X Algorithm 2026: How the Grok Model Ranks

X replaced its old open-sourced recommendation pipeline in January 2026. The candidate-sourcing, scoring, and filtering stages gave way to a single large transformer model, the same family that powers Grok, that reads the actual content of every post and video rather than leaning mainly on engagement counts.

The signals that mattered before still matter, but the model now interprets them in context. Replies and conversations carry the most weight, the first 30 minutes after you post decide most of your reach, and external links in the post body are quietly suppressed for non-Premium accounts.

If you have wondered how does the Twitter algorithm work now that the old pipeline is gone, the short version is this: the model reads your content and watches how fast people engage. Understanding it turns guesswork into a concrete playbook, so you can stop optimizing for last year’s pipeline and start optimizing for the signals the Grok model actually rewards now.

What Changed in 2026: The Grok Transformer Rewrite

In short: In January 2026 X retired its old three-stage recommendation pipeline and replaced it with a single Grok-powered transformer model that reads the content of every post and makes roughly 5 billion ranking decisions a day.

For years X ranked posts through a pipeline of discrete stages: candidate sourcing pulled a few thousand posts, a heavy ranking model scored them, and filters stripped out unwanted content. Much of that pipeline was open-sourced in 2023, which is why the old engagement weights are so well documented.

That system is gone. The new model is a large transformer, from the same family as Grok, that reads the actual text, images, and video of each post instead of relying mostly on engagement counts and follow graphs. It evaluates relevance the way a language model evaluates meaning, so a post is now judged on what it says, not just on who engaged with it.

The scale is the headline number: X engineers describe about 5 billion ranking decisions per day, each returned in under 1.5 seconds. The practical effect is that distribution is more content-aware and less gameable with mechanical tricks. Posts that read as genuinely engaging to a model trained on language tend to travel further than posts stuffed with keywords or engagement bait.

The takeaway for a creator is not to chase the model’s internals. It is that quality of the post itself, and the conversation it sparks, now carry more weight than ever.

What's the Difference Between the For You and Following Feeds?

In short: The Following feed shows posts from accounts you follow in reverse chronological order. The For You feed is the ranked one, and it mixes roughly half in-network and half out-of-network posts the model thinks you will engage with.

Following is simple: every post from the accounts you follow, newest first, no ranking. It is the predictable feed, and it is where your existing followers are most likely to see you in order.

For You is where the model earns its keep. It is the default tab for most users, and it blends posts from accounts you follow with posts from accounts you do not. The long-standing split is close to 50/50 between in-network and out-of-network content, which is what makes reach beyond your follower count possible at all.

That out-of-network half is the opportunity. A reply you leave on a larger account’s thread can surface in the For You feeds of people who have never heard of you, because the model is actively looking for relevant out-of-network content to show. This is why the reply guy strategy is one of the most useful questions a creator can act on: replies are the cheapest way into the out-of-network half of someone else’s feed.

Treat the two feeds as different jobs. Original posts hold your existing audience in the Following feed. Replies and conversation are how you break into For You feeds you do not own.

How Does the Twitter Algorithm Weight Replies vs Likes?

In short: Replies and conversations dominate. Two framings circulate: the open-source formula that scored replies at 13.5 versus likes at 1, and the widely cited rule of thumb that a reply is worth about 27x a like and a full conversation about 150x.

The original open-sourced ranking model used explicit multipliers, and they are worth knowing because they reveal the model’s priorities: likes counted as 1, retweets as 20, replies as 13.5, profile clicks as 12, link clicks as 11, and bookmarks as 10. Even in the formula, a reply was worth more than thirteen likes.

The other framing you will see, that a reply is worth roughly 27x a like and a back-and-forth conversation roughly 150x, comes from creators and analysts measuring real reach outcomes rather than reading the source code. The numbers differ because they measure different things: one is a raw scoring multiplier, the other is observed distribution after the full model and feedback loops do their work.

Both point the same direction. The single most valuable thing you can produce is a reply that earns a reply back, because a conversation signals to the model that your post is keeping two people on the platform, not just earning a passive tap. Retweets and bookmarks are strong secondary signals; likes are the weakest of the bunch.

The strategic read is blunt: broadcasting earns likes, conversation earns reach. If you only have 20 minutes a day, spend them starting conversations, not polishing a single post nobody replies to.

Engagement Velocity: Why the First 30 Minutes Decide Everything

In short: Early engagement velocity is the strongest single signal. The model watches how fast a post earns interaction in the first 30 to 60 minutes, and a small burst, often around 10 meaningful engagements, can trigger wider amplification.

The model does not wait days to decide whether a post deserves reach. It samples how a post performs in a small initial audience, then decides whether to widen distribution based on how fast that audience engages. A post that earns ten replies in ten minutes looks very different from one that earns ten replies over two days.

This is why timing is not a vanity metric. Posting when your audience is awake and active compresses your engagement into the window the model is watching. Our best time to post tool exists for exactly this reason: getting the first wave of replies fast is more valuable than the same replies spread thin.

Velocity also reframes what a reply is worth. A reply you leave in the first few minutes of someone else’s post, while their thread is still hot, rides their velocity instead of fighting your own cold start. Early replies on active threads are some of the highest-leverage actions on the platform.

The hard part is being there in the first 30 minutes, consistently, without sitting in the app all day. That tension between what the model rewards and what a busy founder can sustain is the real problem worth solving.

How Much Does X Premium Boost Your Reach?

In short: A Premium subscription gives verified accounts a real distribution advantage, often described as a 4x to 8x boost in reply ranking and feed prioritization. It is a multiplier, not a substitute for engagement.

Premium accounts get their replies ranked higher in conversations and their posts prioritized in the For You feed. Estimates of the size vary, but a 4x to 8x effective reach advantage over an otherwise identical non-Premium account is a common figure.

It matters most in two places: replies in busy threads, where a Premium reply sits closer to the top where people actually read, and the recovery of link reach, since the link suppression that hits non-Premium accounts is far lighter for subscribers.

It is a multiplier, not a foundation. A Premium account that posts nothing worth replying to still gets little reach. The honest framing is that Premium amplifies whatever engagement you are already earning, so it pays off most for accounts that are already replying and posting consistently.

What Kills Your Reach on X?

In short: External links in the post body, slow early engagement, coordinated engagement pods, and content the model reads as low quality all suppress reach. Hashtags no longer help, and as of March 2026 external links are penalized harder than ever for non-Premium accounts.

Links are the big one. X has long down-ranked posts with external links in the body because they pull users off the platform, and in early 2026 that suppression tightened further: non-Premium posts with links now see close to zero median engagement. The standard workaround is to put the link in a reply to your own post rather than in the post itself.

Hashtags are effectively dead as a reach tactic. Because the Grok model reads and categorizes content semantically, it already understands what your post is about without tags. Stuffing hashtags signals low-effort content and does nothing to widen distribution.

Engagement pods and reply-for-reply rings are risky. The model is content-aware enough to notice engagement that does not match the post, and coordinated inauthentic activity can suppress an account rather than boost it. Slow early engagement, high block or mute rates, and posting into dead hours round out the list of reach killers.

If your reach dropped and you are not sure why, audit your own patterns first. You can analyze your tweets to spot whether links, timing, or thin engagement are the culprit before you change your whole approach.

The Reply Advantage: A Practical Playbook for Busy Founders

In short: Reply early and often, in your own voice, on active threads in your niche. The algorithm rewards exactly this, and it is also the single hardest thing to sustain when you are building a product instead of living on X.

Everything above points to one move: be the person who shows up early in other people’s threads with a reply worth replying to. That captures the conversation weighting, the out-of-network half of the For You feed, and the first-30-minutes velocity window all at once. The reply guy strategy is the proven version of this, and it works because it is aligned with how the model actually ranks.

The catch is volume and timing. Doing this well means landing maybe a dozen or more thoughtful replies a day, fast, on the right threads, every day, without it eating your morning. Reply to 50 posts a day sounds great until you try to do it by hand around shipping a product.

This is the honest connection to what we build. XreplyAI is trained on your own post archive, so it drafts reply in your own voice that you can review and send in seconds, not minutes. The point is not to automate your presence away. It is to make the fast, voice-matched, early replies the algorithm rewards something you can actually keep up.

So the playbook is simple to state and hard to do alone: pick a handful of accounts in your niche, reply early while their threads are hot, keep your links out of post bodies, and show up daily. The model rewards consistency and conversation. Sounding like you, not like everyone else using AI, is what keeps those replies worth reading.

The 2026 X algorithm is a content-aware Grok model, but the strategy it rewards is the same one creators have always suspected: start conversations, reply early, keep links out of your post bodies, and show up every day. The model reads what you say and watches how fast people respond, so quality and velocity beat volume and tricks.

The honest problem is that the behavior the algorithm rewards is exactly what eats a founder’s day. XreplyAI is trained on your own post archive and drafts replies that sound like you, not like everyone else using AI, so you can land fast, voice-matched replies in the first 30 minutes without living on X. Try it free and see how many more conversations you can start per day.

• X (Twitter) open-source algorithm repository: github.com/twitter/the-algorithm
• Sprout Social, How the Twitter/X algorithm works in 2026: sproutsocial.com/insights/twitter-algorithm
• OpenTweet, How the Twitter/X algorithm works in 2026: opentweet.io/blog/how-twitter-x-algorithm-works-2026