Test your tweet against the Twitter algorithm. The published code, not rumors.

What the X algorithm actually is in 2026

Twitter has published its ranking code twice. In March 2023 it released twitter/the-algorithm, with the heavy ranker's engagement weights listed in a companion repository. In January 2026 it published a new codebase for the For You feed, xai-org/x-algorithm, whose weights file was last synced on October 20, 2026. In the 2026 code, posts from accounts you follow and posts found by retrieval are ranked together by a transformer model called Phoenix. It predicts how likely you are to take each action on a post, and the final score is a weighted sum of those predictions: positive actions add, negative ones subtract.

The weights the code publishes

The 2026 repository lists default weights in its params file. These are the ones that matter for a single post:

• Reply 5.0, like 0.66. A predicted reply counts for over seven times a predicted like, and the file sets a separate boost (15.0) for replies between accounts that follow each other.

• Share via copy link 20.0, share via DM 5.0, quote 5.0, repost 1.0. Posts people pass on carry the largest positive weight in the file.

• Follow author 4.0. A post that earns a follow outweighs one that earns a repost.

• Report -234.0, mute -58.8, not interested -52.272, block -31.2. Negative signals weigh far more than likes.

Two caveats come from the code itself. The file says its values mirror configuration defaults, so live values can differ. And the README warns that weights scale predicted probabilities, not raw counts, so “one report cancels 468 likes” is the wrong reading. For comparison, the 2023 heavy ranker weights (dated April 5, 2023) put a reply at 13.5 against a like at 0.5, and a reply the author then engaged with at 75.0.

The seven checks, and the code behind each

• Reply pull, 20 points. A specific question or an arguable position. Basis: the reply weight.

• First line, 20 points. The 2026 model predicts dwell, and “not dwelled” carries a negative weight. Your first line decides whether anyone stops.

• Negative-signal risk, 20 points. Insults aimed at people, shouting in capitals, spammy asks. Basis: the report, mute, block and not-interested weights.

• Share value, 15 points. A number, a named tool, a step someone would send to a friend. Basis: the share weights.

• Follow reason, 10 points. First person plus a real detail. Basis: the follow-author weight.

• Hashtags, 10 points. Twitter's help center recommends no more than 2 per post, and lists hashtag abuse among the reasons accounts get filtered from search.

• Length, 5 points. Up to 280 characters, counted Twitter's way: every link is 23 characters and every emoji is 2.

Links are not scored. The 2026 params file gives links no negative weight, and opening one is a positive action (0.2). We would rather leave a check out than invent a penalty.

What no draft checker can see

The rest of the ranking happens outside your text. The code drops posts older than 48 hours, removes duplicates and posts containing a viewer's muted keywords, discounts repeat posts from the same author, and multiplies posts from accounts a viewer does not follow by a factor below 1. Timing, audience and account standing matter as much as the draft. For those, see when to post, your TweepCred estimate, and the viral tweet analyzer for what strong posts do.

The code rewards replies. Be worth replying to.

Read the weights again and the pattern is plain: the algorithm pays for conversations and for posts worth passing on, and punishes posts that make people mute, block or report. That is a content problem, not a hack. VoiceMoat works on it from both ends. audenAI, the brain inside VoiceMoat, learns from your own content and drafts posts in your voice, and Engage finds the conversations worth joining and drafts replies that sound like you. Replies are one of the heaviest positive weights in the 2026 file, so they are worth doing well.