How the Twitter algorithm works in 2026: a plain-English guide to the code

Updated 23 October 2026. This guide now also checks the most common Twitter myths against the public ranking code, including whether your followers see every post, whether big or Premium accounts rank higher, hashtags and mass reporting. The weights and change log are as of commit 7f2ca2b, read 21 October 2026.

The Twitter algorithm (the For You feed) gathers recent posts from accounts you follow and accounts you don’t, throws out anything older than 48 hours or already seen, predicts how likely you are to like, reply, share, follow, mute or report each post, multiplies those predictions by fixed weights, and shows you the highest scores. The code is public on GitHub at xai-org/x-algorithm. This guide explains it in plain English and links to our deeper Twitter algorithm series for each part.

Everything below comes from that repository as of 21 October 2026, with file references in our source notes. Where the code does not settle a question, we say so rather than guess.

How the Twitter algorithm works, step by step

• Step 1: Read your recent activity. The system loads up to 1,024 of your most recent actions (likes, replies, time spent on posts and so on), plus who you follow, who you have muted or blocked, mutual follows and some profile details such as language and country.

• Step 2: Collect candidate posts. Several sources run in parallel. Thunder brings in posts from accounts you follow (up to 1,200, newest first). Phoenix retrieval finds posts from accounts you don’t follow (up to 1,000, plus up to 200 “cold start” picks). SimClusters adds up to 800 more based on what you engaged with over the last 90 days, and a popular-posts list adds up to 500 recent originals from widely followed accounts. More on this in how out-of-network retrieval finds posts from people you don’t follow.

• Step 3: Filter. Up to 21 filters run in order, some switched off by default. The ones that matter most: posts older than 48 hours are dropped, along with duplicates, your own posts, posts you have already seen or been served, accounts you block or mute, your muted keywords, subscriber-only posts you can’t see, and replies or reposts from accounts you don’t follow. The 48-hour rule is explained in the 48-hour window and Thunder.

• Step 4: Predict and score. The Phoenix ranking model predicts the chance you will take each of a couple of dozen actions on each post. Those predictions are multiplied by the weights in the table below and added up into one score.

• Step 5: Adjust. Three adjustments follow: a boost that can lift one small-account post per feed load, a discount on later posts from the same author, and a discount on posts from outside your network. A separate diversity step can also drop near-duplicate posts from the top of the list.

• Step 6: Select and check. The top 50 posts are kept, then visibility filtering removes any that a safety or legal rule says to drop, and only the best post from each conversation is kept. The feed is cut to 35 posts.

• Step 7: Blend. Ads, a Who to Follow module and prompts are added. Ads take slots inside those 35 posts; modules are added on top, up to 48 items in one response. See what competes with your post for feed slots.

Twitter algorithm weights in 2026

Here are the weights in the production parameter file. The most important thing to know about them: they multiply your predicted chance of doing something, not counts of what people did. A report has a huge weight because the model’s predicted chance of a report is tiny. So “one report cancels hundreds of likes” is wrong; the code comments say so directly.

Positive weights in the Twitter For You scoring code (production defaults, read 21 October 2026).

Negative weights in the Twitter For You scoring code (production defaults, read 21 October 2026).

After the sum, a small offset makes sure any post the model expects you to like more than dislike always ranks above any post it expects you to dislike. Bookmarks are predicted and logged, but they have no weight in the formula. The full list of 25 documented actions (29 weighted terms in code) is in Phoenix’s engagement heads: a creator’s guide, and the negative side, including visibility labels, is in negative signals: mute, block and report.

If you read older guides, including an earlier version of this page, you will see weights from Twitter’s 2023 open-source release. Those came from a different codebase and do not match today’s numbers, so we have removed them.

Parameters people search for

Five parameters from xai-org/x-algorithm, explained.

What changed in the Twitter algorithm in 2026

• 19 January: the xai-org/x-algorithm repository was created.

• 10 to 24 July: a reply boost for original posts from mutual follows was tested on 10 July, launched at +20 on 13 July, and cut to +15 on 24 July after feedback (World Cup posts from accounts people did not follow were being crowded out). It is still +15, which is why the reply weight becomes 20.0 for those posts.

• 13 August: the release that added the weights, visibility filtering, the labeling systems, real Phoenix training code and SimClusters to the public repo. Any “official weights” quoted before this date did not come from this repo’s parameter file.

• 14 August: the README began stating that weights multiply predicted probabilities, not counts. A Brazil 2026 election filter was added the same day.

• 18 September: the Under the Hood report started showing legal withholding, including which country.

• 13 October: the Brazil election filter list was cut to the candidates in the 25 October runoff.

• Like weight: 0.5 in a July diff, 0.66 today. The exact change date is not in the code.

For the bigger architectural shift (a learned ranking model replacing hand-tuned rules), see the May 2026 Twitter algorithm: why voice wins when the ranker becomes a transformer.

Twitter algorithm myths vs the code

Most Twitter myths are about ranking, and the public code settles many of them. Here is each common claim, checked.

Common claims about the Twitter algorithm, checked against the 2026 code.

The Twitter myths that change what you should post

Some myths are trivia. These six change real decisions:

• “My followers see everything I post.” They don’t. Your posts compete for a 35-post feed with everything else your followers might see, scored on how each person is predicted to react. Write for the reaction of the people who do see it, and use the first two hours: for accounts with 50,000 followers or fewer, that is when an original post can get the small-account lift.

• “More followers means more reach.” Follower count is not scored; predicted reactions are. Followers who never react add nothing to the formula, which is why buying Twitter followers backfires: the count rises while the replies, shares and follows the weights reward do not, and bought accounts invite the behavior that bot detection looks for. A smaller audience that actually responds is worth more; how to get followers who stay covers building one.

• “Post as often as you can.” There is no frequency penalty, but your own posts discount each other within one feed load (x0.625 for the second, about x0.44 for the third, with code defaults). Spread posts out and put your best one where you can show up for it.

• “Premium buys reach.” There is no verified or subscription term in the scoring formula. Decide on a subscription for its features, not for a ranking multiplier the code does not contain.

• “Hashtags are how strangers find you.” Hashtags carry no weight. Out-of-network discovery comes from retrieval systems such as SimClusters, which seed on what each viewer engaged with over the last 90 days. Being the kind of post people in your niche engage with is what gets you found.

• “A wave of reports will bury me.” Partly true. Each viewer’s predicted chance of reporting carries the largest negative weight (-234.0), and a separate system turns reports per like into account labels. The fix is structural: don’t write the bait that earns reports in the first place.

Myths that are not about ranking, such as whether deleted posts truly vanish or whether people are told when you unfollow them, don’t change how the feed scores your posts, so we leave them out here.

Dwell is covered in depth in how dwell time affects ranking, and posting frequency in author diversity, DPP and how often to post.

What the code means for creators

The weights reward a specific kind of post: one people want to pass on, talk about with you, and follow you for. Here is what that means in practice, read voice-first.

• Write posts worth sending. Share via copy link (20.0) and share via DM (5.0) are among the largest positive weights. A post someone sends to a friend is the strongest signal you can earn, and it usually means the post says something specific enough to be worth passing on.

• Conversation beats applause. A reply (5.0) is worth several times a like (0.66), and a predicted follow is worth 4.0. Posts that invite a real answer, from people who know your work, do more than posts built for a quick like.

• Mutual follows are your strongest audience. On original posts from people who follow each other, the reply weight rises to 20.0. Building real two-way relationships in your niche pays off in ranking, not only in goodwill.

• The penalties are the biggest numbers. Report (-234.0), mute (-58.8) and not interested (-52.272) dwarf every positive weight. Bait, outrage and off-topic hot takes raise the predicted chance of exactly these actions among the people who see them.

• Small accounts get one shot per feed load. If you have 50,000 followers or fewer, an original post (not a reply or repost) that is under 2 hours old and has fewer than 200 home-timeline views can be lifted to about 16th place for one viewer at a time. Original posts in your first two hours matter.

• Space your posts. With the code defaults, a second post from the same author in one feed load is multiplied by 0.625, a third by about 0.44, falling toward 0.25. Five posts in an hour compete with each other.

• Replies grow you through the conversation, not the feed. Replies from accounts someone does not follow are filtered out before scoring, so your reply on a big account is seen under that account’s post, not ranked into strangers’ For You feeds. Good replies still build an audience; they just do it in the thread. The smart reply guy strategy covers how.

• For artists and visual posts: image posts are scored on the same predicted actions. Photo expand is only 0.05 and video quality views are 0, so shares, replies and follows still carry the weight.

The model also represents posts and authors with learned embeddings, which is how it finds your work for people who don’t follow you; how Phoenix represents creator identity explains what that means for staying recognizable. You can’t game any of this directly. What you control is the post itself and who you build relationships with.

Is there a Twitter algorithm tester?

No outside tool can read the score the ranking model gives your posts, because the predictions are made for each viewer at the moment they load the feed. Two things you can check:

• Under the Hood. Twitter’s own report at x.com/i/jf/under_the_hood shows labels on your posts and account, what each means, the effect on visibility, and since 18 September 2026, any legal withholding by country. On the web you can download it as JSON.

• Search visibility. Our free shadowban checker runs the one visibility test you can verify from outside: whether your posts still show up in search. The method is explained in the Twitter shadowban test.

“Tweepcred” calculators do not read anything from this code. The 2026 repo contains a PageRank-based account score called user-cred-v2, scaled 0 to 100, where 54 or more counts as “high PageRank”. It is used in labeling rules (high-PageRank accounts are exempt from some, such as duplicate-text rules), not as a term in the ranking formula.

Visibility labels remove, they don’t push down

Visibility checks run after ranking. They can drop a post, but “downrank” and “avoid” verdicts do not change its score anywhere in the feed code. Some labels, such as “do not amplify”, spam or malicious URL labels, hide a post from people who don’t follow you while your followers can still see it. Separate systems compute account-level rates of blocks and reports relative to likes, and a bot-detection model looks for automated behavior; the thresholds for both are not in the public repo.

Does the Twitter algorithm punish AI-generated content?

There is no term for “written by AI” in the scoring formula. What the formula does punish is the reaction generic posts tend to get: few shares or real replies, and more scroll-pasts, “Not interested” taps and mutes from people who have seen the same post a hundred times. So the honest answer is that the algorithm is indifferent to how a post was drafted and very sensitive to how readers react. Content that reads like one specific person earns the reactions the weights reward; interchangeable content earns the ones they penalize.

Can you reset the Twitter algorithm?

The ranking model reads your most recent actions, up to 1,024 of them, so your feed follows what you do now more than what you did months ago. To steer it: engage with the posts and accounts you want more of, tap “Not interested” on what you don’t, mute or block accounts (they are filtered out before scoring), and add muted keywords (also filtered out). New activity changes the inputs the model reads on every feed load.

The voice-first playbook for the 2026 algorithm

• Post things people forward. Specific, useful, quotable posts earn the copy-link and DM shares that carry the largest positive weights.

• Build mutual relationships. Reply thoughtfully in your niche and follow back the people who engage, because mutual-follow originals carry a 20.0 reply weight.

• Stay out of the penalty zone structurally. Don’t write bait. The register that earns reports and mutes also erodes the audience that came for you.

• Use the first two hours. For accounts with 50,000 followers or fewer, fresh original posts are the ones eligible for the small-account lift.

• Space your posts. One strong post beats three crowding each other in the same feed load.

• Don’t buy shortcuts. Bought followers, pods and reply bots invite exactly the signals that bot detection and the negative weights are built to catch. See how to grow on Twitter without buying followers and the case against reply bots.

On posting cadence specifically, how often to post on Twitter and LinkedIn walks through what the frequency studies say, and Twitter Blue vs X Premium covers whether a subscription is worth it on its own merits, given the scoring code has no Premium term.

Where audenAI fits

audenAI, the brain inside VoiceMoat, learns from your own posts and checks your drafts for Style and Stance, so what you publish reads like you across the cadence the algorithm rewards. It does not change, predict or game any ranking; nothing can. VoiceMoat is a Personal Brand OS for LinkedIn and Twitter that drafts, schedules, publishes and shows analytics, so you can see which posts earned replies, shares and follows. audenAI suggests. You decide.

Read the full Twitter algorithm series

• The May 2026 Twitter algorithm: why voice wins when the ranker becomes a transformer

• Phoenix engagement heads: every signal and weight, for creators

• Negative signals: mute, block, report and visibility labels

• Phoenix embeddings and creator identity

• Dwell time and ranking

• Out-of-network retrieval: how strangers find your posts

• Recency, Thunder and the 48-hour window

• Author diversity, DPP and posting frequency

• Feed slots: ads, prompts and Who to Follow