Most advice about the X algorithm still treats engagement like a points system. One like earns one point. A reply earns 150. A report cancels hundreds of likes. Post at the right time, add a question, avoid links and hope the machine notices.
That model is wrong.
In 2026, X has published enough of its For You ranking system to explain the real mechanics. The public code shows a personalised prediction model called Phoenix, a weighted scoring layer, author-diversity adjustments, out-of-network discounts, cold-start boosts and separate visibility filters. It does not reveal a simple viral formula, but it does reveal what the system is trying to predict.
This is the part most guides miss: X does not reward engagement counts directly. It rewards the predicted probability that a specific viewer will take a specific action on a specific post.
That distinction changes how you should write, measure and distribute content.
What changed in the X algorithm during 2026?
The biggest change was not a new “reply hack” or a tweak to the Premium boost. X published a substantially more complete version of the For You recommendation system in August 2026.
The repository now includes:
- The Phoenix retrieval and ranking model
- Production-style ranking code
- Action weights used to combine predictions
- Visibility filtering systems
- Author-diversity scoring
- Out-of-network adjustments
- Cold-start and low-impression boosts
- Training code and synthetic data generation
X’s own repository says the For You feed combines posts from accounts you follow with posts retrieved from accounts you don’t follow. Phoenix then predicts how likely you are to take a range of actions on each candidate post. Those predictions are combined using weights in the code. (github.com)
That is much more useful than vague advice about “quality content”.
But it still needs careful reading. The numbers in the code are not points handed out for raw actions. They are coefficients applied to predicted probabilities and continuous values such as dwell time. X explicitly added a warning because people were interpreting the weights as direct engagement-count equivalents. (github.com)
How this differs from other X algorithm guides
The current crop of articles tends to make one of five mistakes:
- It repeats old advice from the first open-source release.
- It treats the weights as a literal points table.
- It focuses only on likes, replies and reposts.
- It ignores the filters that decide whether a post can be recommended at all.
- It confuses what is in the published code with what has been proven to improve a creator’s reach.
Some guides also claim that replies are worth 150 likes, or that one report cancels 468 likes. The first statement may be a rough ratio from an earlier interpretation of the code. The second is explicitly rejected by X’s own comments.
This article takes a narrower approach. It explains every published ranking weight, what each one means, where it applies and what you should do with the information.
It is not another list of generic posting tips.
How does X rank a post in 2026?
The basic scoring equation is:
Final Score = Σ (weight_i × predicted probability of action_i)
Phoenix predicts the probability of many possible actions. The ranking scorer multiplies each probability by its corresponding weight, then adds the results together. Positive actions increase the score. Negative actions reduce it. (github.com)
A simplified example might look like this:
Like prediction × 0.5
+ Reply prediction × 5.0
+ Repost prediction × 1.0
+ Copy-link prediction × 20.0
- Not-interested prediction × 43.2
= preliminary score
That does not mean one copy-link click is worth 40 likes in a universal sense. It means the model gives the predicted probability of a copy-link action a coefficient 40 times larger than the predicted probability of a like.
The predicted probabilities are personalised. They are influenced by the viewer’s behaviour, interests, follows, past interactions, language, location and other contextual features. A copy-link prediction may be close to zero for one viewer and much higher for another.
So your post does not have one fixed algorithm score.
It has a potential score for each viewer.
The ranking system is not the whole distribution system
The published code describes several stages:
- X retrieves possible posts from followed and unfollowed accounts.
- It hydrates the post text, media and author data.
- It removes ineligible or repeated content.
- Phoenix predicts likely actions for each viewer.
- The ranking scorer calculates a weighted score.
- Author diversity and network adjustments change the result.
- Cold-start logic can lift some smaller accounts.
- VMRanker can reorder posts to reduce repetition.
- Visibility filters can remove a post after ranking.
This matters because a high score cannot rescue content that fails an eligibility filter. X separates ranking from visibility. A post can be allowed to exist on the platform but still be excluded from recommendations to people who do not follow the author. (github.com)
What are the exact positive ranking weights?
The following weights are the current production defaults published in home-mixer/params/param.rs. They are the values worth understanding, not a claim that every viewer will experience the same outcome.
The values are published in X’s parameter file. The action categories are listed in the repository README, which also includes post clicks, profile clicks, quoted-post clicks, video-quality views, active seconds and negative actions. (github.com)
Likes are useful, but they are not the main prize
The like weight is 0.5. That is positive, but relatively small compared with replies, quote posts, DM shares, copy-link shares and follows.
That does not mean likes are worthless. A like still helps Phoenix learn what a viewer enjoys, and aggregate likes can improve retrieval and topic understanding. X’s official recommendation explanation says it uses likes, reposts, shares, replies, watched media and other signals to personalise recommendations. (help.x.com)
But a like is a low-friction action. People can tap it without changing their behaviour, opening a conversation or sending the post to someone else.
The practical conclusion is simple:
Do not write for applause alone. Write something people want to respond to, send, quote or save for later.
A post that earns 100 likes and no further action may be less valuable than one that earns 20 likes, five thoughtful replies and several direct shares. You cannot infer the final ranking score from the visible totals, but the weight structure makes the direction clear.
Replies and quote posts carry a strong positive coefficient
Replies and quote posts both have a weight of 5.0. That is ten times the like coefficient.
Again, this is not a guarantee that one reply equals ten likes. The model is multiplying the predicted probability of the viewer taking each action. The viewer’s personal history matters more than the raw coefficient suggests.
Still, the writing implication is obvious. Posts that create a reason to answer have more ranking potential than posts that merely announce something.
Good prompts for replies include:
- A specific disagreement with a common practice
- A decision between two credible options
- A small mistake with a useful lesson
- A clear opinion backed by an example
- A question that can be answered from experience
Bad prompts include “Thoughts?” added to the end of a finished advertisement. People can tell when a question is decorative.
Reposts are positive, but less valuable than conversation
The repost coefficient is 1.0. That is twice the like coefficient and lower than replies, quote posts and several sharing actions.
A repost still matters because it puts the post in front of a new audience. It also creates a new opportunity for other viewers to engage. But asking for reposts is not a complete strategy.
Make the post useful enough to pass along:
- A concise explanation of a confusing issue
- A checklist people can apply
- A strong observation about a fast-moving topic
- A result from a real experiment
- A sentence someone wishes they had written themselves
“Please repost” is weaker than giving people a reason to do it.
Copy-link shares have the highest positive action weight
The copy-link weight is 20.0, the largest positive coefficient among the standard interaction actions published in the file.
That is a fascinating signal. Copying a link usually means the viewer intends to take the post somewhere else. They may send it to a colleague, add it to a document or share it in a private group.
But it is also one of the easiest weights to misunderstand. You cannot see how many people copied a link from a public post, and the weight does not prove that every copied link creates a large reach boost.
The useful lesson is not “tell people to copy the link”. It is to create posts with utility outside the public feed.
Examples:
- “Send this to the person who owns your onboarding process.”
- “Keep this checklist for your next product launch.”
- “This is the five-minute explanation I wish I had before starting.”
- “Bookmark this before you hire your first growth contractor.”
Do not force those lines into every post. That quickly becomes obvious engagement bait.
Following the author is a high-value outcome
The follow-author weight is 4.0. A follow is a stronger signal than a like because it tells X that the viewer wants future content from the account.
This is where positioning matters. A post may earn attention without earning a follow if the viewer cannot work out what the account is about.
Your profile and content need to agree. If you post about product strategy on Monday, football on Tuesday and generic motivation on Wednesday, Phoenix may find individual posts interesting while struggling to identify the right audience for your next one.
A clear account promise improves the chance that a good post turns into a good follow.
Use a profile bio that answers three questions:
- Who is this account for?
- What does it consistently talk about?
- Why should someone follow rather than simply read one post?
You can use Xpert’s free bio generator to test different positioning angles, but the final version should sound like you, not a template.
Sharing by DM is more valuable than a normal share
The share-by-DM coefficient is 5.0. That matches replies and quote posts.
This signal reflects a private recommendation. Someone is telling a particular person that the post is relevant to them.
For creators, this means content aimed at a recognisable person or role can outperform content aimed at “everyone”. A founder might send a post to another founder. A developer might send a technical explanation to a teammate. A marketer might send a campaign teardown to a client.
Write with a real recipient in mind.
What do clicks, dwell time and video actions do?
The ranking code includes several passive and semi-passive attention signals.
They matter, but less in the simplistic way many creators assume.
Post clicks have a 0.4 weight
A post click means the viewer opens the post detail view. The coefficient is slightly below the like weight.
Opening a post can indicate interest, but it is not the same as a reply, share or follow. People often click to read a thread, inspect an image, view a quote or check the context before moving on.
Use a strong first line and enough curiosity to make the reader want the full post. Do not hide the useful information behind artificial suspense. If the click reveals nothing worthwhile, the next signal may be negative dwell or “not interested”.
External link opens have a 0.2 weight
The published weight for opening an external link is positive at 0.2. That is important because “X always penalises links” is too crude.
The algorithm code does not show a universal negative link coefficient. A link can contribute a positive predicted action. But posts containing links may still perform differently for other reasons:
- The reader leaves X quickly.
- The post gets fewer replies.
- The link reduces the space available for an argument.
- The destination looks promotional.
- The content is filtered or labelled by another system.
- The post attracts clicks from people unlikely to engage further.
So the correct advice is not “never post links”. It is:
Make the post valuable before the click, then use the link as the next step.
A bare link with a headline copied from a website is weak. A clear claim, useful context and an honest reason to click is much better.
When you need a repeatable workflow for ideas, hooks and drafts, Xpert’s AI hook generator can help you explore openings before you add the link.
Dwell has a small positive weight
The standard dwell weight is 0.05. Continuous dwell time has a separate weight of 0.004.
Those numbers look small, but dwell is not a binary event. A viewer can spend a meaningful amount of time reading a post, expanding a thread or watching media without liking it.
Dwell should be treated as a quality-control metric, not a writing target. You cannot reliably manipulate it by making every post longer. A long, unclear post may produce scrolling rather than reading.
Better ways to earn attention include:
- Put the tension in the first two lines.
- Use concrete nouns and verbs.
- Cut throat-clearing introductions.
- Break dense explanations into short paragraphs.
- Give the reader a reason to continue.
- End when the idea is complete.
Video and photo actions have modest weights
Photo expand is weighted at 0.05. Video open is weighted at 0.07. The action list also includes video-quality-view signals and active seconds.
This does not prove that video is favoured over text. It means X can predict several media-specific actions and include them in the score.
Video works when the opening earns attention quickly. A clip that takes six seconds to explain why it matters has already lost many viewers.
Use:
- A visible first frame with context
- Captions for sound-off viewing
- The payoff early
- A clear visual change
- A reason to watch beyond “here’s a video”
For practical video experiments, compare completion and active viewing against replies, shares and follows. Do not judge a clip on views alone. Xpert’s video engagement guide covers the creative side without pretending that raw views equal audience growth.
What negative signals hurt a post?
The published code includes five negative actions:
The numbers are large because these behaviours are relatively rare. X’s code comments say the report probability is more than 1,000 times lower than the like probability at baseline, so the coefficient is needed for the prediction to affect the final score. (github.com)
Do reports cancel hundreds of likes?
No.
The repository specifically warns against reading the values as raw count conversions. “One report cancels 468 likes” is not a valid interpretation of the ranking system. A report by a viewer who is unlikely to report content has a very different predicted probability from a report by someone who frequently reports similar posts. The effect is also personalised to viewers with similar behaviour. (github.com)
Mass-reporting schemes are not a reliable way to suppress a post either. X says an action needs to occur on a post served in the Home timeline to count in this recommendation system. Directly navigating to a post through a group chat does not manufacture the same ranking signal. (github.com)
Not interested is a bigger practical risk than a low like count
A low number of likes is not automatically a problem. Some useful posts are read, shared privately or followed without receiving much public engagement.
Repeated “not interested” feedback is different. It tells the system that the post is a poor match for that viewer.
This is why audience fit matters. A provocative post can receive replies and reposts while still accumulating negative feedback from viewers who find it irrelevant, repetitive or misleading.
The best growth strategy is not to please every viewer. It is to attract the right viewers and avoid making your intended audience feel that your account is noise.
Mutes and blocks are strong warnings
Mute and block weights are negative. They are also behavioural signals that can harm the account’s future distribution to similar viewers.
This does not mean you should avoid disagreement. Disagreement can generate replies and quotes. It means you should avoid the lazy forms of provocation:
- Personal attacks
- Repeated outrage bait
- Misleading screenshots
- Fake urgency
- Aggressive sales posts
- Posts that make the reader feel tricked
A good contrarian post attacks an idea, habit or assumption. It does not need to attack a person.
“Not dwelled” is a small but useful signal
The not-dwelled weight is only -0.02. On its own, it is not a giant penalty.
But it can accumulate across many viewers. If people repeatedly scroll past your posts without reading, Phoenix receives a consistent indication that your opening does not match the audience or that the content is not holding attention.
This is one reason you should not judge a post by its first few minutes. A post with a weak first line may underperform immediately, while a post with a specific, useful opening can continue earning attention after the initial burst.
Does X reward replies more than likes?
The code gives replies a coefficient of 5.0 and likes a coefficient of 0.5. So replies have a larger per-prediction coefficient.
That does not mean every reply is worth ten likes. The probabilities are different, the predictions are personalised and additional adjustments apply later.
The safer conclusion is this:
Conversation is a stronger ranking opportunity than passive approval, but only when the conversation is predicted to be genuine and relevant.
A reply-bait question is not magic. “Agree?” does not create a meaningful conversation by itself. A useful disagreement, a specific choice or an experience-based prompt gives people something to say.
The same principle applies when you reply to other accounts. A reply should add a new example, a correction, a useful resource or a sharp observation. “Great post” creates almost no reason for anyone to visit your profile.
X’s published code also includes a 15.0 bidirectional-follow reply boost. That parameter relates to replies between accounts that follow each other in both directions. It is not a universal multiplier applied to every reply. (github.com)
Does X penalise external links?
Not as a simple rule.
The open-source ranking weights include a positive 0.2 coefficient for opening a link. That means the model can treat a link open as a useful predicted action. The published code does not support the blanket claim that every external link receives an automatic negative score. (github.com)
However, links can still reduce performance if they weaken the post:
- The link is the entire post.
- The reader has no reason to trust the destination.
- The opening line gives away nothing.
- The post pushes the reader away before they interact.
- The content attracts low-intent clicks.
- The account posts links far more often than useful native content.
A strong link post usually has this structure:
- A claim or observation
- Evidence, context or a short example
- The link as supporting material
- A specific reason to click
If your link is the main product, write a native summary first. If people cannot understand the point without leaving X, your post is asking too much.
What are author diversity and out-of-network discounts?
The weighted score is adjusted after the action predictions are combined.
Author diversity prevents one account dominating a feed
X applies a decaying multiplier when several posts from the same author appear in the candidate pool. The public ranking code describes this as author diversity. The current parameters are a decay factor of 0.5 and a floor of 0.25. (github.com)
The resulting multipliers are:
This is not a penalty on your account. It is a diversity control for the viewer’s feed.
The practical lesson is to avoid publishing five near-identical posts in a short period. Even if one performs well, the next posts may face a lower effective score when competing for the same viewer’s attention.
Use a content calendar that varies the subject, format and angle. Xpert’s content calendar generator can help map one idea across several distinct posts without repeating the same argument.
Out-of-network content receives a discount
Posts from accounts the viewer does not follow are multiplied by an out-of-network factor below 1. The current standard factor is 0.75. Topic-scoped requests can use a lower 0.5 factor. The same adjustment can apply to replies and reposts from followed accounts under the relevant configuration. (github.com)
This sounds harsh for smaller accounts, but it explains why relevance must be obvious. An unfollowed account needs to overcome a distribution discount by creating a strong predicted match.
The answer is not to post more often. It is to make the topic, audience and value clear enough that retrieval finds the post for the right viewers.
Use specific language:
- “For solo SaaS founders hiring their first marketer”
- “A five-minute debugging method for TypeScript errors”
- “What agency owners should measure before raising prices”
Specificity helps both humans and recommendation systems understand where the post belongs.
VMRanker also reduces repetition
After scoring, X can use VMRanker to reorder results with a diversity-oriented system. The repository describes this as a determinantal point process over post embeddings, which gives up some score to reduce similarity between neighbouring posts. (github.com)
This is another reason not to copy the same post structure repeatedly. If your feed contains six posts that look and sound alike, the system may spread them apart even when their individual scores are strong.
Variety does not mean abandoning your niche. It means approaching the niche through different forms:
- A personal lesson
- A short tutorial
- A disagreement
- A case study
- A response to a current event
- A direct answer to a common question
Is there a boost for small or new accounts?
Yes, the published code contains cold-start logic.
The current parameters include:
- A follower cap of 1,000
- An impression threshold of 1,000
- Original posts only
- A maximum post age of 48 hours
- A target position around slots 15 to 16
- A low-impression position ratio of 0.85
An eligible original post from a smaller account can be lifted towards a target position if it already ranks within the eligible candidate pool. Replies and reposts are not eligible for this particular boost. (github.com)
This is not a free pass for a new account. The post still needs to be retrieved, pass filtering and rank well enough to qualify.
It does mean that starting from zero is not as hopeless as many creators assume. You do not need a huge following for every post to receive a chance at discovery. But the post must give the system enough information to identify its audience.
A vague post from a new account is still vague. A sharp, useful post can earn distribution beyond the account’s follower base.
What content gets filtered before ranking?
Some posts never reach the ranking stage.
The public repository lists filters for:
- Duplicate posts
- Missing post metadata
- Posts older than 48 hours
- The viewer’s own posts
- Unfollowed-account replies and reposts in specific contexts
- Muted keywords
- Blocked or muted authors
- Posts already shown or served
- Ineligible subscription posts
- Certain video requests
- Topic exclusions
- Some new-user low-engagement cases
Post-selection filters can then remove posts rejected by visibility filtering, posts with rejected parents or quoted content and duplicate branches of the same conversation. (github.com)
Visibility filtering is separate from ranking. X’s official recommendation policy says content may remain available to followers and on the author’s profile while being excluded from recommendations to people who do not follow the author. It also identifies spammy accounts, recently rule-breaking accounts and some harmful or marginally abusive content as ineligible for recommendation. (help.x.com)
This is why “the algorithm rewards replies” is incomplete advice. A post can generate replies and still fail to reach people outside your existing audience if another system decides it is not eligible for recommendation.
What actively hurts distribution?
The practical risks are not limited to negative clicks. Content can also suffer when it looks automated, repetitive, misleading or unsafe.
Avoid:
- Posting the same text across several accounts
- Copying viral posts with minor word changes
- Using irrelevant trending topics
- Repeating engagement-bait prompts
- Making claims you cannot support
- Stuffing posts with unrelated keywords
- Publishing a stream of thin AI-generated text
- Coordinating artificial engagement
- Ignoring replies after asking for discussion
X’s official recommendation policy specifically mentions spammy behaviour, harmful content and automated systems that may identify possible rule violations as reasons content or accounts may be kept out of recommendations. (help.x.com)
The safest growth strategy is boring in the best possible way: original ideas, clear audience fit, useful context and recognisable human judgement.
What should you optimise for in practice?
The ranking weights point towards five priorities.
1. Optimise for meaningful action, not raw engagement
Likes are easy to earn and easy to misread. Track replies, quote posts, shares, follows and link actions alongside likes.
A simple engagement rate can still help:
(likes + replies + reposts + quotes) ÷ impressions × 100
But use it as a health check, not a complete representation of the ranking score.
Benchmarks vary sharply by account size, audience and formula. One September 2026 dataset covering 108,694 accounts with more than one million followers found a median follower-based engagement rate of 0.112%, with the upper quartile at 0.544%. It also found a median of 0.321% for accounts in the 1,000 to 10,000 follower band. Those figures are useful for comparison, but they should not be treated as universal creator targets because the dataset is concentrated in large accounts. (playersells.com)
The metric that matters most is whether your own median is improving over time.
2. Give people a reason to reply
Do not finish every post with a question. That becomes predictable.
Instead, make the central idea incomplete without another person’s experience. Share a decision, trade-off or observation that naturally invites a response.
Weak:
What do you think?
Stronger:
We cut our onboarding from 12 steps to 5. Activation rose, but support tickets increased. The trade-off was worth it. Would you accept that exchange?
The second version gives people something concrete to discuss.
3. Make sharing useful
The highest positive coefficient is attached to copy-link sharing. You cannot see that metric publicly, but you can create content people want to pass along.
Write for a specific use case. Checklists, explanations, templates, comparisons and strong observations travel better than generic encouragement.
4. Make your topic legible
Out-of-network content has a discount, so your posts need to be easy to classify. Clear nouns help.
“Three lessons from our failed pricing test” is easier to place than “This changed everything.”
Your post does not need to sound robotic. It needs to make its subject obvious.
5. Reduce negative feedback
You do not need to become bland. You do need to stop relying on cheap provocation.
A strong opinion supported by evidence can earn discussion. A misleading claim may earn the same public engagement while increasing reports, mutes and blocks.
The goal is not universal approval. It is high relevance among the audience you want.
A practical 2026 posting framework
Use this sequence for your next 20 original posts.
Choose one audience
Do not write for “people interested in business”. Choose a narrower reader:
- Early-stage SaaS founders
- Freelance designers raising prices
- Developers learning AI tooling
- Marketing leads at small B2B companies
- Writers trying to build a consistent publishing habit
Choose one action
Decide whether the post should primarily earn:
- A reply
- A quote
- A follow
- A share
- A link click
- A longer read
Do not ask one post to do all six.
The opening should make the topic clear and create tension.
Examples:
Most onboarding advice ignores the part that causes users to leave.
I thought daily posting would grow our account. It mostly made our content worse.
The X algorithm does not reward replies as raw points. It predicts who is likely to reply.
Add evidence
Use a result, number, example, before-and-after comparison or direct observation.
Evidence creates a reason to read and gives replies something to engage with.
Add one natural next step
Invite a response only when a response would improve the post. Ask for a personal example, a choice between two options or a challenge to your conclusion.
Then reply to the people who take you up on it.
Review the post for negative signals
Before publishing, remove:
- Unsupported certainty
- Unnecessary insults
- Fake urgency
- Generic filler
- Repeated phrases from earlier posts
- A link with no context
- A question that exists only to bait replies
Use Xpert’s tweet grader or viral post checker as editing tools, not as promises of performance. A score cannot replace judgement, but it can reveal a weak hook, unclear structure or vague payoff.
What the X algorithm does not prove
The public code is valuable, but it does not answer every growth question.
It does not prove that:
- Replies always beat likes in every feed
- Premium subscribers receive a fixed reach multiplier
- Every post receives the same distribution window
- A particular posting time works for every audience
- Hashtags have no value in every context
- Links are always penalised
- A post with a high coefficient will go viral
- The published defaults never change between experiments
- The visible code represents every ranking system on X
X says that experiments run on a portion of timeline traffic, and some tunable values are read from configuration rather than hard-coded. The repository’s production defaults are useful, but they are not a permanent contract. (github.com)
The code also focuses on the For You feed. Search, notifications, profiles, Following and other surfaces can use different systems.
So use the weights as a map of what X is optimising for, not as an invitation to manufacture clicks.
A good post still needs to satisfy a human reader first.
FAQ
What does the X algorithm reward most in 2026?
The X algorithm rewards the predicted probability of several actions, not raw engagement counts. The strongest published positive coefficients include copy-link sharing at 20.0, replies at 5.0, quote posts at 5.0, DM shares at 5.0, follows at 4.0 and ordinary shares at 2.0. These weights apply to personalised predictions, so they are not direct count equivalents. (github.com)
Are replies worth more than likes on X?
Replies have a published coefficient of 5.0, while likes have a coefficient of 0.5. That makes replies more influential in the weighted score when the predicted probabilities are comparable. It does not mean one reply always equals ten likes, because Phoenix predicts each action separately for each viewer.
Does X penalise links in 2026?
The published ranking code gives opening an external link a positive weight of 0.2. That means there is no simple universal negative link coefficient in the public ranking weights. Links can still perform poorly when the post provides no native value, earns little conversation or sends readers away before they interact.
Is one report worth 468 likes on X?
No. That is a misreading of the published weights. The report coefficient is -234.0 and the like coefficient is 0.5, but those values multiply predicted probabilities, not raw counts. X’s own code comments say that interpreting them as “one report cancels 468 likes” is incorrect. (github.com)
Does X reward copy-link shares?
Copy-link shares have the highest positive standard action coefficient in the published parameter file, at 20.0. That does not prove that every copied link creates a large reach boost, and the metric is not publicly visible in the same way as likes or replies. The practical takeaway is to create content people want to send elsewhere.
Does X boost small accounts?
The published code includes cold-start logic for eligible low-impression authors. The current parameters include a 1,000-follower cap, a 1,000-impression threshold, a 48-hour maximum post age and a target position around slots 15 to 16. The boost applies to eligible original posts, not every reply or repost. (github.com)
Does X reward Premium accounts?
The public For You ranking code does not provide a simple universal Premium reach multiplier. Subscription status appears in candidate metadata, but the published ranking weights do not establish that Premium automatically multiplies every post’s score. Treat claims about a fixed Premium boost cautiously unless X publishes a specific parameter or announcement.
The public ranking weights do not list a hashtag multiplier. Hashtags can still help with topic clarity, search and event discovery, but adding irrelevant hashtags does not create a guaranteed ranking boost. Use them when they genuinely identify the conversation.
How often should you post on X in 2026?
There is no single frequency that the published ranking code recommends. Posting too many similar updates can run into author-diversity decay, while posting too rarely gives you fewer chances to learn what your audience responds to. Start with one or two strong original posts per day, add useful replies and review your median performance over four weeks.
What is the best metric for measuring X growth?
Use a group of metrics rather than one number. Track impressions, replies, quote posts, reposts, follows, link clicks and your engagement rate by impressions. Watch the median across a set of posts because one viral post can distort the average. The most useful signal is whether the right people are taking meaningful actions more often.
Is there a video explaining the 2026 X algorithm?
This video gives a visual explanation of the 2026 For You ranking system and the Phoenix model:
It is useful as an introduction, but the GitHub repository remains the better source for exact weights and implementation details. (youtube.com)
Final takeaway
The X algorithm is not rewarding “engagement” as one big bucket.
It predicts likes, replies, reposts, quote posts, shares, link actions, follows, attention and negative feedback for individual viewers. It combines those predictions with published weights, then adjusts for author repetition, network distance, cold-start conditions, content eligibility and feed diversity.
The best strategy in 2026 is not to chase the largest visible number. It is to create posts that a specific audience is likely to read, respond to, share or follow from.
Write for one person. Make the value clear quickly. Give people a genuine reason to act. Avoid content that attracts attention from the wrong audience. Then measure meaningful actions over enough posts to see a pattern.
That is less exciting than a magic multiplier.
It is also much closer to how the system actually works.
