How the X (Twitter) algorithm works in 2026

The X algorithm ranks your For You feed by predicting, for every candidate post, how likely you are to reply, dwell, repost, or like it, then ordering posts by those odds and subtracting heavily for "not interested," mute, block, and report. Roughly half the candidates come from accounts you follow and half from accounts you don't.

The For You feed is a sorting machine with four moves. First it gathers a pool of candidate posts, around 1,500 of them, split roughly half from accounts you follow and half from accounts you don't. Then a model predicts, for each candidate, how likely you are to take each action: reply, like, repost, watch the video, click the profile, stay and read, or tap "not interested." Then it scores every candidate by combining those odds with a set of weights. Then it filters the result (drops muted authors, stops one account from flooding you, removes duplicates) and serves what is left in ranked order.

Two facts do most of the work. Replies and dwell time are weighted far above likes, so the content that travels is the content that pulls people into a conversation or holds their attention. And negative signals are brutal: a single report historically counted for hundreds of times a like, with a minus sign. Everything else is detail on top of those two.

Before any of the ranking matters, there is one lever you fully control, and most guides bury it. X has two timelines. For You is the algorithmic feed, the sorting machine above. Following is reverse-chronological and shows only the accounts you follow, newest first, with no ranking and no out-of-network recommendations.

If you want the old Twitter back, the Following tab is it. Nothing gets injected, nothing gets re-ordered. The catch is that X keeps defaulting you to For You, so you have to pick Following each session (or pin it). Everything else in this article is about the For You feed, because that is the one doing the deciding.

The 2026 codebase is organized around a coordinator called Home Mixer, which runs the candidates through a pipeline. The stages, in order:

1. Candidate sourcing. Posts from people you follow come out of Thunder, an in-memory store that keeps recent in-network posts ready to fetch in under a millisecond. Posts from people you don't follow come from Phoenix, a two-tower retrieval model that turns you and every candidate post into vectors and pulls the closest matches. The design target is a roughly even blend of the two.
2. Ranking. The same Phoenix system, now in its transformer-ranker role, predicts the probability of each engagement action for every candidate. It scores each post in isolation (a post's score does not depend on the others around it), which is what lets X cache and parallelize the work at the scale it runs. Musk has claimed the model reads "every post" and watches "every video," more than 100 million a day.
3. Scoring. Those predicted probabilities get combined into one number with a weighted sum. More on the weights in the next section.
4. Filtering. A few heuristics survive the rewrite: an age filter, an author-diversity step that stops one account from monopolizing your feed, removal of muted and blocked authors, and de-duplication. Content you can't see (paywalled subscriber posts you don't subscribe to) is dropped here too.
5. Serving. Home Mixer assembles the survivors in ranked order and hands them to your app.

If you read the older write-ups, this is where the famous names lived: SimClusters, RealGraph, TweepCred, the heavy ranker and the light ranker. Those belong to the 2023 system. The 2026 release folds most of that machinery into the two-tower retrieval and the single transformer. The job is the same; the parts have different names.

Here is the part everyone wants and almost nobody sources correctly. The score for a post is a weighted sum: each predicted action probability multiplied by a weight, all added together. The 2023 release published the actual weights. They are the only hard numbers we have ever gotten from X, so they are worth reading carefully, with one caveat we will come back to: these are the 2023 values, and the 2026 system hides its own.

Source: X's 2023 open-source heavy-ranker README (the-algorithm-ml, the recap model). The weights apply to predicted probabilities, not to actions you have already taken, so read "reply 13.5" as "a reply is valued about 27 times a like," not "a reply equals 27 likes."

The shape of the table is the whole lesson. Replies are worth far more than likes, and a reply the author engages back with is worth more still. Holding someone's attention (the dwell-time and profile-click rows) beats a fly-by tap. Reposts are roughly two likes, not the twenty you have read elsewhere. And the negative rows are huge: "not interested" roughly cancels the single biggest positive signal, and a report outweighs anything positive several times over, which is why one angry tap can sink a post that was doing fine. When that downranking turns sustained and account-wide instead of a single unlucky post, people call it a shadowban, and we cover how to tell a real one from a slow week and how to fix it. If you want the tactical version of all this, what to actually post and how to use the reply weighting in practice, we wrote that up separately in how to get more engagement on X. This piece is about the machine; that one is about the moves.

The reason this article exists is that the 2023 description, the one still sitting at the top of most search results, is now describing a retired system. On January 20, 2026, xAI published a new repository, xai-org/x-algorithm, under an Apache 2.0 license, with a public pledge to update it every four weeks. Reporting from TechCrunch and Decrypt covered it the same day, framing it against an EU transparency fine of about 140 million euros that landed days earlier. The repo has been maintained since, with a large update on May 15, 2026.

The architecture is genuinely different. The old design was two ranking models stacked (a cheap one to shortlist, an expensive one to score). The new one is a single transformer, the same family as xAI's Grok model, adapted for recommendations. Elon Musk had previewed the move in October 2025, writing that X was "aiming for deletion of all heuristics within 4 to 6 weeks" and that Grok would "literally read every post and watch every video." The shipped code is more honest than the preview: its own README says it eliminated "every single hand-engineered feature and most heuristics." Most, not all. A handful of filters and the author-diversity rule survive.

The stated goal behind it is worth quoting, because it explains a lot of the recent behavior. Musk has described the objective as maximizing "unregretted user-seconds," and added that "too much negativity is being pushed that technically grows user time, but not unregretted user time." A content-understanding service in the new code, called Grox, reads post text, images, and video to classify topic and catch spam, which is the machinery that lets the system judge a post on what it says, not just on who is engaging with it. Here is the before and after:

What did not change is the spine. It is still a weighted sum of predicted engagement probabilities, replies are still weighted above likes, negative feedback is still heavily punished, and the feed is still a blend of in-network and out-of-network posts. The model that produces the predictions got smarter; the logic for what to do with them is recognizably the same.

"The algorithm is open source" is doing a lot of quiet lying in most coverage. You can read the code. You cannot read the part that decides your reach. The 2026 scorer references a set of named weights (a reply weight, a like weight, a report weight, and so on) exactly like the 2023 one, but the file holding the actual numbers is not in the public repository. The structure is open; the values are redacted.

The omissions go further. The published model is a small demo version, not the production one. The release is a frozen snapshot, while the real system keeps training on live data. The training data itself is not included, for obvious privacy reasons. The 2023 release drew the same criticism, and the README admitted then that the weights "can be adjusted at any time" and "have periodically been adjusted." A researcher at the Knight First Amendment Institute put the limit plainly after the 2023 drop: the scoring formula "isn't actually in the code," it is tuned separately and frequently, so reading the repo tells you the design, not the live math.

There is a smaller lesson in the 2023 history too. The first release contained a tag that grouped users by whether they were, among other labels, Elon Musk. It looked sinister and got pulled within about two hours, because it was A/B-test bucketing for measuring metrics, not a reach booster. The honest read on both releases is the same: transparency theater is a real risk, the gesture is bigger than the disclosure, and you should treat any specific number you read about the current weights as an educated guess.

Most of the durable advice about the X algorithm is half-true at best. The ones worth correcting:

• "A repost is worth 20x a like." The published weight is 1.0 for a repost and 0.5 for a like, so a repost is about twice a like. The 20x number is a misreading that refuses to die. A reply is the action that carries real weight.
• "Hashtags boost reach." The model reads your text to work out the topic; it does not need a hashtag to do it. One or two are harmless. Stacking them can trip the spam classifier, and Musk has flat-out asked people to stop using them.
• "Posting more often means more reach." The author-diversity step caps how much of one account shows up in a given feed, so your tenth post of the day mostly competes with your own other posts. Volume past a point buys nothing.
• "Engagement pods beat the algorithm." X's authenticity policy prohibits "coordinating to exchange engagement in any X features, such as Likes, Polls, Replies, Reposts, Lists, Views, or Follows." Pods are a policy violation that can get the posts pulled from recommendations, not a growth hack.
• "The first 30 minutes decide everything." Early engagement does feed out-of-network expansion, and the legacy code used a relevance half-life of about six hours, so velocity matters. But "everything is decided in 30 minutes" is an overstatement, and the often-quoted 18-minute half-life has no source behind it.
• "A fresh account, or editing and deleting, resets the algorithm." There is no instant reset. The system is a slow-moving model of your behavior, and mass follow or unfollow is itself against the rules. No evidence supports the claim that editing or deleting a recent post penalizes reach.

This is the most-argued question in the space, and the honest answer is more careful than either side wants. There is no explicit rule that throttles links. Musk said as much in April 2025: "there is no explicit rule limiting the reach of links." What is true is that the system optimizes for time spent on X, and a link that sends someone off-platform works against that goal, so link posts tend to get less reach as a side effect, not as a punishment.

X has been actively softening even that. In October 2025 its head of product said link posts would stop pushing the link to the top and instead collapse it to the bottom, so people can read and react before they click away, paired with an in-app browser test. The practical takeaway: the popular "put the link in the first reply" trick is a reasonable hedge, not a law of physics, and the made-up precise penalties you see ("links cost you 30 to 50 percent of reach") have no source. A link with a real description and an image does better than a bare link; that part holds up.

There is no "reset For You" button. X does not give you one, and the guides that promise a clean wipe are selling a trick that does not exist. What you actually have is a set of signals that nudge the model, the same signals that show up as those big negative weights in the scorer. Used deliberately, they retrain your feed over days, not minutes. The quick version is below; the full step-by-step is in how to reset the X algorithm.

• Use the negative signals on purpose. "Not interested in this post," "not interested in this topic," mute, and block all feed back as the negative weights that push similar content down. Report is the strongest, and you should reserve it for content that actually breaks the rules.
• Feed it positives. Follow and genuinely engage with the accounts and topics you want more of. The feed is built from what you do, so deliberate engagement is the most reliable retraining tool you have.
• Build Lists. A curated List is a hand-made feed the algorithm doesn't touch. It is the cleanest way to guarantee you see a specific set of accounts.
• Switch to Following when you want a break. The chronological tab sidesteps ranking entirely. If your For You feed has gone sideways, this is the reset that actually works.

That is the consumption side: tuning what reaches you. The mirror image is the production side, deciding which conversations are worth reaching into, and the two need opposite tools. Muting and "not interested" hide posts; the same patterns, read the other way, are how we spot the bots clogging a brand's replies. And where filtering your feed means hiding things, monitoring means surfacing them: a saved keyword monitor watches X and the other networks for the words you care about and drops the matches into one inbox, with bot scoring keeping the spam out of the queue. The applied playbook is in how to monitor brand mentions, and if you want to clean up your own feed first, start with muting the right words.