How-To Guide
How does the X algorithm rank posts, and what should a brand change because of it?
X published its For You ranking code and weights in 2026. See how a post is scored, which vendor claims the files do not support, and what a brand should do.
By the Superdirector editorial team. X files read November 7, 2026.
At a glance
How does the X algorithm rank posts, and what should a brand change because of it?
X's public For You code describes a per-viewer candidate, filter and scoring pipeline, including a 48-hour candidate-age filter. The source was read November 7, 2026; the production model, deployed settings and other surfaces are not established by it.
- Share via copy link carries the largest positive weight (20.0), then the extra weight on replies between mutual follows (15.0), then reply, quote and share via DM at 5.0 each; report, mute and not interested carry the largest negative weights.
- Vendor claims of a 20x repost multiplier, a six-hour half-life, a followers-first test and a Premium boost have no source in X's files.
- Check Under the Hood for labels before treating low reach as a ranking problem; if you do not have it yet, use the median check in the last step.
The five steps X's README describes, with the lever a brand has at each. Original diagram from X's published For You code, read November 7, 2026.
Make it work for your brand
Find it. Remake it. Post it. Superdirector turns the viral formats in your niche into your brand's carousels, X posts and video scripts, and keeps the calendar full.
How the X algorithm ranks posts
X's published For You feed code, read November 7, 2026, describes gathering candidates, filtering posts older than 48 hours and seen, muted or blocked posts, and scoring predictions. The source does not establish the current production model, deployed settings or behavior of other surfaces.
The score is adjusted for repeat authors, strangers and new authors, and a separate visibility check decides whether a post can be shown at all. X released the ranking code and its weights on August 13, 2026. The repository's quickstart trains only a small model on synthetic data and the files read do not include X's production model, so no one outside X can say how much a single feature of your post moves a score.
| Step | What X's code does | What it means for a brand post |
|---|---|---|
| 1 Gather | Pulls recent posts from followed accounts and finds posts from accounts the viewer does not follow, then ranks both together with one model | Strangers can see you, but only after a model predicts their reaction |
| 2 Filter | Inspected code filters candidates older than 48 hours and seen, muted or blocked posts | Code example only; live settings and other surfaces unknown |
| 3 Predict | Estimates the chance of favorite, reply, repost, quote, share, clicks, dwell, follow and the negative actions | Every viewer gets a different prediction for the same post |
| 4 Score | Multiplies each prediction by a signed weight and adds them | Replies, quotes and sends carry more weight than likes |
| 5 Adjust | Scores down repeat posts by one author, discounts strangers, lifts authors with few impressions, then re-ranks for variety | Five posts in a row compete with each other |
| 6 Visibility | Allows each post, adds a notice or interstitial, or drops it, separately from ranking | A dropped post is not a low-scoring post |
The diagram below shows where a brand can act at each step. This page covers only the For You feed; the files read do not describe the Following tab, Search or Explore.
How the published rules changed
| Date | What X published or changed | Source |
|---|---|---|
| Apr 5, 2023 | Weights published: reply 13.5, repost 1.0, like 0.5, report -369.0 | the-algorithm-ml |
| Jul 10 to Jul 24, 2026 | Mutual-follow reply boost tested at 5, 10, 15, 20; 20 from July 13, 15 from July 24 | Boost record, TechCrunch |
| Aug 13, 2026 | For You code, filters and weights released under Apache 2.0, with Under the Hood | TechCrunch |
| Aug 14, 2026 | README note that weights multiply predicted probabilities, not counts | README |
| Oct 8 and Oct 13, 2026 | Ranking estimates time spent better; retrieval also trained for posts read all the way through | README |
| Nov 6, 2026 | Weights file last synced from feature-switch defaults (read Nov 7) | Parameters file |
Check Under the Hood before you change a post
Ranking sets the order and visibility filtering decides whether a post can be shown, and the README keeps them as separate services. A post can score well and still be dropped, so rule that out before you rework copy (a shadow ban is one name for a drop like this). X's Under the Hood page reports which visibility-limiting labels have been applied to your account and posts, in a downloadable JSON file. TechCrunch reported on August 13 that accounts which posted 10 or more times in the past month can download the file and that it started as a pilot for accounts at least a year old, so you may not see it yet.
- Open Under the Hood if you have it. A label on your account or posts explains a drop that no rewrite will fix.
- No access yet? Use your own numbers. The file cannot be read without access, so compare the post with your last ten at the same age (the median check below) and treat a repeated shortfall as a signal to look again once Under the Hood opens for you.
- Read the label against the code. X says the labels can be matched to the repository to see how visibility is affected.
- Know what is hidden. X withholds some moderation prompts and rules, and TechCrunch says systems that use Grok to flag rule-breaking are not in the release.
Check the post's age and who it could reach
The inspected For You source includes age and previously-seen filters and a mutual-follow adjustment. These describe the files read on November 7, 2026, not verified live settings or other surfaces. A brand can examine its own audience responses; the files alone do not establish a guaranteed reach effect.
What you will read, checked against X's code
Guides that rank for this question state multipliers, test batches and boosts. This table sets the common claims against X's repository and parameters file as read on November 7, 2026.
| You will read | What X's published files say |
|---|---|
| Reposts are worth about 20 times a like, replies 13.5 times and bookmarks 10 times (Sprout Social) | The 2026 file gives repost 1.0, like 0.708 and reply 5.0, and its list of predicted actions has no bookmark. The 13.5 matches the reply weight X published on April 5, 2023, when like was 0.5 and repost 1.0. |
| A post loses half its visibility every six hours (Sprout Social) | No half-life for For You scoring appears in the README or the files read. The inspected source filters posts older than 48 hours before scoring; live settings and other surfaces are unverified. The only half-life found, 8 hours, belongs to a shared popular-posts list. |
| X tests each post on 5 to 15 percent of your followers first (Buffer) | The README's pipeline has no follower-test stage. Followed and unfollowed accounts' posts are gathered in parallel and ranked together by one model. |
| X Premium gives a 2 to 4 times boost (Sprout Social) | No scoring, weights or re-ranking code read contains a Premium or verified term. The code does attach subscription status to each post, and the files read do not show what X's production model does with it. Buffer measured about 10 times more reach for Premium accounts across 18.8 million posts, which is an observation of accounts, not a rule in the code. |
| Links cut reach by 50 to 90 percent (Sprout Social) | The weights file has no link penalty. The code derives a has-link feature for each post and the model sees it, and a predicted link open carries a small positive weight of 0.2. The effect of a link is learned by a model X does not publish. Buffer's data shows a median engagement rate of 2.25 percent on link posts against 3.56 on text posts. |
| One report cancels hundreds of likes | X's README calls this reading incorrect: weights multiply predicted probabilities, not counts. |
Treat guides as routines to try and the repository as the record of what X ranks. Where the two disagree, the repository wins, and where it is silent, say so.
What the published weights say
X publishes the weight on each predicted action in a parameters file whose header says it is mirrored from X's feature-switch defaults, last synced on November 6, 2026. The README writes the score as the sum of each weight times the viewer's predicted probability of that action. The table lists the main entries.
| Predicted action | Weight in the file | Read as |
|---|---|---|
| Share via copy link | 20.0 | Highest positive weight |
| Reply from a mutual follow (extra weight) | 15.0 | Added for original posts from accounts the viewer and author both follow |
| Reply, quote, share via DM | 5.0 each | Conversation and sending |
| Follow author | 4.0 | The viewer follows after seeing the post |
| Share | 2.118 | |
| Repost | 1.0 | |
| Like | 0.708 | Lighter than repost, share and reply |
| Click, open link, dwell | 0.274, 0.2, 0.05 | Attention and clicks |
| Report | -234.0 | Largest negative weight |
| Mute author, not interested, block author | -58.8, -53.07, -31.2 | Negative feedback |
X's README warns against reading these as exchange rates. The weights multiply your own predicted probability of each action, not raw counts, so a report is not worth hundreds of likes. The file's comments add that the baseline chance of a report is more than 1000 times lower than a like, which is why its weight is large. A comment in the inspected Home Timeline scoring file describes a limitation for actions on posts opened directly from group chats. Current deployment and effects on other surfaces are unverified.
The 15.0 mutual-follow line is a change you may have seen discussed, and it is added to the 5.0 reply weight, not a replacement for it: the scoring file's own test adds the boost to the base reply weight, so a reply between mutual follows is weighted 5.0 plus 15.0, or 20.0 in total. In July 2026 X tested a mutual-follow reply boost at values of 5, 10, 15 and 20, rolled out 20 on July 13 and cut it to 15 on July 24. TechCrunch reported that X's head of product described the change as making the reply section feel less like a battleground with people you do not recognize. The full list, with sources and dates, is in the weights sheet.
Negative actions carry far larger weights than positive ones, drawn on a compressed scale. Original diagram from X's parameters file, read November 7, 2026.
Worked example: one post scored for two viewers
Take one promotional post from a skincare brand and two people who might see it. The first is a customer who follows the brand. The second is a stranger who tends to mute ads. X does not publish anyone's predicted probabilities, so the ones below are made up to show the arithmetic; the weights are the published ones.
| Action (weight) | Customer: probability x weight | Stranger: probability x weight |
|---|---|---|
| Like (0.708) | 0.08 x 0.708 = 0.057 | 0.02 x 0.708 = 0.014 |
| Reply (5) | 0.01 x 5 = 0.050 | 0.001 x 5 = 0.005 |
| Share via copy link (20) | 0.002 x 20 = 0.040 | 0.0002 x 20 = 0.004 |
| Mute author (-58.8) | 0.0005 x -58.8 = -0.029 | 0.01 x -58.8 = -0.588 |
| Report (-234) | 0.00005 x -234 = -0.012 | 0.0005 x -234 = -0.117 |
| Score before adjustments | 0.106 | -0.682 |
With these made-up numbers the same post scores well for the customer and badly for the stranger, and the difference is almost all one line: a 1 percent chance of a mute outweighs every positive action the stranger might take. The report line is small for both, because a report is rare. The arithmetic shows how steep the mute weight is, but X publishes nothing about which wording earns mutes. Treat what follows as our reading of the weights: write the post your followers want to answer or forward, and watch for post styles your own audience mutes.
The probabilities are invented; the weights and the sum are X's. Original diagram from X's parameters file, read November 7, 2026.
Read the reactions the weights pay for
After share via copy link (20.0) and the mutual-follow reply extra (15.0), reply, quote and share via DM carry the largest positive weights at 5.0 each, and report, mute, not interested and block carry large negative ones. Each multiplies a probability that depends on the viewer, so you cannot target a count. You can write a post that someone in your niche wants to answer or forward, and keep it away from people who do not.
- End with something answerable. A question your buyers have a view on, or a result worth a quote post.
- Write something worth sending. Share via copy link is the heaviest positive weight and share via DM is among the next, and a how-to or a number a colleague would forward is what people send.
- Skip bait. A post that earns mutes and reports is predicted to earn them from similar viewers next time.
Check whether readers got to the end
X added two changes in October 2026 that point the same way. The ranking model estimates how long you will spend on each post better (the same idea as dwell time), and retrieval is also trained to surface posts people read all the way through. X publishes no ideal length, so do not pad or cut to a number.
- Put the point in the first line. A reader who learns what the post is about can decide to finish it.
- Cut anything that does not change the reader's next step.
- Keep images and charts that the text needs. The code passes a post's media count, video length and text length to the model as features, and X does not say how it uses them.
Check whether your own posts crowd each other
The inspected For You source describes author and variety adjustments and a 48-hour candidate-age filter. Live configuration and other surfaces are unverified; this source does not establish that a post must be republished after two days. Test a sustainable posting cadence with your own audience.
Compare each post with your own median
Read impressions at the same post age and compare against the median of ten valid earlier observations. If fewer than ten are available, the new value is missing, or the median is zero, the comparison is inconclusive; a recorded zero remains zero, not a win. These are editorial thresholds, not X targets or proof of a cause. With fictional values 410, 520, 380, 610, 450, 700, 390, 480, 560 and 430, the median is 465; 1,020 is about 2.2 times the baseline and 220 is below half. Either pattern suggests a follow-up test, not proof that the topic or format caused it.
To test links, post the same point twice in different weeks, once with the link in a reply and once in the post, and compare each against your median instead of any published penalty.
| Field | Where it comes from | What it tells you |
|---|---|---|
| Impressions at 48 hours | X analytics | How far the post travelled while it was eligible |
| Replies and quotes per 1,000 impressions | X analytics | Whether viewers answered or forwarded |
| Ratio to your median of ten | Your sheet | Whether the format beats your own normal |
| Labels | Under the Hood | Whether ranking was ever the problem |
Half and double the median are our rule of thumb, not an X target. Original diagram with invented impression counts, from this guide's method.
Where Superdirector fits
Superdirector, which publishes this guide, does not change how X ranks anything. Its headline is "Autonomous social media growth." What it makes and posts today is carousels and X posts, while AI video and full autopilot are still to come. For X it watches the accounts, competitors and keywords you name, spots which posts and formats are taking off in your niche, and drafts X posts in your brand's voice with a similar hook and structure. At $199 a month the offer is one AI-run social account. Eligible Instagram carousels or a single link-free X text post can be published to a connected account under its standing posting permission, subject to account readiness and plan entitlement. The $29 a month tier is the hands-on entry: you drive it yourself through MCP from your own AI assistant. Either way, the median check above is how you judge whether the posts are working. See what runs today.