The X algorithm in 2026 gathers candidate posts from accounts you follow and from similar accounts you don't, uses a Grok-style transformer to predict how likely you are to like, reply, share, follow, mute or report each one, multiplies those predictions by published weights, filters the results and ranks your feed. Replies, shares and follows count most; reports and mutes count heavily against you.
For most of X's history, creators had to reverse-engineer the feed from rumours and screenshots. That changed in 2026. X open-sourced its new ranking system in January, added the scoring weights in August, and now posts dated notes when it changes things. This guide walks through the system step by step using that public code, compares it with the 2023 version, breaks down what it means by account type, content format and goal, and finishes with a step-by-step method, the common mistakes and answers to the questions we hear most.
Key findings
- X released the code for its For You ranking system on January 20, 2026, describing it as built on "the same transformer architecture as xAI's Grok model," per Decrypt. An August 13, 2026 update added the ranking weights, per the x-algorithm repository.
- In the published weights, a reply is worth 5.0, a follow 4.0, a repost 1.0 and a like 0.5. Sharing by copying a post's link is worth 20.0, and a reply between mutual followers gets an extra 15.0, per X's param.rs.
- Negative signals are far larger: a report is weighted -234.0, a mute -58.8, a "not interested" tap -47.52 and a block -31.2.
- The feed drops posts older than 48 hours, discounts each extra post from the same author and discounts posts from accounts you don't follow, according to the repository README.
- A peer-reviewed study of tweet impressions found a median half-life of about 80 minutes, with roughly 95% of tweets getting almost no meaningful impressions after 24 hours, per Pfeffer, Matter and Sargsyan (2023).
- Regular accounts typically get under 100 impressions per post, Premium about 600 and Premium+ over 1,550, in Buffer's study of 18.8 million posts.
- Since late 2025, the Following tab is also ranked by Grok by default, with a chronological view available on request, per Social Samosa.
What's changed in the X algorithm in 2026
January: a new open-source ranker
X's 2023 open-source release described a "heavy ranker" neural network fed with thousands of hand-built features. The 2026 system replaces that with an end-to-end transformer that does not rely on hand-engineered features. Musk acknowledged its limits at launch ("We know the algorithm is dumb") and promised updates every four weeks with notes explaining what changed, as Decrypt reported.
August: the weights and the safety code
The August 13, 2026 release added the key configuration parameters, including the blend weights used for scoring, plus the code for visibility filtering and the labelling systems that feed it. It also added SimClusters as an additional source of posts from accounts you don't follow and launched a transparency tool called Under the Hood. On August 14, X added comments clarifying that weights scale predicted probabilities, not raw engagement counts.
September: legal visibility notices
A September 18, 2026 update made the Under the Hood reports show whether a post's or account's visibility was limited to comply with law, and in which country. X also runs a filter for Brazil's 2026 elections that removes posts from accounts reported to Brazil's Electoral Court, unless the viewer follows them.
The Following tab became ranked
In late November 2025, Musk announced that "Posts of people you follow are now ranked by Grok," and the app began opening to For You even for users who had left it on Following. Your followers no longer see your posts strictly in time order unless they choose the chronological view.
The link debate, resolved on paper
In April 2026, X's head of product Nikita Bier called claims of a link penalty "fake news" and advised posting links in the body of a post with context, per Roboin. The weights file has no explicit link penalty; it gives a small positive weight (0.2) to opening a link. But earlier Buffer data still shows link posts underperforming, which we cover below.
How the X algorithm works, step by step
Every time you open the app, the system builds your feed in five stages. Here is each one in plain language.
Step 1: gathering candidates
X collects a pool of recent posts that might interest you from several sources:
- Thunder (in-network): an in-memory store of recent posts from accounts you follow, with up to 1,200 results in the default configuration.
- Phoenix retrieval (out-of-network): a model that finds posts from accounts you don't follow but are likely to engage with, up to 1,000 results.
- SimClusters: community-based recommendations from interest clusters, limited to posts up to 48 hours old.
- Popular posts: a shared list of up to 500 widely engaged posts.
These numbers are defaults from X's configuration file and can be changed at runtime.
Step 2: filtering
Before scoring, the system removes duplicates, posts older than 48 hours, your own posts, posts from accounts you have blocked or muted, and posts you have already seen. Subscriber-only posts you cannot access are also removed.
Step 3: understanding each post
As posts are published, classifiers called grox detect categories "such as spam, adult content and violent media" and produce numeric representations of the text and images. The system also attaches details such as the author's labels, language, engagement counts and subscription status.
Step 4: predicting your actions and scoring
The Phoenix ranker, a transformer model, estimates the probability that you will take each of a long list of actions on each post: like, reply, repost, quote, share, click the post, open the profile, open a link, expand a photo, open or keep watching a video, dwell, follow the author, tap "not interested," mute, block or report. A scorer multiplies each probability by its weight and adds them up. The next section lists every weight.
Step 5: diversity, discounts and final filters
After scoring, each post after an author's first in your feed is multiplied by a decaying factor, so one account cannot flood your timeline. Posts from accounts you don't follow are multiplied by a factor below 1, and so are replies and reposts from followed accounts. A new-author boost lifts posts from accounts whose impressions are below a threshold. Finally, visibility filtering gives each post one of four outcomes (allow, notice, interstitial or drop), and a conversation filter collapses several branches of the same thread.
| Stage | What happens | What it means for you |
|---|---|---|
| Candidates | Posts pulled from follows, similar accounts, interest clusters and popular posts | Topic consistency helps the system find people like your followers |
| Filters | Old, seen, muted and blocked content removed | Posts have about 48 hours to be distributed |
| Understanding | Spam, adult and violent content detected; text and images embedded | Spammy patterns are caught at publish time |
| Scoring | Predicted actions times weights, summed | Replies, shares and follows matter most; reports and mutes are costly |
| Diversity and filters | Repeat-author decay, out-of-network discount, new-author boost, safety labels | Posting many times in a row has diminishing returns; small accounts get a lift |
The ranking weights: what every action is worth
These are the default weights in X's param.rs. Each multiplies the model's predicted probability that a viewer will take that action, not the number of times it happened. A post with a 10% chance of a reply gets 0.1 x 5.0 = 0.5 from replies, the same as a 100% chance of a like.
| Predicted action | Weight | Relative to a like |
|---|---|---|
| Share via copied link | 20.0 | 40x |
| Reply between mutual followers (boost) | +15.0 | Adds 30x |
| Reply | 5.0 | 10x |
| Quote | 5.0 | 10x |
| Share via DM | 5.0 | 10x |
| Follow the author | 4.0 | 8x |
| Share | 2.0 | 4x |
| Repost | 1.0 | 2x |
| Like | 0.5 | Baseline |
| Click dwell time (continuous) | 0.4 | Per unit of time |
| Click into post | 0.3 | 0.6x |
| Open link | 0.2 | 0.4x |
| Video open | 0.07 | Small |
| Dwell | 0.05 | Small |
| Photo expand | 0.05 | Small |
| Profile click | 0.0 | None by default |
| Not dwelled (scrolled past) | -0.02 | Slightly negative |
| Block author | -31.2 | Strongly negative |
| Not interested | -47.52 | Strongly negative |
| Mute author | -58.8 | Strongly negative |
| Report | -234.0 | Most negative |
Three lessons jump out. First, conversation is the engine: replies, quotes and DM shares all sit at 5.0, and replies between people who follow each other get a large extra boost. Second, reference value pays: a copied link, which usually means someone is sending your post elsewhere, is the single most valuable positive action. Third, annoying people is expensive: one predicted report cancels out the value of hundreds of predicted likes.
2023 vs 2026: how the weights changed
X's 2023 open-source release also published weights for its heavy ranker, dated April 5, 2023, in the the-algorithm-ml repository. The scales differ, so compare the order rather than the raw numbers.
| Signal | 2023 weight | 2026 weight | What changed |
|---|---|---|---|
| Like | 0.5 | 0.5 | Still the baseline |
| Repost | 1.0 | 1.0 | Unchanged |
| Reply | 13.5 | 5.0 | Still a top signal, now alongside quotes and DM shares |
| Reply engaged by the author | 75.0 | Not a separate weight | Replaced by a 15.0 boost for replies between mutual followers |
| Good profile click | 12.0 | 0.0 (profile click) | Profile visits no longer score by default |
| Good click | 11.0 | 0.3 click, 0.4 click dwell | Clicks matter less; time spent after clicking still counts |
| Follow the author | Not listed | 4.0 | New explicit signal |
| Share via copied link | Not listed | 20.0 | New; the highest positive weight |
| Negative feedback | -74.0 | -47.52 to -58.8 | Split into not interested, mute and block |
| Report | -369.0 | -234.0 | Still by far the most negative |
The broad shape has held for three years: likes are cheap, conversation is valuable and negative feedback is very costly. What is new is the emphasis on sharing and following, which reward posts people want to pass on and accounts people want more of.
How long a post lives on X
X's feed filters out posts older than 48 hours, and in practice most reach happens far sooner. Researchers at the Technical University of Munich tracked impression counts over time and found impressions per second peak about 72 seconds after posting, the median half-life of a tweet is about 80 minutes, and for roughly 95% of tweets almost no meaningful impressions occur after 24 hours, in a paper presented at ICWSM 2023. That study predates the 2026 ranker, but the 48-hour filter suggests the short life still applies.
The practical consequence: the first hour decides most of a post's reach, so post when your audience is online and be there to answer replies. Sprout Social's 2026 analysis of nearly 2 billion engagements found X engagement strongest Tuesday to Thursday from 12 to 6 p.m. local time, per Sprout Social. See our guide to the best time to post on X for the full breakdown.
Breakdown by content format
The ranker treats formats through the actions they lead to. Large datasets show which formats earn those actions.
| Format | Median engagement rate (Buffer, 2026) | Regular accounts (Buffer, 2025) | Premium accounts (Buffer, 2025) |
|---|---|---|---|
| Text | 3.56% | About 0.40% | About 0.90% |
| Image | 3.40% | About 0.20% | About 0.42% |
| Video | 2.96% | About 0.25% | About 0.85% |
| Link | 2.25% | 0% | About 0.28% |
Sources: Buffer's 2026 format study and Buffer's analysis of links on X, which use different methods, so compare within each column. In both, text leads and links trail.
Text posts
Text is the native format of X and earns replies most easily. X counts characters with weights: URLs count as 23, emojis and CJK characters as 2, per X's developer docs. Check drafts with our X character counter.
Images
Photo expands carry only 0.05, so images help through what they lead to: replies on a comparison, shares of a useful chart. Our guide to images on X covers sizes and image types.
Video
The ranker predicts video actions such as quality views and continued watching, but the default weight on a quality view is 0.0 and on a video open 0.07, so video earns its score mostly through the same replies and shares as anything else.
Links
X says links are not penalised, and the weights give opening a link a small positive value. But link posts take people off X, and readers often forget to come back and reply, which Bier himself acknowledged in 2025. If you share links, put the value in the post and explain what the link offers.
Threads
Each post in a thread is scored separately, and the conversation filter collapses branches, so the first post does most of the work. See our X thread strategy.
Breakdown by account type
Regular (non-Premium) accounts
Buffer's data covering August 2024 to August 2025 found regular accounts typically get under 100 impressions per post, and their median engagement rate fell to 0% by March 2025. The weights do not mention Premium, but candidate hydration includes subscription status, and Buffer's data shows the reach gap is roughly tenfold. Growth without Premium leans heavily on replies to larger accounts.
Premium and Premium+ accounts
Premium accounts in Buffer's data got about 600 impressions per post and Premium+ over 1,550. About 27% of accounts in the sample paid for a tier. If X is a business channel for you, Premium is close to a requirement; check current prices in the app.
New and small accounts
The new-author boost lifts posts from accounts whose impressions are below a threshold toward a target position. That gives small accounts a genuine chance, but only if the post is good enough to earn replies from the people it is shown to. Our guide to getting more X followers has a 30-day plan.
Large accounts
Author diversity decay means a large account's sixth post of the day reaches far fewer people than its first. Large audiences also produce more mutes and "not interested" taps in absolute terms, so a poorly judged post costs more.
Brand accounts
Brands face the same weights as people. Posts that read like adverts invite "not interested" taps. Brand accounts do best with useful, specific posts, quick customer replies and a human voice. Sprout's roundup reports that 58% of X users interact with brand content at least weekly, per Sprout Social.
Breakdown by goal
| Goal | Signals to aim for | Tactics |
|---|---|---|
| Reach | Replies, quotes, shares | Clear opinions, original data, questions; post at peak times |
| Followers | Follow (4.0), replies from mutuals | Narrow topic focus, a reason to follow, replying to larger accounts |
| Traffic | Link opens, click dwell | Give context before links; strong reason to click |
| Community | Mutual replies (+15.0) | Reply to everyone who replies; engage with peers daily |
| Brand safety | Avoid reports, mutes, blocks | No bait, no spam patterns, accurate claims |
Breakdown by industry
The same weights apply to everyone, but what earns replies and shares differs by field. Here is how the signals translate for six common account types.
| Industry | Posts that earn high-weight actions | Posts that risk negative signals |
|---|---|---|
| B2B SaaS | Original data, product lessons, honest launch numbers | Repetitive feature promos, generic thought leadership |
| Developer tools and AI | Code snippets, benchmarks, demos, sharp technical opinions | Hype without evidence, misleading benchmarks |
| Media and newsletters | Context on breaking news, charts, quotable findings | Headline-only link posts, clickbait |
| Finance and investing | Clear explainers, data with sources | Unverified tips, promises of returns |
| E-commerce and consumer brands | Fast customer replies, product questions, behind the scenes | Constant discount posts, reply-guy marketing |
| Agencies and freelancers | Before and after results, teardowns, process screenshots | Cold pitches in replies, unrelated trend-jacking |
The pattern holds across fields: substance that people want to answer or pass on wins, and anything that feels like an interruption loses. Track which of your posts earn replies and follows in a simple log, using the definitions in our glossary entry on impressions, and lean into the topics that work.
Who you are reaching
The algorithm decides which users see your posts, but the pool is shaped by who uses X. Pew's 2025 data shows 33% of US adults aged 18 to 29 use X versus 10% of those 65 and older, men are more likely to use it than women (25% vs 16%), and college graduates use it more than those without a degree, per the Pew social media fact sheet. Pew's 2025 report found 10% of US adults use X daily, per Americans' Social Media Use 2025. If your buyers are not on X, no ranking trick will fix that; compare options in our X vs LinkedIn guide and Bluesky vs X comparison.
For You vs Following: two feeds, one ranking approach
X has two main timelines, and they now work more alike than they used to.
| For You | Following | |
|---|---|---|
| Whose posts | Accounts you follow plus recommended accounts | Only accounts you follow |
| Order | Ranked by predicted actions | Ranked by Grok by default since late 2025; chronological on request |
| Default when opening the app | Yes, since the late 2025 change | No |
| Out-of-network discount | Applies to recommended posts | Not relevant |
| What it means for you | Strangers can find you if early engagement is strong | Even followers may miss posts that earn no early engagement |
Social Samosa's report noted that the change may matter little for people who follow fewer than 200 accounts, because there are fewer posts to rank. For heavy users who follow thousands of accounts, ranking now decides most of what they see from you.
Safety, visibility and what X can limit
Ranking is only half the picture. X's code also decides whether a post can be shown at all.
- Classifiers at publish time. Grox classifiers label posts for categories such as spam, adult content and violent media as they are published.
- Four visibility outcomes. Each post gets ALLOW, NOTICE, INTERSTITIAL or DROP, which decides whether it is shown normally, with a notice, behind a warning screen or not at all.
- Legal withholding. Since September 2026, X's Under the Hood reports show whether a post or account was limited to comply with a law, and which country.
- Election filters. The Brazil 2026 election filter removes posts from accounts reported to Brazil's Electoral Court from For You, unless the viewer follows them.
For most businesses and creators none of this applies day to day. But if a post suddenly gets almost no impressions, check for a notice or label before assuming the ranking model has turned against you.
Myths about the X algorithm, checked against the code
"Hashtags boost distribution"
Hashtags do not appear anywhere in the published weights. They can help people find live conversations through search, but they are not a ranking lever. One relevant hashtag at most.
"Profile clicks are a huge signal"
That was true in 2023, when a "good profile click" was weighted 12.0. In the 2026 defaults, profile click is weighted 0.0. Follows (4.0) are what count now, so give people a reason to follow, not just to look.
"Replying to your own post's replies is worth 75x a like"
That number comes from the 2023 weights for a reply the author engaged with. The 2026 code replaces it with a 15.0 boost for replies between mutual followers. Answering replies still matters a great deal, but quote the current numbers.
"X shows your posts to all your followers"
Not any more. Both timelines are ranked by default, and posts that earn no early engagement may not reach many of your followers.
"Posting more always means more reach"
More posts can mean more total reach, but each extra post from the same author in one viewer's feed is discounted. Buffer also found reach per post tends to decline at higher posting frequencies. Spacing and quality matter.
"Old posts can go viral days later"
Rarely in the ranked feed, which filters out posts older than 48 hours. Old posts can still spread through search, links and quote posts, but plan for a short window.
A worked example: applying the code to a real week
Take a fictional freelance data analyst, Marco, with 1,800 followers on a regular account. His posts get 60 to 120 impressions and a handful of likes. Here is what the published system suggests he change.
Before: three posts a day, often within the same hour, mostly links to his blog with a one-line summary, on varied topics from football to productivity to SQL. He rarely replies.
After:
- He narrows to two topics, SQL tips and analytics careers, so interest-based candidate sources can match him with the right people.
- He schedules two posts a day, at 12:30 and 4:30 p.m. on weekdays, to avoid repeat-author decay and hit Sprout's peak window.
- He replaces most link posts with the actual tip in the post, plus a screenshot of the query, and ends with a question ("How would you write this without a subquery?").
- He spends 15 minutes a day replying to larger analytics accounts with genuinely useful answers, and answers every reply on his own posts.
- Once a week, he posts a short checklist people want to save and send, aiming for copied-link shares.
- He considers Premium after a month, once he knows X is worth the investment for his business.
None of these steps games the system. Each one simply makes the actions the model predicts (replies, shares, follows) more likely, and the negative ones (mutes, "not interested") less likely. That is the only durable way to work with any ranking model, as our overview of how social media algorithms work explains.
How X compares with other feeds
X is not the only network that predicts engagement, but its emphasis differs.
| Network | What it emphasises | Typical post life |
|---|---|---|
| X | Replies, shares, follows; strong penalties for mutes and reports | Hours; filtered after 48 hours |
| Dwell time, comments, professional relevance | Often several days | |
| Threads | Predicted likes, replies, clicks and viewing time | Short; ranked by recent engagement |
| Bluesky | User-chosen custom feeds alongside a chronological option | Depends on the feed the reader picks |
Threads' ranking signals come from Meta's Threads feed explainer, and Bluesky's approach of letting users pick their own feed algorithms is set out in its 2023 post on algorithmic choice. For the LinkedIn side, see how the LinkedIn algorithm works. And if you are weighing X as an income source rather than a marketing channel, read how much X pays creators.
What the open-source code does not tell you
Publishing the code is a big step, but it has limits worth knowing before you build a strategy on it.
- Runtime overrides. The weights are defaults. X notes that feature switches can change them at runtime, so the live values may differ from the file on any given day.
- The model itself. The repository includes training code and synthetic data generation, not X's real user data. You can see how predictions are combined, but not exactly how the model predicts that a given viewer will reply.
- Ads. Paid placements follow their own systems, so promoted posts do not tell you much about organic ranking.
- Search and Explore. The repository covers the For You feed. Search results, trends and other surfaces may rank differently.
The safest reading is to treat the weights as a statement of priorities rather than a formula to optimise. They describe what X wants its feed to reward: conversation, sharing and following, with very little tolerance for content people want to hide.
A step-by-step method for working with the algorithm
- Pick two or three topics and stick to them. Candidate sources like SimClusters work through interest communities. A consistent topic helps X find people like your current followers.
- Write for replies. End posts with a real question, a clear opinion or an invitation to share examples. Replies are weighted 5.0, ten times a like.
- Make reference material. Checklists, data and frameworks get shared by copied link, the highest positive weight at 20.0.
- Reply to everyone who replies. Replies between mutual followers get an extra 15.0 boost. Follow back people you genuinely talk with.
- Reply to larger accounts in your niche. Your reply is a post that their audience sees; a useful one earns profile visits and follows.
- Post at your audience's peak times. Tuesday to Thursday afternoons are a strong default. Schedule posts ahead in an X scheduler so timing does not depend on your day.
- Space your posts. Author diversity decay means bursts compete with each other. Spread posts through the day rather than posting five in ten minutes.
- Be present for the first hour. Most reach happens fast, with a median half-life of about 80 minutes in the ICWSM study.
- Avoid anything that earns mutes or reports. Bait, repetitive promotions, misleading claims and off-topic rants carry the largest negative weights.
- Decide on Premium. If X is a business channel, Buffer's data makes a strong case for it.
- Measure the right signals. Track replies, follows, profile visits and impressions per post, not just likes. See X metrics and KPIs.
- Re-read the changelog monthly. X promises updates every four weeks with notes. Check the repository's notable updates and adjust.
Putting it into a weekly routine
The algorithm rewards two things that pull in opposite directions: consistent, well-timed posts and live conversation. The way to get both is to separate them. Write and schedule your posts in one weekly session, then spend your daily X time on replies.
Postprism handles the first half. You write a week of single X posts, with up to four images each, schedule them at your peak times on one content calendar, and see results per post in analytics. The same calendar holds your LinkedIn, Bluesky and other network versions, so one session covers every channel. Buffer found that the gap between top accounts and median accounts in posting frequency is widest on X, per its 2026 engagement report, so a steady schedule matters more here than almost anywhere.
Schedule Your Twitter/X Posts in Advance
Write a week of X posts in one sitting, schedule them for your audience's peak hours, and spend your daily time on the replies the algorithm rewards most.
| Time | Task | Algorithm signal it serves |
|---|---|---|
| Monday, 60 minutes | Write and schedule the week's posts | Consistency, peak timing, spacing |
| Daily, 15 minutes after your main post | Answer every reply | Replies, mutual reply boost |
| Daily, 15 minutes | Reply to 5 to 10 larger accounts in your niche | New-audience exposure, follows |
| Friday, 15 minutes | Check which posts earned replies and follows | Learning what your audience rewards |
Common mistakes (and how to fix them)
- Optimising for likes. Likes carry the lowest positive weight. Fix: write for replies, shares and follows.
- Rage bait. Replies from angry users come with mutes and reports, which cost far more. Fix: argue clearly and fairly, and invite disagreement politely.
- Engagement pods and reply rings. Artificial patterns look like spam to classifiers that run as posts are published. Fix: build real conversations with peers.
- Posting in bursts. Each extra post from the same author is discounted. Fix: space posts through the day, and schedule them ahead.
- Posting and leaving. The first hour is when most reach happens. Fix: schedule posts for times you can be online, and answer replies.
- Ignoring replies to your posts. Mutual replies carry a +15.0 boost. Fix: reply to every genuine reply.
- Bare links. Link posts trail every other format in Buffer's data. Fix: give the value in the post and context for the link.
- Wandering topics. Interest-based sources work best when you are consistent. Fix: stick to two or three themes.
- Treating hashtags as distribution. Hashtags do not appear in the published weights. Fix: use one only when it names a real live conversation.
- Running a business account without Premium. Regular accounts typically get under 100 impressions per post. Fix: subscribe, or invest more heavily in replies.
- Believing old guides. Many still quote the 2023 weights. Fix: read the current repository and its dated notes.
- Not measuring. Without tracking you cannot tell what the algorithm rewards for your audience. Fix: review replies, follows and impressions weekly.
Frequently Asked Questions
How does the X algorithm work in 2026?
It collects recent posts from accounts you follow and from similar accounts you don't, filters out old, seen and muted content, predicts how likely you are to take each of many actions on each post with a Grok-style transformer, multiplies those predictions by published weights and ranks the results with diversity and safety filters on top.
What does the X algorithm value most?
In the published weights, sharing a post by copying its link (20.0), replies (5.0), quotes (5.0), DM shares (5.0) and follows (4.0) count most. Replies between people who follow each other get an extra 15.0 boost. Likes are worth 0.5. Reports (-234.0), mutes (-58.8), "not interested" taps (-47.52) and blocks (-31.2) count heavily against a post.
Is the X algorithm really open source?
Yes. X published the code for its For You ranking system on GitHub in January 2026 and added the ranking weights and visibility filtering code in August 2026. The weights are defaults, and X notes that runtime settings can override them, so treat them as relative priorities rather than exact values.
Does X penalise posts with links?
X's head of product says no, and the published weights include no explicit link penalty. However, Buffer's data shows link posts get the lowest engagement of any format on X, likely because readers leave and do not come back to reply. Put the value in the post and add context before any link.
Does X Premium boost reach?
Buffer's study of 18.8 million posts found Premium accounts got about 600 impressions per post and Premium+ over 1,550, against under 100 for regular accounts. The ranking system's candidate data includes subscription status, though the published weights do not list a Premium bonus directly.
How long does a post stay in the X feed?
The For You feed filters out posts older than 48 hours. Research presented at ICWSM 2023 found a median half-life of about 80 minutes, with roughly 95% of tweets getting almost no meaningful impressions after 24 hours. The first hour matters most.
Is the Following tab chronological?
Not by default since late 2025. Posts from accounts you follow are ranked by Grok, and users can still switch to an unfiltered chronological view. This means early engagement affects how many of your own followers see a post.
How often should I post on X for the algorithm?
Often enough to be consistent, spaced enough to avoid repeat-author decay. Buffer found top accounts post more often than median accounts and that the gap is widest on X. For most people, two to five well-spaced posts a day plus daily replies works; scheduling them ahead keeps the rhythm steady.




