X (Twitter) Algorithm 2026: What’s Confirmed, Contested, and Myth

X publishes the code that ranks its For You feed, and much of what you read about the X algorithm still describes the 2023 version.

By
Dana Pritchard
Marketing Strategist & Tech Journalist
Dana Pritchard covers marketing software, social platforms, and the tools small teams and households pay for every month. She looks hardest at what a plan costs...
- Marketing Strategist & Tech Journalist
23 Min Read
23 Min Read
Disclosure: This post may contain affiliate links. If you click one and make a purchase, we may earn a commission at no extra cost to you.
X algorithm illustration: an open-air Jacquard loom weaves many colored threads into one long ribbon while an old chain of punch cards lies faded on the courtyard stones
Follow meikuio.com on Google

Add us as a Preferred Source to get our latest articles and updates directly in your Google Search & Top Stories.

Highlights
  • Of 18 popular X algorithm claims, X's published code backs 4, leaves 6 open and contradicts 8; the weights multiply your predicted reactions, not a post's like count.
  • No post older than 48 hours reaches For You, a predicted reply between mutual followers is worth 20.0, and small accounts get a two-hour new-author window.
  • There's no reset button: mute, unfollow, choose "Not interested in this post" and change what you engage with for a few days.

The X algorithm is the system that picks and orders the posts in your For You feed. Since January 2026, X has published its code on GitHub. The code scores each post by how likely you are to like, reply, share or report it. Much of the popular advice still describes Twitter’s 2023 version.

Contents
- Advertisement -

I didn’t run an experiment on an X account for this. On October 15 and 17, 2026, I read X’s 2026 repository. Every code reference below points to commit 862a123. I also read the weights file X says is synced to production defaults, its change notes and Help Center pages, and the 2023 code it replaced. Then I took the claims that keep circulating, including the ones Google’s AI answer repeats, and checked each against a specific file. If you only came to clean up your feed, there’s no reset button, but muting, unfollowing and “Not interested” do the work; the reset steps are further down.

How does the X algorithm work in 2026?

The X algorithm builds your For You feed fresh on every request. It gathers recent posts from accounts you follow and accounts you don’t, then drops anything you’ve blocked, muted or already seen. It predicts how you’d react to each remaining post and adds those predictions into a single score. Ranking and permission to appear are separate steps.

X’s own README lays out the pipeline in seven stages, and it describes the system more clearly than anything in X’s Help Center. In plain terms, this is what happens each time you pull to refresh:

  1. It loads your context: your recent engagements (the main input to the model), who you follow, your blocks and mutes, your muted words and what you’ve already been shown.
  2. It pulls candidates. A service called Thunder returns recent posts from accounts you follow. For accounts you don’t follow, Phoenix retrieval finds posts whose vectors sit closest to yours, SimClusters finds posts through clusters of accounts that engage with the same things, and a shared list adds popular recent posts.
  3. It loads each post’s text, media, author details, labels and engagement counts.
  4. It filters before scoring: duplicates, posts older than 48 hours, your own posts, blocked and muted accounts, muted keywords, posts you’ve seen and subscriber-only posts you can’t open.
  5. It scores. A model called Phoenix predicts the probability of each action you might take, and those probabilities are combined with the weights X publishes.
  6. It keeps the top posts by score, after adjustments for repeated authors, accounts you don’t follow and new authors.
  7. It runs visibility rules on the survivors, then blends in ads, Who to Follow cards and prompts.

The detail that matters most for everything below is step five. The model doesn’t look at how many likes a post has and multiply that. It estimates how likely you, specifically, are to like it, reply to it or report it, and the weights multiply those estimates. The same post can score very differently for two people, because their own histories drive the predictions.

X algorithm illustration: two brass balance scales hold the same blank card, one pan sinking and one rising, in front of two different windows

That is also why the Twitter algorithm you read about in 2023 guides is the wrong reference. The 2023 Twitter algorithm was a Scala service with a separate “heavy ranker” and hand-tuned rules. The 2026 code is a different codebase, a mix of Rust, Scala, Python and Java with the feed service itself in Rust. Its ranking model is trained on engagement data and reads your own history on every request. Some names survived, such as Home Mixer and SimClusters, and that overlap is where much of the confusion starts.

Of the 18 claims I checked, the code backs 4 outright, leaves 6 open and contradicts 8. The table gives the verdict for each and the file it rests on, at commit 862a123 unless it says 2023. Tap the + on a row to see the line that decides it.

A verdict here means what the published code says. X’s README states that cron scripts copy production defaults into the repository, but experiments can run with other values. X publishes the code under the Apache 2.0 license, including Phoenix’s training code, but the trained model itself isn’t public. So “confirmed” means the code does it and X says the code reflects production. It doesn’t mean I watched it happen on a live account.

I also checked the answer most people get first. On October 17, 2026, Google’s AI Overview for “x algorithm” made three claims. It said ranking uses “over 1,000 signals,” that “Interaction in the first 30 minutes after posting heavily influences wider reach,” and that external links lose “up to 50%–90%” of their reach. None of those three statements has a source in the 2026 code. The signal count echoes Twitter’s 2023 engineering post, which said its ranker “takes into account thousands of features.” When I traced the other two back through the guides that repeat them, the earliest versions I found cited no experiment or code.

What the X algorithm code confirms

Four claims about the X algorithm hold up line by line. X scores predicted reactions rather than counts, weights a predicted reply more heavily when you and the author follow each other, and caps the feed at posts from the last 48 hours. Muted words are filtered out before scoring. The code also confirms three adjustments that run after scoring: a discount for strangers, a decay for repeated authors and a lift for new authors.

The weights multiply your predicted reactions, not a post’s counts

X’s ranking doesn’t multiply a post’s like count. A comment in its scoring code says each weight “multiplies the *predicted* probability of that action,” and the README adds that the weights “do not scale the raw engagement counts.” The final score is a sum: each predicted probability times its weight, positive actions adding and negative actions subtracting. The arithmetic lives in xai-value-model/scoring.rs; the weights live in home-mixer/params/param.rs.

For a social team, this means the X algorithm doesn’t let a post “bank” engagement. A thousand likes from one audience don’t raise its score for a viewer whose history predicts they’d scroll past. The post earns a better prediction among people who resemble those who already engaged.

Posts from people you mutually follow get a reply boost

On X, a predicted reply to someone you mutually follow is worth 20.0 instead of 5.0, and X documented the change in its own change log. X calls it the bidirectional follow boost and describes it as increasing “the weight on the predicted probability that you’ll reply to a post from one of those authors who you mutually follow.” The base reply weight is 5.0. The boost adds 15.0 for a mutual follow, so a predicted reply from a mutual is worth 20.0.

The history is in docs/BIDIRECTIONAL_BOOST_CHANGE.md. X started an A/B test on July 10, 2026 with boost values of 5, 10, 15 and 20. It rolled out 20 to many users on July 13 and cut it to 15 on July 24. Its stated reason was feedback that people weren’t seeing enough World Cup discussion from accounts they didn’t follow. A boost for mutuals, in other words, came at the expense of strangers, and X lowered it after users said they were missing that discussion.

Posts older than 48 hours never reach For You

For You drops any post older than 48 hours before scoring. The limit is a constant: MAX_POST_AGE = 48 * 60 * 60 in home-mixer/params/config.rs, applied by AgeFilter before anything is scored. A post that is two days old can still be read on your profile, in search or in Following, but it won’t be recommended. For a team planning a campaign, that gives organic reach through the X algorithm’s main feed a maximum shelf life of two days.

X algorithm illustration: fresh bouquets fill the sunlit front of a flower stall while two-day-old flowers wait in the shade behind it

Strangers get a discount, repeated authors decay, new authors get a lift

Three adjustments run after the score, and their values sit in vm-ranker/params.rs and param.rs. Posts from accounts you don’t follow are multiplied by 0.75, and the README says the same discount applies to replies and reposts posted by accounts you do follow. Repeated authors decay on a curve: the second post from one author in the same feed refresh keeps 62.5% of its score, the third about 44%, and later ones approach a floor of 25%. Both are confirmed in the README’s own words as an “Out-of-Network Discount” and “Author Diversity.”

The new-author boost is the part small accounts should know about. The rule sits in home-mixer/scorers/author_cold_start.rs. An original post (not a reply or a repost) qualifies if its author has at most 50,000 followers, the post is under two hours old and it has fewer than 200 views on Home. Only one qualifying post gets the lift each time a viewer’s feed refreshes. If it would have ranked lower, it is lifted to about 16th place in that viewer’s ranked list, and nothing in the rule checks how old the account is. The 200-view limit counts views from everyone, followers included, so a post from an account with a few thousand active followers can pass 200 views, and lose the lift, before many strangers see it. Those thresholds changed recently: the version synced on August 12 used 1,000 views, 1,000 followers and 24 hours.

One fresh original post from an account with up to 50,000 followers can be lifted to about 16th place in your feed during its first two hours.

Muted words are removed before scoring, with one caveat

X removes posts containing your muted words before it scores anything. MutedKeywordFilter sits in the pre-scoring list, so a post matching a word you’ve muted is dropped before the model ever sees it. That makes muting more reliable than tapping “Not interested,” which only feeds the model a signal. The caveat comes from an open GitHub issue filed on October 16, 2026. The issue says the server waits 200 milliseconds for your saved settings and falls back to ViewerData::default() if that lookup times out (home-mixer/server.rs, lines 202 to 212). I confirmed the fallback in the code. What the default contains lives in a library that isn’t in the repository. So the published code can’t settle the issue author’s conclusion that your mutes can silently drop out for that request.

What’s contested about the X algorithm

Six claims stay open. In some, the code says one thing and the platform’s behavior or X’s own documents suggest another. In others, X hasn’t published the deciding piece. Premium is the one most teams ask about, so I start there. The rest are links, replies, Grok, mass reporting and the Following tab.

Does X Premium boost your reach?

Not through any weight X has published. Nothing in the 2026 weights file or the score arithmetic mentions Premium, verification or a checkmark. I searched the full repository at commit 862a123 for the usual names: premium, blue verified, verified and subscription level. The 2023 code was different. Twitter’s HomeGlobalParams.scala carried a multiplier of 4.0 for posts from Blue-verified authors you follow and 2.0 for those you don’t, at commit ec83d01, lines 89 to 101. That is where the “Premium gets a 2x to 4x boost” claim comes from.

Why “contested” and not “myth”: Phoenix’s message schema still carries an author’s isBlueVerified flag and subscriptionLevel inside a record called UserSafetyInfo. I found no published Phoenix or scoring code that reads those fields, but the trained model’s inputs aren’t fully public, so I can’t rule out that it learns from them. What I can say is that a team buying Premium seats is not buying a multiplier the X algorithm code shows.

Premium does change two other things. X now pays creators through Original Content Rewards, which count impressions from Premium users, so Premium readers decide what a creator earns. And regulators have taken the checkmark seriously. On December 5, 2025, the European Commission fined X €120 million, partly because it found the blue checkmark’s design deceives users about who is verified.

There’s no link penalty in the code, and I found no source for the popular percentage. Opening a link carries a small positive weight (OpenLinkWeight, 0.2), not a negative one. No weight or rule in the scoring code penalizes a post for containing a URL. The code does count a post’s external links, but in the published defaults it sends that count only to a reranker “debias” input that is switched off (VMRankerSendDebiasInputs, false). When I traced the “50 to 90 percent” figure back through the creator guides that repeat it, the earliest versions I found cited no experiment or code line.

It stays contested because of what the model learns from. Phoenix’s list of recorded actions includes how long someone stays outside X after clicking an external link, in buckets from under three seconds to over sixty. If people who tap your link leave and don’t come back to reply or repost, the model learns that your link posts produce fewer of the actions it rewards. That is a behavioral effect, not a rule, and it varies by audience. The same debate runs on other networks, as our look at LinkedIn’s link penalty claims shows.

Are replies the strongest signal?

Only between mutual followers, and even there a reply only ties the top weight. A predicted reply weighs 5.0, exactly the same as a quote or a share by DM, and well below a share by copied link at 20.0. With the mutual-follow boost, a reply rises to 20.0 as well. So advice to “bait replies” from strangers overstates what the code rewards, while advice to build conversations with people who follow you back matches it.

Does Grok rank the X For You feed?

Not the Grok chatbot. The ranker is a transformer, and X’s own documents describe its link to Grok differently over time. Phoenix’s README says “Earlier releases shipped a sample transformer ported from the Grok-1 open source release,” while the current release ships “the production implementation itself.” The European Commission, opening a new investigation on January 26, 2026, described “the impact of its recently announced switch to a Grok-based recommender system.” Today’s main README just says posts are ranked “using a transformer model.”

The safest reading is that the ranker is a transformer descended from Grok-1’s architecture, not the Grok chatbot reading your feed. Separately, a system called Grox runs classifiers for spam, adult content and violent media as posts are published. Its language-model prompts are among the files X keeps private. If you want the chatbot side of the story, our Grok vs ChatGPT comparison covers it.

Can mass reporting kill a post?

X says no, and its reasoning is in the code. Because predictions are personalized, “reports from bad actors will primarily affect recommendations for users who are similar to the bad actors,” a comment in scoring.rs argues. The comment adds that a report counts only on a post served in Home Timeline: “Directly navigating to a post (i.e., coordinating via groupchat) has no ranking impact.” That is a plausible design, and it’s still X describing its own system. The ranking code can’t tell you what the moderation side does with a pile of reports. Moderation runs on a separate set of rules, some of them unpublished.

Is the Following tab chronological?

X’s Help Center says yes: “The Following timeline displays posts only from the accounts you follow and they are shown in reverse chronological order.” The code points the other way. For a request limited to followed accounts, it labels the surface HomeTimelineRankedFollowing and now applies a separate muted-keyword filter to that feed. The help page carries no date, so I treat the strict “chronological” promise as unconfirmed for 2026. I’d check the Following tab on your own account.

What’s myth about the Twitter algorithm in 2026

Eight claims about the Twitter algorithm fall apart against the 2026 code. Several of them aren’t invented. They’re real numbers from the 2023 release, or arithmetic done on real weights without the context X itself attaches. That’s why they sound authoritative and keep getting repeated.

“One report cancels out 468 likes”

It’s a myth built on real arithmetic: in the August 2026 weights a report was −234 and a like 0.5, and 234 divided by 0.5 is 468. X added a note to its README on August 14, 2026 saying “it’d be incorrect to see that a report has 468 times higher weight than a like and conclude” that one report cancels out 468 likes. The scoring code explains why: “the baseline probability of a Report is more than 1000x lower than a Like,” so the weight has to be large for the prediction to matter at all.

The ratio isn’t even current. In param.rs as synced on October 16, 2026, a like is 0.66 and a report is still −234, so the same division gives about 355. Mid-August coverage that turned these weights into multiples of a like used the old like weight. The equivalence was misleading on the day it was published and is out of date now.

- Advertisement -

It’s a myth with one true half. Copying a post’s link carries the largest positive weight in the 2026 file, 20.0, tied only with a reply from a mutual follow. The “40 likes” figure comes from dividing by the old 0.5 like weight; with today’s 0.66 it would be about 30. Both versions make the same mistake as the report math, treating a weight on a probability as an exchange rate between counts.

“The first 30 minutes decide a post’s reach”

There is no 30-minute window for posts anywhere in the code. The time limits that exist are 48 hours for any For You candidate and two hours for the new-author boost. The only 1,800-second values in the parameters belong to a filter for brand-new viewer accounts, not to posts. Early engagement can still help, because the model learns from it, but no rule cuts a post off at half an hour.

“The feed starts from 1,500 posts, ranked by a 48-million-parameter network”

Both numbers are Twitter’s own, from its March 31, 2023 engineering post: “we attempt to extract the best 1500 Tweets” per request, and “Ranking is achieved with a ~48M parameter neural network.” They described that year’s system. The 2026 configuration caps Phoenix retrieval at 1,000 results and Thunder at 1,200, and the ranker is the Phoenix transformer. A guide that cites 1,500 and 48 million in 2026 is describing the old Twitter algorithm, not the feed X runs now.

“The algorithm had a special flag for Elon Musk”

It was in the code Twitter released in 2023, it was gone the same day, and it wasn’t a boost. Twitter’s released code contained author_is_elon next to author_is_power_user, author_is_democrat and author_is_republican. Twitter used them to count how often it served posts from those groups. Twitter deleted all four on March 31, 2023 in a commit titled “Remove stats collection code measuring how often Tweets from specific user groups are served.” I found nothing like it in the 2026 repository.

“Hashtags hurt reach” and “video gets a bonus”

Neither has a rule behind it. In the ranking code, hashtags show up only in the filter that removes words a viewer has muted; there’s no hashtag weight or penalty in scoring. Video has no format bonus either: in the current weights, opening a video is worth 0.02 and a video quality view 0.0. What helps a good video is dwell time, which the model weighs like dwell on any other post.

“The 4-1-1 rule is how the algorithm works”

The 4-1-1 rule (four pieces from others, one repost and one self-promotional post) is a content-marketing guideline that predates both repositories. It may be a sensible posting mix for a brand account. I found no trace of it in the X algorithm code, and the code has no notion of the ratio between your own posts and other people’s.

The weights X published, and how to read them

The X algorithm’s weights are in home-mixer/params/param.rs, a file whose first line says it is “mirrored from config feature-switch defaults; last sync 2026-10-16T16:00:32Z.” Read them as how much each predicted action moves a post’s score for you, not as prices of likes. The chart shows the positive weights as of that sync.

The negative side is shorter and heavier: “not interested” −52.272, block −31.2, mute −58.8 and report −234.0. A small “not dwelled” weight of −0.02 penalizes posts people scroll straight past. X’s own comment above these numbers says they reflect “a combination of how much an action is valued in ranking and typical propensities of these actions across the X network (e.g. negative feedback is overall rare).”

Two cautions before anyone builds a posting plan on these numbers. First, they move. Between the August 12 and October 16 syncs, the like weight rose from 0.5 to 0.66 and “not interested” went from −43.2 to −52.272. The new-author thresholds changed too. Second, the repository holds a second set. The reranking service’s vm-ranker/params.rs lists a like at 0.5 and “not interested” at −47.52 under the same setting names. Home Mixer asks that service to recompute the score on every request. Which values win depends on a production config that isn’t public. The README names param.rs as the file mirrored from those defaults, so that’s the set I use. But the duplicate is a reminder that reading one file isn’t the same as knowing what ran. The file’s first line shows when it last synced, so check it before you trust any weight here, mine included. When I rechecked on October 24, the file had synced again on October 23 and every weight in this article was unchanged.

For comparison, the 2023 Twitter algorithm’s heavy-ranker README printed its own table, dated “(April 5, 2023).” It gave a reply engaged by the author 75.0, a reply 13.5, a like 0.5 and a report −369.0, and warned that the weights “can be adjusted at any time.” Those are the numbers behind posts that still claim a reply is worth 27 likes or that one report costs 738 likes: 13.5 and 369 divided by 0.5.

What changed since the 2023 Twitter algorithm

Almost everything people quote about the Twitter algorithm changed. The ranking model, the in-network source, the weight names and values, the paid-account multipliers and the group counters from 2023 are gone or replaced. What’s new in 2026 is a diversity reranker, the new-author boost, an election filter and a tool that shows you counts of your own labels.

The release history explains why the old Twitter algorithm still dominates search results. Twitter switched on a ranked timeline in 2016 with a setting called “Show me the best Tweets first.” It open-sourced its recommendation code on March 31, 2023, then mostly left that repository alone. Its last commit, “update for-you recommendations code,” is dated September 3, 2025. The new repository went up on January 20, 2026, under the account xai-org.

As Reuters reported, Elon Musk said on January 10, 2026, that the release “will be repeated every 4 weeks, with comprehensive developer notes, to help you understand what changed.” The commit history shows a 115-day gap from January 20 to May 15 (two commits that day), then a 90-day gap to August 13. After that the pace flipped: by October 24, 53 of the repository’s 56 commits had come on or after August 13, often daily. On July 15, 2026, Musk went further. As The New Stack reported, Musk said X would make “the entire codebase of X open source, with no exceptions,” with third-party reviewers to confirm the open code is what runs. I found no announcement of that review yet.

When I read it on October 17, X’s own Help Center page on For You recommendations still opened with Twitter’s 2023 wording about “roughly 500 million posts made daily.” Its “𝕏 algorithm: Learn more” link points to github.com/twitter/the-algorithm, the 2023 repository, not the current one at github.com/xai-org/x-algorithm. If X’s own help page routes readers to the old code, it’s not surprising that AI answers and marketing guides describe it too.

A guide that quotes a 4x Premium multiplier is describing the 2023 code that X’s own Help Center still links to.

Why some posts never appear: visibility filtering and Under the Hood

A high X algorithm score doesn’t guarantee a post shows up. After ranking, a separate system decides whether each post is shown normally, shown behind a warning screen or dropped, based on labels that other systems attach to posts and accounts. X’s README calls this “Ranking and Visibility Are Separate,” and it answers most “shadowban” questions: the limits are real, they come from labels, and they apply after ranking.

The labels come from several places in the code: rule engines, models that score accounts from how people react to them, image and video classifiers and an enforcement service. Some rules for posts from accounts you don’t follow can only drop, never allow; the README gives spam “caught at high recall” as the example. So a post can reach its author’s followers and still be kept out of strangers’ feeds.

X algorithm illustration: at a baggage sorting junction, suitcases with colored tags turn onto a short loop while untagged ones ride the long belt out

Since August 13, 2026, X has offered a pilot tool called Under the Hood that shows an account aggregate statistics about the visibility labels on its posts. On September 18, X added legal limits to the report, such as whether any of your posts were withheld in a country after a legal demand, and which country. Users have pointed out its limits in the repository’s issue tracker. One open issue from August 31 notes a report showing 16 posts labeled for adult content. It complains: “There is no list of which 16 posts were labeled, so the account owner cannot review or appeal the specific posts.” Another, from October 12, says the report was still showing August data in October. As of October 17, neither issue has a reply from X.

Law shows up in the feed code too. home-mixer/filters/brazil_2026_election_filter.rs removes posts from accounts reported to Brazil’s Superior Electoral Court under its Resolution 23.610, unless the viewer follows the account. On October 13, 2026, following the court’s instruction, X cut the list to the accounts of candidates in the October 25 runoff, so their posts reach For You only for people who follow them. The filter as published doesn’t check which country the viewer is in. The file now holds 25 account IDs. It’s a rare case where a reader can see exactly how a legal order changes what a recommender shows.

How to reset your X algorithm

No single button resets the X algorithm, and X’s Help Center documents none. You can still change your For You feed a lot, because the model’s main input is your recent behavior and several of X’s controls work as hard filters. These steps use X’s own settings, in the order I’d run them.

  1. Mute the words and accounts you never want to see. Muted words and muted or blocked accounts are removed before scoring, so they work immediately and don’t depend on the model learning anything, apart from the timeout caveat above.
  2. Unfollow accounts that pull your feed where you don’t want it. Who you follow decides what Thunder retrieves and shapes what the model predicts for you.
  3. On unwanted recommendations, use the post menu and choose “Not interested in this post” or “Not interested in this Topic.” X says it uses these “as a signal to recommend less of that type of content,” and param.rs weighs “not interested” at −52.272.
  4. Stop engaging with what you want less of, including quick replies to content that annoys you. The model reads your recent engagement sequence; arguing with a post teaches it you’ll reply to more like it.
  5. For a few days, like, reply to and follow the accounts you do want, so your recent history points somewhere new.
  6. When you just want to read, switch the Home tab to Following, which X says shows only accounts you follow.

What doesn’t help: clearing the app cache, reinstalling the app or logging out and back in. None of those touch the history the model reads, which lives on X’s servers, and I found nothing from X describing any of them as a reset. The full list of controls is in X’s recommendations help page.

Can you completely reset the X algorithm?

Not in one step. The recommendation history is built from your engagements, and X’s Help Center describes no way to wipe it. The closest you get is a combination: mute hard, unfollow, engage deliberately for a few days, and use Following when you want only accounts you follow. Instagram is the counterexample; it now has a reset in settings, which our guide to Instagram’s algorithm reset walks through.

X algorithm illustration: a tiler replaces dark mosaic tiles with bright ones, one section at a time, across a half-finished floor

If you post on X: what the code rewards

For a brand or creator account, the X algorithm rewards posts that people share by link or DM, quote, reply to and follow you from. It also rewards conversations with people who follow you back. It doesn’t reward a format, a hashtag count or a posting time. And because posts age out of For You at 48 hours, a team’s X calendar is really a calendar of two-day windows.

  • Write for the share. A copied link is the heaviest positive weight at 20.0, and a DM share and a quote each weigh 5.0, the same as a reply. A plain share, which the code doesn’t define further, weighs 2.0. X’s comment says weights also reflect how rare an action is, so a large weight means a small rise in that chance moves the score.
  • Build mutual relationships. A predicted reply from a mutual follow is worth 20.0, four times the 5.0 for anyone else.
  • Space out posts by the same account. When two or more of your posts compete in the same feed refresh, the second keeps 62.5% of its score, the third about 44%, and the floor is 25%.
  • Avoid what triggers “not interested,” mutes and blocks. Per unit of predicted probability, each of those costs more than the biggest positive weight, 20.0, can add. Blocks and reports also feed the models that label accounts.
  • If the account is small, treat the first two hours after an original post as its best chance to reach strangers, through the new-author boost.

Posting time still matters for who is awake to see a post in its first two days, but that’s a scheduling question, not an algorithm rule. Our comparison of the best social media schedulers covers the tools. If Buffer is what your team uses now, the Buffer alternatives breakdown shows which ones post to X.

The money side changed in 2026 too. X’s Help Center, as I read it on October 17, still says Creator Revenue Sharing “is being retired; new enrollments ended August 7, 2026, and it closes September 7, 2026.” Its successor, Original Content Rewards, pays on “qualified impressions”: “unique impressions from Premium users” on the Home Timeline feed, “where at least 50% of the post is visible.” For a creator, that makes Premium readers in the For You feed the paying audience. That’s one more reason not to build income on a single platform, and our comparison of beehiiv vs Substack covers owning the list instead.

What research says the X feed amplifies

Two peer-reviewed studies published in 2026 measured what X’s ranked feed does to its readers. One found it shifted political opinion, and the other found it amplified values at odds with users’ own. Both measured real feeds rather than reading code, and the Nature experiment ran in 2023, so it measured the feed that came before today’s code. I’m reporting what each paper’s abstract says and leaving the politics there.

In Nature, published online on February 18, 2026, Germain Gauthier, Roland Hodler, Philine Widmer and Ekaterina Zhuravskaya randomly assigned active US users to an algorithmic or a chronological X feed for seven weeks. Switching from chronological to algorithmic “shifted political opinion towards more conservative positions,” while switching the other way “had no comparable effects.” They also found the algorithm promotes conservative content and demotes posts by traditional media.

In PNAS, published August 17, 2026, Ziv Epstein, Michael Bernstein and colleagues annotated the values in the feeds of 715 US users. Posts from followed accounts matched users’ stated values, but there was “an overall negative correlation (misalignment)” between those values and the ones the algorithm amplified, for Democrats and Republicans alike, and the gap was larger for Democrats. They point to “replying, a heavily weighted form of engagement” as part of the reason. X’s former head of product Nikita Bier posted on the day the paper drew coverage. Bier wrote that “The largest contributor of seeing ragebait was the reply predictor and we were aware that angry replies were causing people to see more of that content,” and said the mutual-follow boost had since reduced it.

Back in 2023, Arvind Narayanan, writing for the Knight First Amendment Institute, called the released Twitter algorithm “a standard engagement prediction algorithm of the kind most major platforms use.” The 2026 code is more modern, but that description still fits. It’s the frame I’d keep: X predicts what you’ll engage with, and engagement isn’t always what you’d choose.

How the X algorithm compares with other platform feeds

Of the big platforms we’ve covered, only X lets you read the code that ranks its main feed. That changes what “confirmed” can mean. Everywhere else, the evidence is company statements and creator testing. The mechanics are similar enough that comparing them helps a team planning across networks. Our guide to social media algorithms puts all of them side by side.

TikTok also leans on predicted watch behavior and largely ignores follower count, as our breakdown of TikTok’s For You algorithm shows. Facebook now fills a large share of its feed from accounts you don’t follow, covered in how Facebook’s feed ranks posts. YouTube separates Shorts from long-form, in how YouTube ranks videos, and Instagram runs different rankings per surface, in Instagram’s algorithm explained.

The X algorithm differs in that you can check and date the claims about it. That cuts both ways. It’s the most transparent feed of the five and the one where advice goes stale fastest, because the weights file was updated 46 times between August 13 and October 24.

Field notes

  • Code version checked: github.com/xai-org/x-algorithm at commit 862a123 (October 17, 2026); weights rechecked October 24, 2026, unchanged after the October 23 sync.
  • Positive weights: copy link 20.0; reply, quote and DM share 5.0; follow author 4.0; share 2.0; repost 1.0; like 0.66; reply from a mutual follow 20.0.
  • Negative weights: report −234.0; mute −58.8; not interested −52.272; block −31.2.
  • Hard limits: no For You post older than 48 hours; posts from strangers ×0.75; repeated authors ×0.625 for the second post, decaying toward ×0.25.
  • New-author boost: original posts, at most 50,000 followers, under 2 hours old, under 200 Home views, one per refresh lifted to about 16th place.
  • Release pace: 56 commits from January 20 to October 24, 2026; gaps of 115 and 90 days before August 13, then 53 commits.
  • Sources: X’s x-algorithm repository (README, param.rs, vm-ranker/params.rs, scoring.rs, author_cold_start.rs, BIDIRECTIONAL_BOOST_CHANGE.md); Twitter’s 2023 the-algorithm and the-algorithm-ml repositories; X Help Center; European Commission; Nature (2026); PNAS (2026).

Frequently Asked Questions

How does the X algorithm work in 2026?

It gathers recent posts from accounts you follow and accounts you don’t, drops posts you’ve blocked, muted or already seen, then predicts how likely you are to like, reply, share or report each one. Those predictions are added into a score using weights X publishes and updates often.

Is the X algorithm actually open source?

Mostly. X publishes the ranking code, weights and training code under Apache 2.0, but not the trained model, its Grox prompts or some moderation rules. Its README says production defaults are copied into the repository, which you can read but not use to reproduce your own feed.

How do I reset my X algorithm?

There’s no reset button. Mute unwanted words and accounts, and unfollow what you don’t want. Mark recommendations “Not interested in this post,” stop engaging with content you want less of, and engage more with what you want for a few days. Use Following when you want only accounts you follow.

Does X Premium boost your reach on X?

Not through any published weight. The 2023 code multiplied Blue-verified authors by 4.0 or 2.0, but the 2026 weights and score arithmetic don’t mention Premium. Premium does matter for creator pay, which now counts impressions from Premium users.

Not by rule. The code gives opening a link a small positive weight and has no link penalty. I traced the popular “50 to 90 percent” figure and found no experiment behind it. Link posts can still underperform if readers leave X and don’t come back to reply or share.

Is the Following tab on X chronological?

X’s Help Center says Following shows only accounts you follow “in reverse chronological order.” The page has no date, and the code labels that feed as ranked. Check the Following tab on your own account before relying on it.

Does one report cancel out hundreds of likes on X?

No. The report weight multiplies the predicted chance that you’d report a post, not a count of reports. X’s README calls the “468 likes” reading incorrect, and its code notes a report is more than 1,000 times rarer than a like.

Does Grok rank the X For You feed?

Not the chatbot. The ranker, Phoenix, is a transformer whose earlier releases were ported from Grok-1, and the European Commission calls it Grok-based. A separate system called Grox uses language models to label content, and its prompts aren’t published.

How old can a post be to appear in For You?

Up to 48 hours. The X algorithm’s AgeFilter drops older posts before scoring. You can still find a two-day-old post on a profile, in search or in Following, but For You won’t recommend it.

My Take

If I were briefing a social team on X today, I’d throw out every playbook that quotes 1,500 candidates, a 48-million-parameter model, a Premium multiplier or a link penalty. Those describe a feed X replaced. I’d keep three things: write posts people want to share by link or DM, build replies with accounts that follow you back, and plan in two-day windows. That’s how long a post can live in For You. I wouldn’t approve Premium seats on the promise of reach. The published code doesn’t support that promise, though Premium does matter if anyone on the team earns creator payouts.

I’d also cut the time spent on hacks. The code makes most of them irrelevant. The weights file was updated 46 times between August 13 and October 24, so a rule that was right in August can be wrong by October. The useful habit is the one this article is built on: when someone says the X algorithm rewards something, ask which file and which date. If they can’t answer, it’s probably 2023.

- Advertisement -
Share This Article
Marketing Strategist & Tech Journalist
Dana Pritchard covers marketing software, social platforms, and the tools small teams and households pay for every month. She looks hardest at what a plan costs over a year and what the pricing page leaves out.
Leave a Comment