How the X Algorithm Works in 2026

“X algorithm” is one of the least precise phrases in this job. People use it for the For You feed, for a reach slump, for a rumor they saw in a reply, and for a feeling that the app is personally annoyed with them.

I have advised 500+ accounts over 10+ years. I still will not invent a score X has not published. This page sticks to what X has put in public, plus a clear label on everything that is only a working interpretation. If you want to turn the published weights into a relative illustration, use the X algorithm simulator. The simulator is not the live ranker. It says so on the page, and it should.

Two feeds, not one machine

X shows a For You feed and a Following feed.

For You is the recommendation feed. It can show posts from people you do not follow. That is the system people mean when they say the algorithm. X has described it, across public engineering posts and later open-source code, as a ranker: it tries to predict how likely you are to take certain actions on a post, then orders what you see.

Following is the feed of accounts you chose. I am not going to claim it is a pure clock. X has not published a complete, unchanging description of every Following ranking tweak. If a post from someone you follow is missing, that is not automatic proof of a secret penalty. It can be recency, muting, the author’s settings, or a client bug. I do not have a public spec that lets me rank those causes.

When someone asks for the Twitter algorithm explained, I start with that split. Advice that mixes the two feeds usually sounds confident and is hard to test.

What “published” means here

There have been at least two public eras.

In 2023, Twitter released an earlier recommendation codebase and an engineering write-up. People still quote figures from that era, often from memory, often wrong. I am not repeating those figures on this page. Mixing 2023 blog-post math with 2026 code is how bad charts get made.

The weights below are from a later public repository, github.com/xai-org/x-algorithm, in home-mixer/params/param.rs, checked on 26 September 2026. The same note is on the simulator. If you are reading this later, open the file. Do not trust a screenshot of this article as the live configuration.

Two limits matter before any number:

The public file is a snapshot. The ranker X is running today can be ahead of it, behind it, or partly different. An open-source drop is not a contract that production will match the file forever.

The weights multiply the predicted probability of an action. They do not multiply the raw number of likes or replies sitting on a post. A reply weight of 5.0 does not mean “one reply is worth ten likes” in your analytics. It means the model’s estimated chance of a reply is scaled by 5.0 inside that scoring function, while the estimated chance of a like is scaled by 0.5. Those are different objects. Treating them as a conversion table is the most common misread I see.

Weights in the public For You file

These are the figures this site already cites from that file. I am not adding new ones.

Positive signals, as published:

• A reply is 5.0.
• A like is 0.5. Replies are weighted far more heavily than likes.
• A repost is 1.0.
• A quote is 5.0.
• A follow is 4.0.
• A copy-link share is 20.0.
• A reply from someone you follow back adds 15.0 to the reply weight. That is an addition inside the reply term, not a separate promise about your follower count.
• DM shares are 5.0 in the same public notes.
• Dwell is 0.05. A video open is 0.07. A photo expand is 0.05. These are small next to a reply.
• Opening a link has a small positive weight, 0.2, on the predicted probability of that action.
• Profile clicks are weighted 0.0 in that file.

Negative signals, as published, again on predicted probability rather than raw counts:

• A report is -234.
• Mute author is -58.8.
• Not interested is -43.2.
• Block author is -31.2.

The size of the negative weights is the part operators should respect. A handful of strong negative actions can dominate a pile of light positive ones inside this scoring function. That does not mean I can look at your account and count how many mutes you received. X does not show you that. It means the published math punishes predicted negative actions very hard, so posting in a way that attracts blocks, mutes, and reports is a bad trade even when the like count looks fine.

Source, again: github.com/xai-org/x-algorithm, home-mixer/params/param.rs, checked 26 Sep 2026.

What the file does not say

This is the section that should slow you down.

There is no bookmark weight in the file this site checked. Bookmarks still matter to you as a creator, because a save means someone wanted the post later. They are in the engagement rate calculator for that reason. They are not, in this public snapshot, a listed For You weight. Do not claim X “counts bookmarks as 50x” or any other multiple. That number is not in the file.

There is no Premium multiplier in the file. Subscribers get product features, including longer posts, that X has marketed separately. I will not attach a ranking percentage to Premium. The simulator leaves Premium qualitative for this reason. If a thread tells you Premium is “a 4x boost” or any other factor, ask for the file and the line. A testimonial is not a weight.

There is no link-penalty percentage. Posts with external links often get less reach in practice, so I still tell clients to test the same idea with the link in the first reply or in the bio. That is a test, not a published bypass. The public code does not say “links lose 70 percent” or any other cut. The link-open weight above is small and positive. A positive weight on opening a link is not the same thing as a penalty for including one. Both can be true in a live system that has other, unpublished filters. I do not know those filters, so I will not describe them.

Author replies under your own post are worth doing. X does not publish a special multiplier for them in the notes this site uses. I still reply, because the conversation is the point and because replies in general are a heavy weight. I do not tell clients that answering their own thread “doubles reach.” I have not seen that claim in the file.

If you want to poke the published weights without pretending you are inside X’s production model, the algorithm simulator applies the reply, like, repost, quote, and media weights to likelihoods you choose, then scales a relative total. Link placement and Premium do not move that number. That limitation is the honest part.

An X algorithm update is not a mood

X algorithm update is what people type after a bad week. Sometimes they are right that something shipped. Sometimes they changed topics, audience, or posting habits in the same week and blamed the feed.

Here is a careful way to talk about updates:

X can change ranking, filtering, and product rules. Some of those changes are discussed in public posts by X or by the people who run it. Some land in the open-source repository. Some may exist only in production. I cannot see production. You cannot see production. A public commit is evidence of what was published, not a complete diary of the live system.

When I review an account after a slump, I look for ordinary explanations before I reach for “update”:

• The posts stopped earning replies and started earning only likes, or neither.
• The topic drifted away from the people who used to respond.
• External links moved into the main post.
• The account picked up behavior that attracts mutes and blocks: repetitive pitches, bait, or obvious automation.
• The comparison window is a viral outlier, so “normal” looks like a crash.

None of those require a secret patch. They are also fixable without a rumor.

If you track this for a company, write down the date you changed the content, the date you changed the profile, and the date a public ranking note actually appeared. Otherwise every dip becomes an update story, and you learn nothing.

Is there an X algorithm reset?

X algorithm reset is a phrase creators use. It usually means one of two ideas. They are not the same, and only one of them has a public footing.

The first idea is that X replaced or heavily revised the ranker. That kind of change happens. The 2023 release and the later xAI repository are already two different public artifacts. Calling a real ranking overhaul an “update” is fair. Calling it a personal reset of your account is not, unless X says the overhaul wiped per-account state. I have not seen that stated in the file this page cites.

The second idea is a folk remedy: your distribution was zeroed, and if you behave a certain way for a set number of days it comes back. I will not publish a day count. X has not, in the material this site relies on, published a reset button, a cooldown length, or a promise that reach returns on a schedule. Anyone selling a “reset protocol” with a timer is ahead of the evidence.

What I do when reach falls, instead of hunting a reset:

• Read the last twenty posts as a stranger. If you would not follow this account, the ranker is not the first problem.
• Check whether replies are still happening in other people’s threads. If your own posts are quiet and your replies are also quiet, look at who you are replying to. The method is in how to find the right people.
• Compare a post with a link in the body to a post with the link moved. One test. Not a law.
• Stop automation if you are using it. Bots and pods are a good way to collect the negative actions the published file weights so heavily.
• Use the simulator to see which actions the public file cares about, then go earn those actions from the right people. A high predicted like on the wrong audience is still the wrong audience.

A slump can last longer than a week for reasons that have nothing to do with a hidden timer. Relevance is slow. That is uncomfortable, and it is still the working explanation I trust more than a reset myth.

What I would not do with this information

I would not build a spreadsheet that converts ten likes into one reply and then calls the result “algorithm score.” The math is not that.

I would not delete a month of posts because a thread said the account needs a clean slate. Deleting can remove the proof a new visitor needed. Do it when a post is wrong or off-strategy, not as a ritual.

I would not buy engagement to “train” the model. You would be training it on a crowd that does not buy, and you would be courting the negative weights if the pattern looks automated.

I would not quote a weight from this article without the date. The file can change. The sentence to use is: as of the 26 Sep 2026 check of param.rs, a reply was 5.0 and a like was 0.5, on predicted probability.

What to do on the account anyway

The practical reading of the public file is almost old-fashioned.

Write posts people can answer. A reply is a heavy positive weight. A like is a light one. Ask for the reply only when you actually want the answer. Empty “thoughts?” lines do not create a real conversation, and they annoy the readers who still give you the benefit of the doubt.

Stay in a topic long enough that the same people recognize you. The ranker can only predict interest in what you keep demonstrating.

Be careful with outrage and with bait that earns reports, mutes, and blocks. The published negatives are enormous next to a like. Even if production differs, I have never seen a commercial account win by collecting blocks from the people it hoped to sell to.

Put the profile and the offer in order so a visit from For You can become a follow. Distribution without a clear profile is a tour that ends at the door. That work is covered in how to grow on X in 2026 and in the X Growth Playbook.

If you want a human to look at the account with you, organic X growth and X consulting are the relevant offers. Neither one includes a private ranking lever. I do not have one. The simulator is the tool. This article is the caution label.