The X algorithm (still called the Twitter algorithm by most people) decides what shows up in the For You feed. Unlike most platforms, X has published how it works, twice. Most advice still quotes Twitter’s numbers from 2023, sometimes from a version that lasted five days. This guide uses the 2026 code: how the algorithm scores a post, what it rewards and punishes, what X keeps back, and how to grow an account with it. Every number has its source and date; where sources disagree, it says so.
How does the X algorithm work?
It predicts what each viewer will do with a post, weights each prediction, and adds them up: score = the sum of (weight × predicted probability) for each action. That was true of Twitter’s 2023 ranker and is true of Phoenix, the ranker xAI published in January 2026 (xai-org/x-algorithm). Phoenix predicts 25 actions. Filters and boosts then reorder the list.
These are the default weights in the 2026 code (param.rs, added 13 August 2026, read 28 October 2026), beside the 2023 table (the-algorithm-ml):
| Action | 2023 weight (5 April table) | 2026 default | 2026, compared with a like |
|---|---|---|---|
| Like | 0.5 | 0.66 | 1× |
| Repost | 1.0 | 1.0 | 1.5× |
| Reply | 13.5 | 5.0 | 7.6× |
| Reply, when viewer and author follow each other | — | +15.0 on top | about 30× in all |
| Reply the author engages with | 75.0 | not a separate action | — |
| Quote | — | 5.0 | 7.6× |
| Share by DM | — | 5.0 | 7.6× |
| Share by copying the link | — | 20.0 | 30× |
| Follow the author | — | 4.0 | 6× |
| Profile click | 12.0 | 0.0 | — |
| Open a link | — | 0.2 | — |
| Block | (in negative feedback) | −31.2 | — |
| Not interested | −74.0 | −52.272 | — |
| Mute | (in negative feedback) | −58.8 | — |
| Report | −369.0 | −234.0 | — |
Small terms for dwell time, video opens and clicking into a conversation are left out; together they show X now rewards time spent. These are defaults: the code says scheduled jobs sync them to production and experiments may differ, and the 2023 README said the weights were “periodically adjusted”. Treat every weight as true as of a commit, not a constant.
“A reply is worth 27 likes” is 13.5 ÷ 0.5 from the 2023 table. Articles that quote a reply weight of 27 are using the day-one table from 31 March 2023, halved on 5 April. In 2026 a reply is about 7.6 likes.
What does the X algorithm reward in 2026?
Shares, conversation between people who follow each other, and time spent. Compared with 2023, X now leans towards copy-link and DM shares, quotes, follows and replies between mutual followers, and away from profile clicks and video views, which dropped to 0.
Read the weights as balance, not exchange rates. The 2023 README says they were set so that, on average, “each weighted engagement probability contributes a near-equal amount to the score”. A big weight mostly makes up for a rare action. The 2026 code says so outright: “one report cancels 468 likes” is wrong, because “the baseline probability of a Report is more than 1000x lower than a Like” (param.rs).
What it punishes: the negative weights are 10 to 50 times the positive ones, so a post that makes some viewers mute or click “not interested” loses far more than its likes win back. Beyond the weights, posts can be dropped or put behind a warning by visibility labels from Grok-based classifiers and account scores, whose criteria X doesn’t publish.
How do you grow on X with the algorithm?
From the published weights and the research below, not from anyone’s private data (the full routine, with growth stories and their numbers, is in how to get more followers on X):
- Reply early, under posts that are live. Since 2026 the feed no longer shows replies from accounts the viewer doesn’t follow, so your reply is seen under the post, where early replies sit higher. Founders who grew on X say most early followers came from replies (Marc Lou in his newsletter, for one; self-reported; see how Marc Lou grew).
- Turn replies into mutual follows. A reply between people who follow each other weighs 20 in the 2026 defaults, against 5 for a stranger.
- Make things people share by link or DM. A copy-link share is the heaviest positive action (20), a DM share 5. Tools, lists, data and screenshots get passed around; opinions mostly get liked.
- Post while the two-hour window is open. The new-author boost covers posts under 2 hours old from accounts under 50,000 followers, and early pick-up is what research says predicts spread.
- Don’t bait. A mute is −58.8 and “not interested” −52.272. Rage and “reply with 🔥” posts can win likes and lose reach. What does break out is in how to go viral on X.
- Skip the pods. Engagement coordinated in a group chat has no ranking impact.
- Measure posts against your own usual. Your past posts predict your reach better than any rule. A post at 3× your median taught you something; a post at someone else’s numbers didn’t. To see your own numbers, export your posts from X; to pick when to post, find your own best time.
Is the X algorithm open source?
Partly. In March 2023 Twitter published its ranking code (twitter/the-algorithm) and, days later, the weight table. In January 2026 xAI published Phoenix, built on the same transformer design as Grok, and on 13 August 2026 it added the default weights, filters and labelling systems.
| What’s public (2026) | What’s withheld |
|---|---|
| The 25 actions the ranker predicts for each viewer | The trained production model |
| How the predictions are added up (the weights) | Each viewer’s engagement history |
| Filters, boosts for new authors, diversity rules | Mutes, reports and “not interested” counts for a post |
| The label systems that can limit reach | The prompts behind the Grok-based content labels (“to reduce the risk of gaming”, per the README) |
| Live weight overrides and experiments |
So the recipe is public; the parts that decide whether a given person sees your post are not.
Does the X algorithm boost small accounts?
Yes, in the 2026 code. A new-author boost gives small accounts some exploratory slots near the top of feeds, using Thompson sampling, a standard way to try new items on a few people. Its limits in param.rs:
| Setting | Value |
|---|---|
ColdStartImpressionThreshold |
200 impressions |
ColdStartFollowerCap |
50,000 followers |
ColdStartMaxPostAgeSecs |
7,200 (2 hours) |
| Slots | 15–16 |
The 2026 code also drops posts older than 48 hours from For You, keeps a discount for posts from outside your network, lowers the score of repeated posts from one author in a feed, and no longer shows replies or reposts from accounts you don’t follow.
Do engagement pods work on X?
No. The 2026 code counts engagement only on posts X served in the Home timeline: “Directly navigating to a post (i.e., coordinating via groupchat) has no ranking impact” (param.rs). Likes from a “go like my post” chat don’t feed the ranker.
Does the X algorithm penalise links?
The sources disagree, and the code sits between them.
| When | Who | What they said |
|---|---|---|
| Oct 2023 | Elon Musk | Links “don’t get as much attention” because people leave X (Axios) |
| Nov 2024 | Elon Musk | Put the link in a reply (Benton) |
| Oct 2025 | Nikita Bier, X’s head of product | “Links are not deboosted” (Roboin) |
| Jul 2026 | Musk and Bier | Post links directly; no penalty “for over a year” (PPC Land) |
Buffer analysed 18.8 million posts and found link posts from non-Premium accounts had 0% median engagement from March 2025. The code has no explicit link penalty; opening a link is a small positive (+0.2). The likeliest reading: someone who leaves for a link doesn’t like, reply or stay, so the post’s predicted value falls on its own. No randomised test of links by creators has been published.
Does X Premium boost reach, and does reach pay?
Premium’s effect is unproven; reach now pays. Buffer (18.8M posts, August 2024 to August 2025) found median reach per post under 100 for regular accounts, about 600 for Premium and over 1,550 for Premium+. That’s about 10×, but it isn’t controlled for who buys Premium. The 2026 For You scoring code has no Premium multiplier like 2023’s (a default ×4 for followers, ×2 for others); Premium shows up in reply ordering instead.
Since January 2026, creator payouts follow “Verified Home Timeline impressions”: views by Premium users in their Home feed. Replies no longer count towards pay, and X said the pool “more than doubled” (PPC Land, Communicate, January 2026). Ranking and creator income are now the same thing.
What predicts how many views a post gets?
The algorithm scores what each viewer is likely to do, and that depends mostly on who you are to them. Three public studies on large data:
| Study | What it tested | What it found |
|---|---|---|
| Martin et al., WWW 2016 | Predicting how far 852 million tweets would spread, before posting | Content alone explained well under 5% of the variation. The author’s features and past success explained about 42–48%. Past success alone did nearly as well as everything together |
| Cheng et al., WWW 2014 | Facebook photo reshares: once a post has 5 reshares, will it reach 10? | 79.5% accuracy, mostly from how fast the early reshares came. Content features alone were close to a coin flip (49–52%) |
| Tan et al., ACL 2014 | Pairs of tweets by the same author with the same link: which wording got more retweets? | A model picked the winner 65.6% of the time (people: 61.3%), against about 55% for a model that ignored who posted |
Read together: before you post, your own track record predicts reach far better than any feature of the post. Once a post is live, how fast it picks up early is the best signal. Wording does matter, but you see it only when comparing two versions from the same person.
Audits of X’s feed agree on what rises (emotional and divisive posts; exposure concentrated on big accounts) but none saw the algorithm’s inputs, none covers the 2026 ranker, and none looked at non-political creator or business posts. Two examples: Milli et al. (PNAS Nexus, 2025: 806 users, divisive content amplified, and users didn’t prefer it) and Ye et al. (FAccT 2025: 120 test accounts, exposure concentrated on a few big accounts).
Do “X algorithm score” tools work?
None has shown that it does. Tools such as Xscore, TweetAlgorithm and various “X algorithm” analysers ask an AI to guess the probabilities, then multiply by X’s published weights. None publishes a check against real reach (searched October 2026). They have three problems:
- They can’t see the inputs that decide it. Your viewers’ histories, dwell time and the trained model aren’t public.
- They can’t see the negatives. Mutes, reports and “not interested” carry the heaviest weights and aren’t in any public count, so a score built from public numbers will overrate rage and reply bait.
- Research says content is the weakest predictor. Martin et al. put it under 5% of the variation before posting.
How do I find which posts beat an account’s usual?
Look at an account’s last 30 or so posts, take the median views, and mark the ones several times above it. Those are its breakouts; compare what they share. Parsly is a free Chrome extension that does this on any public X profile: open the profile, click Collect, and it reads each post’s views, likes, replies and reposts and marks the ones that beat that account’s own usual. It reads only public pages you can already see.
Questions people ask
How does the Twitter algorithm work in 2026?
Twitter is now X, and its For You feed uses a ranker xAI published in January 2026. It predicts 25 actions for each viewer (like, reply, share, mute and more), multiplies each prediction by a weight and adds them up. Filters and boosts then reorder the feed.
When did the X algorithm last change?
Often. The public code got the default weights on 13 August 2026 and near-daily commits after; a commit on 8 October 2026 improved how it predicts time spent, and one on 13 October trained retrieval on read-through. Check the weights against the code before trusting any guide, this one included.
How much is a reply worth on X compared with a like?
In the 2026 default weights, a reply is 5.0 and a like 0.66, about 7.6 times. The often-quoted "27 likes" comes from the 2023 table. X's own code warns that the weights multiply predicted probabilities, so a big weight mostly makes up for a rare action.
How do I get more views on X?
From the published weights and the research: reply early under live posts, turn replies into mutual follows, post things people share by link or DM, post while the two-hour new-author window is open, and avoid bait that makes people mute you. Then measure each post against your own usual.
Does X Premium increase reach?
Buffer found Premium accounts got about 10 times the median reach of regular ones from August 2024 to August 2025, but didn't control for who buys Premium. No 2023-style Premium multiplier was found in the 2026 For You scoring code.
Can a tool tell me my post's X algorithm score?
Tools exist that multiply guessed probabilities by X's weights, but none publishes a check against real reach. They can't see mutes, reports or "not interested", which carry the heaviest weights, so they tend to overrate bait.
Sources
Checked against the real sites on 28 October 2026. Platforms change their buttons often; if something has moved, tell us at [email protected].
- xai-org/x-algorithm: README (GitHub)
- xai-org/x-algorithm: home-mixer/params/param.rs, the 2026 default weights (GitHub)
- twitter/the-algorithm-ml: Heavy ranker README, 2023 (GitHub)
- Buffer: Links on X, 18.8M posts (October 2025)
- Buffer: Does X Premium really boost your reach? (October 2025)
- PPC Land: X doubles creator revenue pool (January 2026)
- PPC Land: X drops year-old link penalty (July 2026)
- Martin, Hofman, Sharma, Anderson, Watts: Exploring limits to prediction in complex social systems (WWW 2016)
- Cheng, Adamic, Dow, Kleinberg, Leskovec: Can cascades be predicted? (WWW 2014)
- Tan, Lee, Pang: The effect of wording on message propagation (ACL 2014)
- Milli et al.: Engagement, user satisfaction, and the amplification of divisive content on social media (PNAS Nexus, 2025)