How the X (Twitter) Algorithm Works in 2026 (Exact Weights)

On January 20, 2026, X open-sourced its full Grok-powered feed algorithm on GitHub, the first complete code drop since the 2023 release. That matters, because most of the advice still floating around was written for the old stacked recommendation system, and that engine no longer exists. If you are wondering how the X (Twitter) algorithm works now, you can actually read the ranking logic instead of guessing.

Here is the short version of what the code and the published weights confirm: the new model rewards conversation over volume. Replies are weighted far heavier than likes, native video crushes text-only reach, and external links are throttled to near-zero for non-Premium accounts. Posting more is not the lever anymore. Talking to people is.

Most of this is good news for small accounts. The old system leaned on raw engagement counts, which favored people who already had reach. The new model reads meaning, so a genuinely good reply from a 200-follower account can land in front of thousands. The lever changed, and it changed in your favor.

What the X Algorithm Actually Is in 2026 (Grok-Powered)

In short: The X algorithm in 2026 is one Grok-powered transformer that reads the full text and video of every post, then ranks candidates for your feed in under 1.5 seconds, instead of the stacked model X ran before.

The old system was a stack of separate models: one for in-network tweets, one for out-of-network, a heavy-recommendation ranker bolted on top. X open-sourced that version in 2023, and most algorithm guides still describe it. That stack is gone.

What replaced it is a Grok-based language model that actually comprehends content. It reads your full post, transcribes and watches attached video, and scores relevance against each viewer instead of leaning mostly on engagement-count proxies. X has said the system makes around 5 billion ranking decisions per day.

The January 2026 GitHub release made the structure public. The code shows the candidate pipeline and the relative weighting between engagement actions, which is why the weight estimates later in this guide are now grounded in published logic rather than pure inference. X still does not ship a printed table of exact multipliers, but the direction of the weighting is no longer a secret.

Candidate sourcing still pulls from three pools: people you follow (in-network), accounts you do not follow but that the graph suggests (out-of-network), and an interest graph that maps topics to your behavior. The model assembles roughly 1,500 candidate posts per refresh, then ranks them for relevance. For a small account, the out-of-network pool is where growth lives, because it is the path to people who do not follow you yet.

The practical shift: because Grok reads meaning, low-effort engagement bait and keyword stuffing work worse than they used to, and genuinely useful or interesting posts surface more reliably even from small accounts.

The Ranking Weights X Published: Every Signal, Ranked

In short: Reach is driven mostly by engagement weights, and replies count far more than likes: relative to a like, a reply is worth roughly 27x, and a reply the author answers can be worth around 150x.

Not all engagement is equal. The model assigns each action a relative weight, and those weights decide whether a post escapes your followers and spreads to out-of-network feeds. Likes are the weakest meaningful signal. Conversation is the strongest.

Here are the published, relative 2026 engagement weights, expressed against a single like as the baseline:

One caveat worth stating plainly: these are relative weights drawn from the open-sourced code and third-party analysis, not a precise multiplier table X stamped as official. Treat them as a directional ranking of what matters, not gospel decimals. The order is reliable; the exact numbers are estimates.

Read that table closely, because it rewrites the playbook. A reply chain you take part in is worth more than a hundred likes. Bookmarks and profile clicks, both signals that someone found you worth saving or investigating, outrank a retweet on a per-action basis.

The takeaway for a small account: chasing likes is the lowest-leverage thing you can do. Starting and continuing conversations is the highest. To see which of your posts already pull the heavy signals, analyze your tweets with our free tool.

Replies vs Original Posts: What the Algorithm Favors

In short: For growth, the algorithm favors replies. A reply carries roughly 27x the weight of a like and lands in front of someone else's audience, while an original post starts inside your own follower ceiling. Original posts still matter: they anchor your profile and host the conversations that generate the 150x signal.

The reach math explains why replies punch above their weight. An original post from a 300-follower account competes inside a 300-person pool until engagement pushes it out-of-network. A reply on a 50,000-follower account's post starts in a 50,000-person room, and every author response multiplies the signal. We ran the numbers in does replying boost reach on X.

Because a reply that earns an author response is worth so much, the people you reply to matter more than how often you post. One specific reply on an account whose audience overlaps yours can outperform a week of posting to your own followers. Pick your targets, then make the reply good enough to earn a response.

The weights also reframe what a "viral" post is. They compound: a post that earns 50 replies, half of which you answer, feeds the model far more signal than a post that earns 500 silent likes. The conversation post will travel further out-of-network, even though the like count looks smaller. Optimize for conversation, not the vanity number.

Original posts still do a job replies cannot. Your profile is what reply-driven visitors judge before they follow, and a profile with no recent posts converts nobody. Post at least once a day so the profile clicks your replies earn have somewhere to land, then spend the rest of your time in other people's threads.

The First 60 Minutes: Why Early Engagement Decides Everything

In short: The first 30 to 60 minutes after you post is the strongest timing signal, and visibility decays fast: a post loses roughly half its reach potential within about 6 hours.

The model treats early engagement as a quality vote. If a post earns replies and reposts quickly, it gets pushed to larger out-of-network audiences. If it lands flat in the first hour, it rarely recovers, no matter how good it is. Engagement velocity, not total engagement, is what the early window measures: ten replies in the first ten minutes beats a hundred spread over two days.

Timing matters because of that early window. You want to post when the people most likely to engage with you are actually online, so the first-hour signal fires before decay sets in. We break the data down in our guide to the best time to post on X, and the free best time to post tool shows your own peak windows.

Decay is steep after that. Half of a typical post’s visibility is gone within roughly 6 hours, and most posts are effectively dead within a day. Evergreen content does not slowly accrue reach on X the way it does on YouTube or a blog. The practical implication: you cannot rescue a post that flopped, so stop refreshing it and put that energy into the next one.

For a time-poor founder, the lesson is to concentrate effort: post once when your audience is active, then spend the next hour replying to the conversation it starts rather than scattering posts across the day.

The first hour also rewards your own replies. When you answer the people who comment, you extend the conversation inside that critical window, and the model reads the back-and-forth as a quality signal worth amplifying. Being present for the first hour after you post is often more valuable than the post itself.

What Does the Algorithm Penalize in 2026?

In short: The biggest 2026 penalty is external links, which can cut a non-Premium post’s reach by 50 to 90 percent. Engagement bait, combative tone, generic AI-slop replies, and rapid follow/unfollow are also suppressed.

The link penalty is the change that catches founders off guard. X has long disliked sending people off-platform, and as of July 2026 the throttle is firmly in place: a post containing an external link gets its reach cut by an estimated 50 to 90 percent unless you have Premium. This is based on observed reach data, not a printed X policy, but the pattern is consistent enough that founders feel it. Share your launch with a raw link and barely anyone sees it.

The survival playbook is simple. Post the value natively (the thread, the screenshot, the takeaway) and drop the link in your first reply instead of the main post. The main post carries the reach, the reply carries the click. A screenshot of the page with the URL typed out also dodges the link detector.

Tone is now a signal too, and this is new with the Grok-powered model. Because the ranker reads meaning, it can register combative, dunking, or rage-bait phrasing and dampen the reach of posts that read as hostile rather than constructive. You do not have to be bland, but a thread that picks a genuine fight will not travel as far as one that adds something. The model is reading for quality conversation, not for heat.

Engagement bait gets penalized too. Because Grok reads meaning, "reply YES to get my free guide" and follow-for-follow loops are detectable and down-ranked rather than rewarded. Rapid follow/unfollow churn trips spam heuristics and can suppress your whole account.

AI-slop replies are a specific and growing risk. Generic, off-topic, clearly-templated AI replies are exactly the pattern the model is tuned to mute, because they look like the bait it down-ranks. A reply that does not reference the actual post reads as spam to a model that comprehends the post. This is the trap most reply-automation tools walk users straight into: volume of generic replies suppresses reach instead of building it.

Inconsistency is a softer penalty. The interest graph needs recent signal to know who to show you to. Go quiet for weeks and you effectively restart from a colder position when you return.

If your reach has collapsed across the board rather than on one post, you may be shadow-banned rather than just penalized. Run a shadow ban check before you blame the algorithm, because the fix for a suppression flag is different from the fix for a weak post.

One more trap: deleting and reposting to fix a typo. Each fresh post starts the first-hour clock over from zero, and the original’s early signal is wasted. Edit if you have the feature, or live with the typo. The reach you lose to a delete-and-repost is almost never worth the fix.

Does Liking (or Replying to) Your Own Tweet Help?

In short: Liking your own tweet does nothing: the algorithm discounts self-engagement, so a self-like adds no ranking signal. Replying to your own tweet can help, but only when the reply adds real content, because the model reads and scores it like any other post. Filler self-replies read as spam.

Self-likes are the easy call. The ranking system weights engagement by what it predicts other viewers will do, and your own like predicts nothing about anyone else. It will not hurt the post, it just does not count. The same goes for liking your post from an alt account: coordinated self-engagement is exactly the pattern spam heuristics watch for.

Self-replies are more useful than most people assume. A reply that adds substance, such as the link you kept out of the main post, extra context, or a follow-up point, extends the thread and gives the model more content to score. That is why the link-in-first-reply play from the penalty section works: a self-reply is a legitimate post in its own right.

The line is substance. A "bump" or "check this out" self-reply gives the model nothing to read, and content-free filler pattern-matches the engagement bait the ranker down-ranks. If a self-reply would look pointless to a human, it looks worse to a model that comprehends the post.

Self-reposting follows the same substance rule: a self-repost can re-expose a post to followers who missed it, but it earns no engagement weight by itself. We break down when it helps and when it backfires in our guide to reposting your own tweets.

Premium, Native Video, and the Boosts That Actually Move Reach

In short: Two levers reliably increase distribution in 2026: X Premium, which adds an estimated 2x to 4x reach and exempts you from the link penalty, and native video, which the model now watches and favors over text and linked clips.

Most of growing on X is free and earned. But two paid or format-based levers genuinely change how the algorithm treats you, and it is worth being honest about both.

The first is X Premium. X has confirmed that subscribers get a distribution boost, and third-party reach analysis pegs it at roughly 2x to 4x a comparable non-Premium account, with some accounts reporting higher in replies specifically. Premium also exempts you from the worst of the external-link penalty, which alone can justify the cost for anyone who links out regularly. It is the one paid setting that materially changes the math.

Premium is not a cheat code, though. It multiplies the reach of content that already earns engagement; it does not rescue posts the model ignores. A boosted bad post is still a bad post. Treat it as an amplifier on a working system, not a substitute for one.

The second lever is native video. Because the Grok-powered model now watches and transcribes video, clips uploaded directly to X get a strong initial distribution edge, often several times the reach of an equivalent text post. The key word is native: a YouTube link gets the link penalty, while the same clip uploaded straight to X gets the boost. Short and watchable wins, with clips in the sub-60-second to roughly two-minute range performing best because more of the audience finishes them, and completion is a quality signal the model reads.

The combination founders sleep on: a short native video, posted at peak time, with any link in the first reply. That stacks three favorable signals (format, timing, no link penalty) on one post. You do not need a studio. A 40-second screen recording with a clear point clears the bar.

What’s the Highest-ROI Move in 2026? Replies, Not More Posts

In short: Because replies carry roughly 27x the weight of a like and reply chains up to 150x, thoughtful daily replies on other people’s posts are the cheapest, fastest way to grow reach in 2026.

Put the engagement table together with the link penalty and one strategy falls out. The algorithm now pays you in reach for conversation, and a good reply puts you in front of someone else’s audience without needing your own to be large yet. This is the foundation of the reply guy strategy.

A reply that earns a response from the original author can be worth 150x a like. A handful of those a day, on accounts whose audience overlaps yours, compounds faster than posting into a void ever will. You are borrowing the reach of accounts bigger than you, one good comment at a time.

The catch is voice. Generic AI replies read as generic AI replies, and as the penalty section explained, the model is now tuned to mute them. The reach only comes when the reply is specific, considered, and sounds like a real person who actually read the post.

The fix is specificity. Reference the actual post, add something the author did not say, and write it the way you talk. If you use AI to draft, it has to work from your writing: XreplyAI's optional reply assist, on paid plans, drafts from your own post archive so the draft reads like you instead of like a template. You review and post; the reach still comes from being in real conversations.

How to Work With the Algorithm If You Only Have 20 Minutes a Day

In short: Spend most of your 20 minutes replying to 8 to 12 relevant posts, post one native (link-free) update at your peak time, and put any link in the first reply.

You do not need to live on X to grow on it. You need to spend a short, focused block on the highest-weighted actions and ignore the rest. Here is a 20-minute routine built around the 2026 weights.

First, ten minutes on replies. Open posts from 8 to 12 accounts whose audience overlaps yours and leave a specific, useful reply on each. This is where 80 percent of your reach return comes from, because replies are the heaviest signal. The bar is that each reply references the actual post: that is what separates a reach-building comment from the AI slop the model mutes.

Next, five minutes on one post. Publish a single native update (a take, a lesson, a short clip) with no external link in the body. If you need to link out, put the URL in the first reply. Time it for when your audience is online so the first-hour signal fires.

Last, five minutes circling back. Reply to anyone who replied to you, because an author response is the 150x signal, and it tells the model your post sparked real conversation.

Consistency is the part founders drop first, and the routine only compounds if it happens daily. XreplyAI removes the setup cost: schedule posts across 14 platforms from one calendar and your native post goes out at your peak time whether or not you are at your desk, leaving the 20 minutes for the reply blocks. On paid plans, an optional AI reply assist drafts from your own post archive so the ten reply drafts stay specific instead of generic.

Done daily, that routine compounds. You are not gaming anything; you are spending a fixed 20 minutes on the actions the algorithm weighs heaviest instead of the ones it barely counts. Over a few weeks the interest graph learns who you are and starts surfacing you to the right out-of-network audiences without you doing anything extra. If you want to skip the manual reply grind entirely, browse our free X/Twitter tools to speed up the rest.

If you take one thing from how the X (Twitter) algorithm works in 2026, make it this: it rewards real conversation above all. The open-sourced code confirms what the weights imply. Replies outweigh likes by a wide margin, links are throttled, combative and generic-AI content gets muted, and posting volume decays in hours. The growth move that actually works now is showing up in replies daily, in your own voice, on posts whose audience overlaps yours.

If you want to run that playbook without living on X, try XreplyAI free. Schedule your posts across 14 platforms from one calendar so the timing signal fires every day, and on paid plans the optional reply assist, which drafts from your own post archive, speeds up the daily reply block.

• X (Twitter) open-sourced its feed algorithm on GitHub (January 20, 2026 release)
• Sprout Social: How the Twitter/X algorithm works (published engagement weights)