | Takeaway |
|---|---|---|
| Posts older than 48 hours | A post's For You life is two days |
| Replies and reposts from accounts the viewer doesn't follow | Your reply to a big account reaches your followers' feeds, not theirs |
| For new accounts, out-of-network posts below an engagement threshold | Brand-new accounts must earn reach before strangers see them |
| Posts the viewer has already been shown | Reposting the same thing won't resurface it to the same people |
After scoring, three adjustments apply. Out-of-network posts are multiplied by a factor below 1. Repeated posts from the same author decay toward a floor. And a new-author boost lifts posts from authors with few impressions toward a target position. The parameters behind it, ColdStartImpressionThreshold (200), ColdStartFollowerCap (50,000) and ColdStartMaxPostAgeSecs (7,200 seconds), suggest it targets small accounts and posts under two hours old.
That's the real basis for "post when your audience is awake." It isn't a mythical 30-minute test pool of a few hundred accounts; no such mechanism appears in the code. It's a two-hour cold-start window for small authors, inside a 48-hour eligibility window for everyone. If you're working out your windows, the nine posting times that actually work is a faster start than guessing.
Replies still matter, just not the way most guides say#
Because out-of-network replies are filtered from For You, a reply on a large account works through the conversation view. People who open that thread see it, and some of them tap through to you. That's still a solid way to get discovered. But it's a separate channel from For You ranking, and the 2026 weights no longer pay directly for the profile click it produces.
Links: there's no explicit penalty in the weights#
"X buries links" is the most repeated claim about the algorithm. In the published weighted sum, opening a link carries a small positive weight (0.2). There is no negative weight for containing a URL.
What the code can't rule out is the indirect effect: if viewers are less likely to reply, dwell or share on a link post, its predicted engagement is lower and so is its score. The weights alone don't settle whether "put the link in the first reply" helps. Test it on your own account, and read how to post links on X without killing reach for the measured trade-offs.
X algorithm reach and visibility: political and sensitive content#
Not every reach problem is a ranking problem. Visibility filtering runs after ranking and answers one of three things for each post and viewer: ALLOW, INTERSTITIAL (shown behind a tap-through warning) or DROP.
The labels it reads come from systems that are now in the repository:
agathalabels accounts from how others respond to their posts: blocks, reports and spam reports relative to favorites.bdsmreads an account's action sequence over time for signs of inauthentic behavior.user-cred-v2runs PageRank over follow and engagement edges to give each account a credibility score.groxand the media models classify posts for spam, adult content and violent media as they're published.
One rule matters for reach more than the rest: some visibility rules only drop a post when it's a recommendation from an account the viewer doesn't follow. The same post is allowed for followers. That matches the pattern people describe as being "throttled". Impressions arrive almost entirely from existing followers, the non-follower share collapses, and engagement rate looks fine.
Political content has a documented case too. Under Brazil's electoral law, For You runs a Brazil2026ElectionFilter. It removes posts from accounts reported to Brazil's Electoral Court for the 2026 election, unless the viewer follows the account.
Check Under the Hood before you guess#
X now runs a pilot transparency tool, Under the Hood, that shows aggregate statistics about the visibility-limiting labels on your account and posts. Since 18 September 2026 it also reports content withheld in a country after a legal demand, and names the country. Pair any label it shows with the rule in visibility-filtering/rules/registry.rs and you know what's actually happening.
That beats every third-party "shadowban checker." If the tool isn't available on your account yet, the logged-out search tests in our shadowban guide are the fallback.
X Premium: subscription status is read, but there's no multiplier in the score#
Candidate hydration includes "subscription status," and one filter removes subscriber-only posts a viewer can't access. We found no Premium multiplier in the published weighted sum or its three adjustments. If Premium changes reach, it does so somewhere this repository doesn't show, such as reply ordering inside a conversation, which isn't part of the For You code.
Treat any precise "Premium gets X% more reach" figure with suspicion unless it links to a source you can check.
What to do with this: a checklist built from the code#
Every line below maps to a weight or filter the Twitter X algorithm 2026 code actually contains:
Build mutual follows in your niche. Original posts get a reply weight of 20 from mutuals versus 5 from everyone else.
Write posts people pass on. Copy-link shares (20) and DM shares (5) outweigh a like (0.5) by 10–40x per predicted action.
Ask for replies, not likes. A predicted reply is worth 10 likes, before any mutual boost.
Protect your negative-feedback profile. Mutes (−58.8), "not interested" (−47.52) and blocks (−31.2) come from showing posts to the wrong people. Stay on topic for the audience you have.
Publish for a two-day window. Nothing older than 48 hours is ranked, and the small-author cold start lasts two hours.
Reply to your commenters. Not for a 75x bonus that's gone, but because each reply is a fresh post your mutuals can engage with.
Diagnose visibility separately from ranking. Non-follower share first, Under the Hood second.
If you want the full growth routine built on these signals, How to Grow on X walks through it day by day.
Where a tool helps, and where it can't#
Most of that list is human work. Building mutuals, writing replies and choosing what to post can't be automated without crossing X's automation rules. Any tool that fires replies at strangers for you is a fast route to the negative labels above. Our guide to what X allows for auto replies covers where that line sits.
What a tool can take off your plate is the mechanical half around it. ReachMore is a web app with a composer and scheduler, so original posts land inside your audience's two-day window. AI drafts start from your own voice. A mention queue drafts a response to everyone who @-mentions you and holds it until you approve it, so the people already talking to you get an answer while the conversation is still alive. Knowing how your tool watches for mentions decides whether you can keep that pace.
Frequently asked questions#
How does the X algorithm work in one sentence?#
A transformer called Phoenix predicts how likely you are to like, reply, share, mute or report each candidate post, and the feed sorts posts by a weighted sum of those predictions. Visibility rules then remove or blur anything labelled as not showable to you.
Is the X algorithm open source in 2026?#
Yes, most of it. The For You pipeline, the scoring weights, the visibility-filtering rules and the Phoenix training code are in xai-org/x-algorithm under Apache 2.0. Grox's LLM prompts and some botmaker rules are kept private to reduce gaming, which X says openly in the README.
Does liking your own post help?#
No. SelfTweetFilter removes your own posts from your For You feed, and a like is the smallest positive weight (0.5) anyway. Time spent getting one genuine reply is worth far more.
Do profile visits still matter for the algorithm?#
Not directly. The 2026 profile-click weight is 0, down from 12 in 2023. A profile visit still matters because it can turn into a follow, and "follow author" carries a weight of 4.
Key takeaways#
The weights are public. Reply 5, reply from a mutual on an original post 20, copy-link share 20, like 0.5, report −234. They multiply predicted probabilities, not raw counts.
The 2023 numbers are stale. The 75x author-reply bonus and the 12x profile-click weight aren't in the 2026 weighted sum.
Mutual follows are a ranking input now. The July 2026 bidirectional boost made them the biggest lever on reply weight.
Your window is 48 hours. Small authors get a cold-start lift for posts up to two hours old.
Visibility isn't ranking. Check your non-follower share, then Under the Hood, before changing what you write.
The algorithm rewards conversations with people who know you, and posts worth passing on. Keep that part human, and let software handle the queue around it. Open ReachMore for a composer, scheduler and mention approval queue on a credit wallet: one-time top-ups, credits that never expire, and charges only when an action succeeds.