Most "how the algorithm works" guides recycle a 2023 diagram and call it a day. The real thing is weirder. X open-sourced its ranker in 2023, replaced the heuristic scorer with Grok in late 2025, and still runs the public weight table underneath (Social Media Today, 2025). That means two things — the old weights still matter, and semantic relevance matters on top.
If you want to stop guessing why some posts get 10,000 impressions and others get 80, you have to read the actual signals. This post walks through every one — the positive weights, the negative multipliers, what Grok layered on top, and which signals you can move.
Key Takeaways
- A reply that triggers an author response carries +75 in the ranker — 150x the weight of a like (+0.5) (Social Media Today, 2025).
- Grok replaced the heuristic ranker in late 2025, adding semantic matching on top of the open-source weight table.
- Free-account posts with outbound links saw a 1,700% reach lift when the link was removed in A/B tests (hashmeta, 2025, citing Buffer data).
- Negative signals are stronger than positive ones — a block multiplies score by -3.0, a report by -1.5 (X Engineering, 2023).
What Is the X Algorithm in 2026?
The X algorithm in 2026 is a two-layer system — the open-source Heavy Ranker from 2023, with a Grok transformer layered on top as of late 2025 (Social Media Today, 2025). The Heavy Ranker still calculates a score per candidate post. Grok decides which posts become candidates and how semantically relevant each one is to you.
Here's the flow. When you open the "For You" tab, X pulls about 1,500 candidate posts from your network and beyond. Grok now handles candidate selection — it reads the content of every post and video, matches it to your semantic interest profile, and forwards the top slice to the Heavy Ranker. The Heavy Ranker then assigns each candidate a score based on predicted engagement probability.
The full source code of the ranker sits in the twitter/the-algorithm GitHub repo, open-sourced in March 2023 (X Engineering, 2023). The core scoring formula from that drop looks like this:
Score = 0.5 × P(Like)
+ 1.0 × P(Retweet)
+ 0.3 × P(Reply)
+ 0.15 × P(Profile_Click)
- 1.5 × P(Report)
- 3.0 × P(Block)
Each P(Event) is the model's predicted probability that you'll perform that action on the post (X Engineering, 2023).
Three things this formula tells you. Retweets are weighted 2x higher than likes in the prediction layer. A predicted block costs six times what a predicted like rewards. And profile clicks — a signal most creators ignore — show up at 15% of a like's positive weight, meaning a strong bio and profile picture matter more than most people think.
Grok doesn't replace this formula. It feeds it better candidates. The math that wins still runs underneath.
What Are the Exact Ranking Weights in the X Algorithm?
The public weight table, published when X open-sourced the algorithm, scores individual engagement signals on a different scale than the Heavy Ranker's probability formula. A like is +0.5. A reply that triggers an author response is +75 — a 150x multiplier (Social Media Today, 2025). These weights reward the signals X believes indicate genuine interest, and they punish everything performative.
Here's the full table, sorted from lowest to highest weight:
Signal Algorithmic weight Video 50% completion +0.005 Like +0.5 Retweet +1.0 Dwell time (content read) +10 Conversation click +11 Profile interaction +12 Direct reply +13.5 Reply with author response +75 Source: X open-source algorithm weights via Social Media Today (2025).
A few non-obvious takeaways from this table.
Dwell time is worth 20 likes. If a reader pauses on your post for more than a couple of seconds, that single action moves your ranker score more than 20 likes combined. Long-form posts (the 4,000-character Premium feature) print dwell time by design — they average 44.87 engagements per post versus 15.28 for threads in Metricool's 2.1-million-post dataset (Metricool, 2024).
Video completion is nearly free weight. A video watched to the halfway point is worth +0.005 — effectively zero. Don't over-invest in video unless it drives profile clicks or replies downstream.
Profile interaction (+12) rewards the unglamorous work. When someone clicks your name to open your profile, the ranker treats that as intent-level signal. Pin a post that converts and you're effectively banking +12 every time a reply of yours triggers a profile click.
The asymmetry is the lesson. Most creators optimize for likes because they're the visible vanity metric. The ranker barely cares. It cares about dwell, clicks, and conversations — the signals that are harder to fake.
- Discover the 100-reply-a-day strategy that actually works
Why Are Replies Weighted 150x More Than Likes?
A reply that sparks an author response is weighted +75 in the ranker. A like is weighted +0.5 — a 150x gap (Social Media Today, 2025). X treats reply-with-response as the strongest proxy for "this content caused a real conversation," and conversations are what keeps users on-platform.
The math gets brutal when you apply it to a growth scenario. If you post to 500 followers and the post earns 10 likes, you've banked +5 in ranker weight. If instead you reply to a mid-sized account and they respond, you've banked +75 on that single reply — and your reply now appears to both audiences, not just your 500 followers.
Why does X weight it this way? Two reasons:
- Reply-with-response is statistically rare. Most replies get ignored. When an author actually responds, the ranker treats it as a quality signal the way search engines treat editorial backlinks.
- It generates dwell time for both users. The original author reads your reply, types a response, and their followers read the exchange. That's compounded attention, which is X's core business metric.
Why generic replies never get responses: The +75 weight only triggers when the post author replies. Authors don't reply to "great thread!" They reply to specific pushback, a contrarian data point, or a question that extends their thinking. Optimize for the response, not the reply.
The practical consequence is reply leverage. A 30-reply daily habit targeting mid-sized accounts (10K-100K followers) outperforms a 3-post daily habit for any account under 5,000 followers. The ranker math doesn't let you skip it.
- Discover: (Reply templates vs AI replies - what converts)[/blogs/reply-templates-vs-ai-replies-what-converts]
How Does Grok Score Semantic Relevance?
Grok's job is candidate selection, not final ranking. In late 2025, Elon Musk announced that Grok would "read every post and watch every video" to match users to content (Social Media Today, 2025). In practice, that means Grok builds a semantic embedding of your interest profile and scores candidate posts against it before they reach the Heavy Ranker.
What changed for creators:
- Keyword matching got weaker. Old X indexed posts by exact tokens. Grok indexes by meaning. A post about "serverless cold starts" surfaces to someone who reads about Cloudflare Workers, Vercel functions, and edge compute — even if the exact phrase never appeared in their reading history.
- SimClusters now drive distribution. Grok groups users into "SimClusters" — communities that discuss overlapping topics (Sprout Social, 2026). Your post gets routed into the SimClusters whose semantic fingerprint matches yours, not to random followers.
- Niche language pays off. Writing "bootstrapped SaaS founder under $10K MRR shipping solo" gives Grok enough semantic signal to route the post precisely. Writing "entrepreneur tips" routes it nowhere.
Does Grok override the Heavy Ranker weights? No. It filters which posts get scored. The weights still decide who wins among the candidates Grok forwards. This is why a niche account with 500 followers can now outperform a generic account with 50,000 — if the niche post matches a SimCluster tightly and generates replies, the ranker math does the rest.
One tactical implication: generic "growth tips" content is semantically ambiguous and lands in no SimCluster cleanly. It may have worked in 2023. In 2026, it underperforms anything specific by an order of magnitude.
How Often Does X Update the Algorithm?
X has made at least four major ranker changes since open-sourcing the code in 2023 — the biggest being the Grok transformer swap in late 2025 (Social Media Today, 2025). Smaller adjustments happen constantly. Weight rebalancing, signal additions, and penalty thresholds all shift without announcement.
The changes you can track from public sources:
- March 2025: Free-account text posts dropped to near-zero median engagement. Premium+ held at ~0.9% (Buffer, 2025).
- Late 2025: Grok transformer replaced the heuristic candidate selector. Semantic matching became a dominant signal (Social Media Today, 2025).
- Early 2026: Link suppression intensified. A/B tests of identical tweets with and without a link showed a 1,700% reach difference in favor of the link-free version (hashmeta, 2025, citing Buffer data).
There's a second layer of change that nobody sees — the model retraining. The Heavy Ranker is a machine-learning model that's retrained continuously on fresh engagement data. What worked last month may not work next month if user behavior shifted. That's why "I had a banger six weeks ago and now nothing lands" is a common creator experience. It's not personal. The model updated.
The practical takeaway: check your own analytics every 14 days, not every 14 months. If a format's median engagement drops 40%+ over two weeks, the ranker reweighted. Pivot to what's working for similar accounts in your niche.
- Discover: Weekly X content audit template
What Signals Penalize Your Reach?
Negative signals in the X ranker are stronger than positive ones. A predicted report multiplies the score by -1.5. A predicted block multiplies it by -3.0 — six times the weight of a predicted like (X Engineering, 2023). Outbound links trigger a separate suppression layer, and coordinated behavior patterns trigger shadowbans that last days to weeks.
Here's the negative-signal landscape:
Signal Impact Predicted report Score × -1.5 Predicted block Score × -3.0 Outbound link (free account) ~95% reach loss vs identical link-free version Mass unfollow event Up to 3-month shadowban Rapid mass-following / unfollowing 48-72 hour shadowban (first offense) Duplicate link posting (shorteners, affiliate) Shadowban escalation Sources: X Engineering (2023), hashmeta (2025) citing Buffer.
The outbound-link penalty is the one most creators stumble on. Post a link to your blog from a free account and median engagement collapses. The practical workaround is a two-step pattern — write the post with the insight, then drop the link as the first reply under your own tweet. That reply inherits most of the parent's reach without triggering the suppression layer.
Most automated shadowbans lift within 48-72 hours once the triggering behavior stops. Repeated violations extend to 7-14 days, and mass-unfollow patterns can trigger three-month restrictions.
Commeta team observation: Across our beta users who switched from manual mass-reply tools to human-in-the-loop AI drafting, reach recovery from minor shadowbans averaged under 72 hours. The pattern that worked was slowing reply cadence, keeping each reply contextual, and pausing link-dropping for 48 hours.
The cleanest rule: if a signal looks automated to a human reviewer, it looks automated to the algorithm. Slow down, personalize, and don't link from a fresh free account.
- Discover: is AI reply automation safe on X?
Frequently Asked Questions
Does X Premium boost algorithmic reach?
Yes, substantially. Buffer's analysis of 18.8 million posts across 71,000 accounts found free accounts get under 100 median impressions per post, Premium hits around 600, and Premium+ crosses 1,550 (Buffer, 2025). The lift is strongest under 5,000 followers and shrinks for established accounts with warm audiences.
Does the X algorithm punish outbound links?
Yes — for free accounts. A/B tests of identical tweets showed a 1,700% reach increase when the link was removed (hashmeta, 2025, citing Buffer data). Workaround: post the insight in the main tweet and drop the link as the first reply under your own post.
How can I check if I've been shadowbanned on X?
Open a logged-out browser and search from:yourhandle — if recent posts don't appear, you're likely in a search-suggest shadowban (OpenTweet, 2026). Most automated bans lift within 48-72 hours of stopping the triggering behavior, though mass-unfollow patterns can trigger restrictions lasting up to three months.
Does Grok read images and video content on X?
Yes. The late-2025 Grok update specifically "reads every post and watches every video" to build semantic matches (Social Media Today, 2025). Alt text, on-image text, and video captions now contribute to your post's semantic fingerprint for SimCluster routing.
When was the last major X algorithm change?
The biggest recent shift was the Grok transformer rollout in late 2025, which replaced the heuristic candidate selector (Social Media Today, 2025). Link-suppression tuning intensified in early 2026. Smaller weight rebalances happen continuously because the Heavy Ranker is a live machine-learning model.
Conclusion
Three things actually move the needle in the 2026 X algorithm. Replies with author responses carry 150x the weight of likes, dwell time carries 20x the weight of a like, and Grok's semantic matching rewards niche specificity over generic volume. The ranker still runs on the open-source weight table underneath, which means the math hasn't changed — only the candidate filter has.
If your X growth has stalled, the playbook isn't a mystery. Post long-form content with clear dwell-inducing structure, reply to 30-50 mid-sized accounts daily to farm the +75 weight, and keep your semantic footprint narrow enough for Grok to route you cleanly. And if manual reply targeting is what's eating your day, try Commeta free — it surfaces the right accounts and drafts contextual replies you edit in seconds, no credit card required.
- Discover: How often should you post on X in 2026?