Skip to content

← All posts

Published in Guides

X Algorithm 2026: How to Go Viral, From 393 Posts and X’s Ranking Code

Growth advice is everywhere on X right now.

Threads about hooks. “The algorithm changed” posts. Guides written by accounts with 5,000 followers explaining how to get a million views. Screenshots of dashboards with no context.

Almost none of it comes with evidence. It is opinion, repeated confidently by enough people that it starts to sound like fact.

So I did the opposite. I pulled every post I published over five months, all 393 of them, tagged each one by hook, format, length, media and timing, and then read X’s published ranking code to check what the platform actually rewards.

Here is what everyone repeats but nobody verifies, including the parts that make me look bad.

The 500x test: same story, three hooks

On 12 June 2026, SpaceX filed to go public after twenty years of refusing. I posted about it three times across two days. Same story, same account, same research, same week. Three different first lines.

  1. “Elon just created 4,400 millionaires in a single day.” 7,600,000 views
  2. “SpaceX’s 11th employee just became a billionaire.” 438,400 views
  3. “Elon spent nearly 20 years promising SpaceX would never go public.” 15,248 views
Impressions from X’s own post analytics, as of 2 November 2026. All three posts are still live: each line links to its post.

Look at the third line again, because this is the part that surprised me. It is not badly written. It names Elon, it sets up a promise that got broken, it has real tension in it. A year ago I would have been pleased with it.

It got 0.2 percent of the views of the first one: about 500 times fewer.

One thing I will not overclaim. The 15,000-view post went out less than two hours after the 7.6-million one, and X discounts each additional post from the same author in a viewer’s feed by half, down to a quarter (I come back to this below). So some of that gap is me competing with myself. That discount can explain a 2x or 4x gap. It does not explain a 500x one.

Here is the part that actually changed how I think about this.

Per person who saw it, the third post did not do badly. 2 percent of the people shown it engaged. On the 7.6-million post, 3.3 percent did. So a 1.6x difference in how people responded is sitting underneath a 500x difference in how many people were shown it.

The hook is not mainly changing whether people like your post. It is changing how far the post is allowed to travel before anyone gets the chance to decide. That distinction runs through everything below.

That is the whole argument of this guide. Going viral on X is not a mystery and it is not luck at scale. It is a small number of decisions, most of them made before you write a single word of the body, and almost all of them measurable. I have the measurements.

Why trust these numbers (and where you shouldn’t)

Almost every guide to going viral on X is written by an account with 5,000 or 20,000 followers. I have 181,000, and I have been posting here since before the platform changed hands, back when none of it paid anything at all.

I did not start posting to make money. I started because I like writing. That matters more than it sounds. Plenty of people here post to earn, and there is nothing wrong with that. But when someone earns from the advice itself, it is hard for them to tell you which parts are true and which parts sell.

So here is my position, plainly. I am eligible for X’s Original Content Rewards programme. I have not applied for it. And I built Postey, the scheduling tool I mention later, so discount that section accordingly. Everything before it stands on the data alone.

What this platform gave me was never payouts. It is the only place I know of where you can write something true and have it read by the people who actually run the industry you are writing about. That is available to anyone who writes something worth reading, which is the entire reason I am writing this down instead of keeping it.

The baseline nobody publishes

Every growth guide shows you the winners. None of them show you every post, including the ones that flopped. Here is mine, for the 393 top-level posts I published between 4 May 2026 and 8 October 2026.

MeasureMy 393 posts
Median views8,041
Median likes31
Median replies6
Reached 50,000 views10%
Reached 100,000 views7%
Best post17,470,000 views
  • 7% passed 100,000 views
  • 3% landed between 50,000 and 100,000
  • 90% stayed under 50,000
Ninety percent of what I post does not go viral. Each square is 1% of 393 posts.

That is the honest number and the most useful one here, because it reframes the whole question. Virality is not a quality you add to a post. It is a hit rate. The only thing worth improving is how often you land in that top 10 percent, and the only way to improve it is to know why the other 90 missed.

So the rest of this guide is two things: what the 393 posts say about why some land, and the system I run to make landing happen more often.

How to build a hook (the only part that really matters)

Every guide says the hook matters. None of them tell you what a hook actually is, so people write a punchy-sounding sentence and wonder why it died.

A hook is one claim: the most recognisable actor in your story, doing the most surprising true thing in your story, in as few words as possible.

That definition is narrow on purpose, and my data is narrow with it. I tagged 101 posts by the shape of their first line. Across those, 12 percent reached 50,000 views. The groups below are small (10 to 44 posts each) and some overlap, so read them as direction, not law.

  • A big name alone does almost nothing. Of 44 posts that opened with a major name, 14 percent reached 50,000 views, barely above the tagged set’s 12. Some of the surprising posts below are inside those 44, so the name on its own is worth even less than that.
  • A big name doing something surprising, with stakes, changes everything. 12 posts. Median 33,705 views. 42 percent reached 50,000.
  • A big name saying something buys nothing. 10 posts that opened with a reporting verb (says, argues, on, asked): one of the ten reached 50,000. Tiny, but it matches what I see every week.

Two of mine, both real, both still up. “Jensen Huang argues the narrative on automation destroying jobs is wrong” is a dead post: the biggest name in chips, attached to a verb that promises nothing. “NVIDIA just dropped Nemotron 3.5 Lightning” did about 4,000 views, because dropping a model is a thing NVIDIA does every month. Nothing in either line violates expectation, so the name does no work.

How I build line one

This takes me about ten minutes per post, and it is the highest-return ten minutes in my week.

  1. List every verified fact in the story. Not the interesting ones. All of them. The hook is almost always a fact I already had and nearly threw away.
  2. Find the biggest gap between what a reader expects and what is true. Not the most important fact, the most surprising one. Those are different facts, and the important one usually makes the worse hook.
  3. Make the most recognisable actor the subject, doing it. Active verb, strong and true: created, told, filed, banned, bought. Never says, argues, on, discusses.
  4. Add a compressed time frame if it is true. In a single day. In six months. In 48 hours. Speed is free surprise.
  5. One number at most, and the most extreme one. A number alone gave me no lift across 28 posts. A number as the payoff of an action is what works: 4,400 millionaires is the payoff of created.
  6. Delete everything before it. No setup, no “Why this matters”, no “Here is what happened”. If the sentence starts with why or here, the hook is in the middle of it and the first half is dead weight.
  7. Keep line one true as written. If the claim needs a qualifier to be honest, the qualifier goes in line two. Never soften line one and never overstate it. A hook you have to walk back costs more than it earns.

The test

Read line one on its own, as a stranger scrolling past at speed, with no idea who you are. Would they stop?

Most of my first lines are under ten words. That is a habit, not a finding: shorter lines are easier to read at scrolling speed, and every word you cut is one fewer reason to keep scrolling.

I write five hooks before I pick one. Not as a creativity exercise: because the first hook I write is almost always the summary, and the summary is the 15,000-view version.

How the X algorithm works: what its ranking code rewards

X publishes its ranking code. I first read it in October and last checked it on 12 November 2026, so you do not have to. Here is how it works, without the jargon. For every single person who might see your post, X’s model predicts how likely that particular person is to reply to it, send it to a friend, follow you, or mute you. Each prediction is multiplied by a weight and the results are added up. That total is your score for that one viewer.

Your post is not scored on what it already got. It is scored on what X thinks the next person is about to do.

Four things in the code changed how I write:

  • It is looking for posts people send to someone. Copying the link to share it is the single heaviest positive weight. So the target is not writing things people approve of. It is writing things people send to one specific person.
  • Likes are one of the lightest signals. Replies weigh far more, and a reply between two accounts that follow each other gets an extra boost. I stopped looking at my like count.
  • Irritating people is the expensive mistake. Mute, “not interested” and report carry big negative weights. That is why engagement bait stops working. There is no secret penalty: if the model expects some viewers to hide your post, it stops showing it.
  • Bookmarks are not in the list at all. They will not get you in front of one extra person. I am saying this in a guide I would quite like you to bookmark.

One warning: do not turn these numbers into exchange rates. “A report cancels 468 likes” is the most repeated version, and the file itself says that reading is wrong. The details are below if you want them.

For the curious: the actual weights, and why they keep changing
  1. Copies the link to share it +20
  2. Replies (more when you follow each other) +5
  3. Sends it by DM +5
  4. Quotes it +5
  5. Follows you +4
  6. Shares it another way +2.118
  7. Reposts it +1
  8. Likes it +0.708
  9. Blocks you −31.2
  10. Taps “Not interested” −53.07
  11. Mutes you −58.8
  12. Reports the post −234
X’s default ranking weights, checked 12 November 2026. Above zero, the model wants more of that action; below zero, it pushes the post down.

The weights live in home-mixer/params/param.rs in the xai-org/x-algorithm repository (commit f06f45a). A comment above them explains that each weight multiplies a predicted probability, not a count. Rare actions like reports get big weights because they are rare: a report is more than a thousand times less likely than a like, so without a big weight the prediction could never move the ranking. That is why ratios between weights are not trades between actions.

The defaults also move. The like weight alone was changed three times between October and November 2026: 0.5 (until 9 Oct 2026), 0.6 (13 Oct 2026), 0.66 (4 Nov 2026), 0.708 (5 Nov 2026). My post data runs from May to October, when the like weight was 0.5, so I use the code for its shape, which has held, not its exact values. The file also warns that live values can differ from these defaults. And ignore anything quoted from the 2023 version of the algorithm: “reposts are 20x a like” and “bookmarks are 10x a like” appear in neither version.

The chart draws each bar at the square root of its weight so small weights stay visible next to the report’s −234. The printed values are exact.

Two multipliers outside the weights

  • Strangers cost more. Posts shown to people who do not follow you are multiplied by 0.75 on purpose. Reaching new people is harder by design, not by bad luck.
  • You compete with yourself. Every extra post of yours in the same person’s feed is multiplied by 0.5 again, down to a floor of 0.25. This matters later, when we get to the calendar.

Five things my own posts proved me wrong about

1. A post without a picture is not finished

  1. Posts with an image or video (194) 10,196 views
  2. Posts with no media (38) 3,547 views
Median views of my 232 original posts, with and without an image or video. The other 161 of the 393 were quote posts, covered under format below.

Not one of my 38 text-only original posts reached 100,000 views. Not one. I used to think the writing should be able to carry it alone. In my data, it could not.

2. A link in the body kills the post

Twelve posts with a link in the main text: median 3,147 views, less than half my overall median, and none past 50,000. Twelve is a small sample, and the public ranking code has no link penalty in it, so I cannot prove why. But moving the link costs nothing, so my links go in the first reply now.

3. Hashtags did nothing

Thirteen posts with hashtags, not one past 50,000 views. Another small sample, but I saw no upside, so I stopped using them.

4. The templated opener is the most expensive habit in content advice

You know the shape:

  • “X was never Y. It was Z.”
  • “X isn’t just Y, it’s Z.”
  • “Most people do X.”
  • “Every company has X.”

I ran 33 of them. Median 20 likes, against my overall 31. Zero reached 50,000 views, against 10 percent of everything else. Three of mine, so you can see how bad it gets:

  • “Most robots memorize steps”: 11,032 views, 9 likes
  • “Every company has a copilot”: 4,778 views, 7 likes
  • “The real bottleneck for knowledge creators was never making content”: 36,667 views, 11 likes

That last one is the worst of the three, and it looks like the best. Thirty-six thousand people were shown it and eleven of them liked it. Those lines feel like insight while you are writing them. That is the trap. They read as filler, and lately they read as AI-written, which is worse.

5. Writing to “you” in the opening made things worse

133 posts that opened with you or your: 5 percent reached 100,000, against 9 percent for everything else. That is a small difference, so treat it as a lean, not a law. I think the reason is simple. Opening with “you” is almost always the start of advice, and nobody is on X hoping to be advised by a stranger. Opening with a fact is the start of a story.

None of these five are clever. That is rather the point. The gap between a post that reaches 8,000 people and one that reaches 800,000 is mostly made of boring decisions made correctly, over and over, while everybody looks for the trick.

What you can fix today, and what takes months

This is the part nobody separates, and it is why most people give up. Growth advice bundles things you can change in ten seconds with things that take half a year to build, then presents all of it as one list. People try the hard things first, fail, and conclude the platform is rigged. There are two levels. Do level one first.

Level 1: fix these today (minutes, and they cost nothing)

Every one of these is a decision, not a skill. You can apply all seven to your next post.

  1. Attach a picture or video to every post. With media: 10,196 median views. Without: 3,547, and not one of 38 reached 100,000.
  2. Move every link into the first reply. 12 posts with a link in the body: 3,147 median views, none past 50,000.
  3. Delete the hashtags. 13 posts with hashtags, none past 50,000 views.
  4. Stop opening with “you”. 5 percent reached 100,000, against 9 percent without.
  5. Kill the templated opener. 33 posts in the “X isn’t just Y” family: zero reached 50,000.
  6. Write five first lines, then throw away four. Ten minutes. On the SpaceX story, the first line was the biggest difference between a 15,000-view post and a 7.6-million one.
  7. Space your posts a few hours apart. The reason is in the code: the ranker multiplies each additional post from you in one person’s feed by 0.5. The code has no fixed gap. Four hours is my rule of thumb, and 75 percent of the 363 posts I measured follow it.

Time investment: minutes per post. These are the cheapest gains available, and almost nobody does all seven.

Level 2: these take real work (weeks to months)

These are the ones the screenshots are selling you. They are real. They are also slow, and they only pay off once level one is already habit.

  • Developing an actual beat. My China posts reach 100,000 views 13 percent of the time. My Claude and Anthropic posts, 24 percent. Those topics are not magic. My best guess is that I posted about them consistently enough for X to learn who to show them to. It takes months on a narrow subject, and the discipline to turn down topics that are not yours.
  • A research habit that produces facts nobody else has. The hook is almost always a fact, not a phrase. Finding facts other people missed is the actual job: primary sources daily, filings and papers instead of other people’s posts about them.
  • Earning the right to a one-line reaction. My highest-ceiling format is a quote post of twelve words or fewer. It only works because people already know who I am. A stranger posting “this is ridiculous” over a video reaches nobody.
  • Being somewhere worth writing from. First-hand beats commentary every time: going places, testing things, meeting people.
  • A system that survives a bad week. Consistency is not motivation. It is a calendar and a backlog.

Level one costs nothing but attention. Level two takes months. Most people attempt level two on day one, skip level one entirely, and end up with neither.

If you’re not a 181K-follower news account

My data comes from one AI and technology account with a large, mostly American audience. You probably are not that. Here is how I would translate it.

What should carry over to any account

The structural rules: media on every post, the link in the first reply, no hashtags, no templated openers, and space between posts. They line up with how the ranker works, not with my topic, so I expect them to hold for a 300-follower account, a brand and a fitness creator alike. Test them anyway; they cost nothing.

What probably does not

Celebrity hooks, one-line reaction posts and the breaking-news habit. Those depend on an audience that already trusts you and a beat with news in it. At a smaller size, the one-liner reaches almost nobody, because nobody yet knows whose reaction it is.

The hook, when the story is your own

The rule does not change: the most surprising true thing, about the most recognisable thing in the story, in under ten words. When there is no Elon in it, the recognisable thing is whatever your reader already knows: a familiar problem, a named tool or company, a number with a unit people feel. These are illustrations of the rewrite, not posts from my data. Use your own real numbers.

WhoThe summary (what most people post)The fact (the hook)
Founder“Shipped dark mode this week!”“312 users asked for dark mode. It took one afternoon.”
Brand“Introducing our new reporting dashboard.”“Our customers spent 6 hours a week on reports. Now it’s 10 minutes.”
Fitness creator“You’re deadlifting wrong.”“I deadlifted wrong for 8 years. One cue fixed it.”
Executive“Some thoughts on remote work.”“We closed our office in 2023. Hiring got easier.”

Notice what changed. The summary announces a topic. The fact reports something that happened, with a number or a stake in it. Opinion posts work the same way: lead with the fact that made you hold the opinion, then give the opinion.

If you have a small account

Do not measure yourself against 50,000 views. Measure against your own median. If your typical post gets 300 views, a post that gets 3,000 is your viral post, and the job is to find out why that one did and do it on purpose. Tag every post by its hook shape from day one. In three months you will have your own version of this guide, which is worth more than mine.

If you post for a brand, or for clients

Everything structural applies. Swap the celebrity for your customer: their number, their problem, their result. Report hit rate and replies upward instead of likes; the ranking code above is the argument for why. If you run several accounts, Postey keeps a separate set of accounts per client and lets clients comment on drafts before anything is scheduled ( how agencies use it ).

If you are coming from Instagram

  • Hashtags do not carry over. On Instagram they help discovery. On X my 13 hashtag posts showed no upside at all.
  • “Link in bio” becomes “link in the first reply”. Same idea, one tap closer.
  • Your Reels are already X media. My data does not split video from images, so I cannot tell you which wins. Together, they beat text alone by a wide margin.
  • You do not have to post more. Spacing matters only once you post more than once a day. One good post a day, with a fact for a first line, is a fine start.

If you delegate your posting

Split the work by what only you can do. Your assistant (human or AI) gathers facts, drafts five first lines per post and schedules. You pick the line, because taste is the part that does not delegate. Opinion posts work the same way as news: lead with the fact that made you hold the opinion. And if you cannot give replies a full hour, give them ten minutes after a post goes out: answer the five best replies yourself. That is where most of the reply value is.

If you post to earn

As reported at launch, X’s Original Content Rewards measures eligibility on impressions from verified users (500,000 in 90 days) and asks for 500 verified followers, so reach among verified accounts is what counts, not raw views. Everything structural in this guide still applies. The main risk is the rules: they prohibit repeatedly instructing people to engage, so the “comment YES” habit can cost you eligibility as well as reach. Requirements have changed more than once, so check X’s Help Center before you plan around them. There is more in my guide to X Premium and payouts.

Format decides your ceiling before you write a word

The same material in the wrong format caps out. These are my numbers, not general ones, but the direction has been consistent enough that I now pick the format before I draft.

The one-line quote post is my highest-ceiling format. Quote posts of twelve words or fewer: 8 posts, median 14,421 views, two of the eight past 100,000. Longer quote posts: 153 posts, median 7,419 views, 5 percent past 100,000. Small sample on the short ones, so hold it loosely, but the gap matches how the format works: the quoted post is the evidence and my job is the reaction, not the explanation. “y’all better read Fahrenheit 451 and 1984 while you can” did 448,000 views and took about eleven seconds to write. My rule: if the one-liner needs a sentence of explanation under it, it is not a one-liner. Write a single post instead.

The single post with media is my most reliable format. This is where the research goes: a hook, a second surprising fact before the fold, then names, numbers and dates in short paragraphs, one sentence pre-empting the obvious objection, and the sharpest fact last. Three of mine built that way, all still up:

  • “WTF. AI companies are purchasing large quantities of used and rare books”: 17.47 million views
  • “Chinese models are 112x cheaper than Anthropic per million tokens”: 840,000 views
  • “A Chinese robot just performed a flying side kick, a sword fight, and a fencing match”: 557,000 views

What those three first lines share: a subject you already recognise, a verb that is doing something rather than saying something, and a fact you would not quite believe without the picture attached. None of them explain. All of them report.

They also bend my own rules: all three run past ten words, one opens with “WTF.” before the fact, and one leans on a bare number with a weak verb. They worked because the fact underneath was strong enough to survive it. I would not bet on that twice.

Threads mostly do not travel, and the code explains why.

  1. First post of a thread (17) 10,665 views
  2. Every post after it (41) 776 views
Median views in my threads: a 93 percent drop between the first post and the rest.

X filters replies from accounts a viewer does not follow out of their recommendations entirely, and extra branches of one conversation get collapsed together. So the payoff goes in post one, never the promise of a payoff. If post one does not work as a complete post on its own, the thread is already dead, and everything after it will reach a small fraction of the people who saw the first post.

Articles are the save format, not the reach format. My 13 X Articles: median 11,052 views, 15 percent past 50,000, against 10 percent overall. A better hit rate than average and a much lower ceiling than a viral single post. An Article is what people keep. The post announcing it is what people see. Those are two different pieces of writing, and both have to work.

The system I run every day

Everything above is what the data says. This is what I actually do with it, five days a week.

Step 1: find the story before anyone else has an angle on it

The hardest part of posting consistently is not writing. It is arriving at the desk with something worth writing about. So the first job of the system is to make sure I never start from an empty page. I run three inputs every morning:

Not a news account? Keep the idea, swap the inputs: your product metrics, support tickets and customer calls if you are a founder or brand; your training log and client questions if you coach; your board notes and deal flow if you invest. The job is the same: find the fact nobody else has.

  1. Primary sources only, for my beats. Vendor announcements, system cards, technical reports, filings, repos, lab pages. Not other people’s posts about them. If I am reading a take instead of a source, I am already second.
  2. What happened in the last 24 hours. Breaking framing only works inside about 48 hours. After that the word “just” is a lie and the post reads like a recap.
  3. What is already saturated. Before I pick an angle I check what has already been posted. A crowded story is not a reason to skip it. It is a reason to find the fact nobody has led with.

Then the step most people skip: list every verified fact before choosing one, each with its source URL. The 4,400 millionaires number was not the headline of the SpaceX story. The headline was the IPO. The number was buried in the filing. The headline did 15,000 views and the buried number did 7.6 million.

One rule I do not break: every number, name and quote traces to a primary source I opened that day. Not model memory, not another creator’s post. On a sophisticated audience one wrong number costs more than the post earns, and the quote post correcting you will outperform your original.

Step 2: five hooks, one draft, and an AI that cannot publish

I connect Claude to my posting calendar through Postey’s MCP server. I describe the story and the angle in plain language, and five hook options come back against the rules I have given it: no hashtags, no templated openers, no “you” in the opening, one extreme number, under ten words where possible. The drafts land on the calendar as drafts.

Claude drafting into Postey over MCP. Product demo, not my account.

The important part is the permission setting, because “let AI run your account” is terrible advice. Postey lets you choose which actions your AI is allowed to take. Mine can draft. It cannot publish. Nothing reaches X until I have read it, cut it, and pushed it myself. I rewrite most of what comes back, and the value is not the finished sentence. It is that I never stare at a blank composer, and that I get five options instead of marrying the first line I thought of.

That matters because of the hook test above: the first line I write is almost always the summary, and the summary is the 15,000-view version. Generating five and killing four is a mechanical fix for a psychological problem.

My honest criticism of this setup, and discount me accordingly because I built it: the AI is good at structure and bad at taste. It reliably produces five technically correct hooks and has no instinct for which one a human would stop on. It cannot tell you that “Elon just created 4,400 millionaires” is the line, because that judgement comes from having watched a thousand posts land. Use it to widen the options. Do not use it to choose.

Before anything is scheduled, I read line one out loud, alone, with the rest of the post hidden. If I would not stop for it, nothing else gets fixed. I go back to the fact list.

Step 3: the calendar, and why spacing beats frequency

  • Batch the writing, drip the posting. Writing and posting are different jobs, and doing them in the same session produces worse writing and worse timing. I write in blocks and the calendar releases them.
  • Space posts at least four hours apart. Each additional post in one viewer’s feed is multiplied by 0.5, down to 0.25. You are not competing with other creators for that slot. You are competing with your own previous post.
  • Never put a strong post directly behind a weak one. The weak one eats the discount the strong one needed.
  • Post into your audience’s window, not yours. Find it in your analytics: the hours your followers are most active. Mine is 12:00 to 17:59 UTC because my audience is American and I am not. Before I used a calendar I posted whenever I finished writing, which often meant missing that window entirely.
  1. Four or more hours after my previous post 8,658 views
  2. Under an hour after it 6,281 views
Median views of my posts by the gap since my previous post. A modest gap on its own; the stronger reason to space posts is the 0.5 discount in the code.
A week laid out on Postey’s calendar. Product demo.

The reason I run this on a calendar rather than in my head is the reactive half. When something breaks in my beat I have about a 48-hour window, and I need the slot to be free and the rest of the week already handled. A full calendar is what buys you the freedom to drop everything for a story that is actually worth it, so I always leave slots empty for news.

Step 4: the first hour decides the next three days

Publishing is the middle of the job, not the end. Four things happen after a post goes live, and I automated the three I kept forgetting.

  • The link goes in the first reply, automatically. I have it post itself as a comment once the post starts moving, so the link arrives when there is an audience to click it rather than sitting under a post nobody saw.
  • Replies for the first hour, by hand. A predicted reply is one of the heaviest positive weights, with an extra boost between accounts that follow each other. This is the one part I do not automate and never would. Replying is where the account is actually built. Every reply adds a fact the commenter did not have. Never just agreeing.
  • The auto-repost. A post that worked once gets a second showing to the half of the audience that was asleep. Scheduled, not manual, because manual means it does not happen.
  • A keyword reply, used honestly. If a post promises something real (a repo list, a template, a set of links), I send it to anyone who comments a keyword. It drives replies, and replies are what the ranker is looking for.
The link, auto-commented once a post takes off. Product demo.

I want to be careful about the keyword reply, because it is exactly the mechanic that turns into slop. X’s Original Content Rewards rules prohibit repeatedly instructing users to engage, which covers asking for likes, replies, bookmarks, follows or reposts again and again. And the ranking code is brutal on anything that irritates people. So the line I hold: the thing being sent has to exist, it has to be worth the comment, and the ask appears once. “Comment REPOS and I will send you all five links” is a trade. “Comment YES if you agree” is bait, and bait is a negative-weight machine.

Measure the right thing, which is almost never views

My most-viewed post of the measured window did 17.47 million views. It is not my most valuable post. Views are the number everyone optimises because it is the number everyone can see. A post can go wide inside a crowd that has no interest in anything else you will ever write. 500,000 views that bring no followers and no replies is a post that happened to you, not a post you can repeat. These are the five I look at every week:

  1. Hit rate. The share of posts that cleared 50,000 views. Mine is 10 percent. It is the only number that tells me whether the system is working, because virality is a frequency, not an event.
  2. Replies per post. The best early signal that a post is going to travel. My median is 6.
  3. Followers per post, not per view. A post that gained 500 relevant followers on 50,000 views beat one that gained 20 on 500,000. The second is a lottery ticket. The first is a repeatable result.
  4. Which hook shape won. Every post gets tagged with its hook type when it goes out. That is the entire reason I could write the hook section of this guide.
  5. What died, and what shape it was. Failures teach more than wins. I only found the templated-opener problem by looking at my worst posts instead of my best.

What this data can’t tell you

Every guide to going viral is written as though the author has it solved. I have not, and this data has real holes in it.

  • It is one account, one niche, one audience. 393 posts from an AI and technology account with a largely American audience. The structural findings (media, link placement, spacing) probably transfer, because they line up with the ranking code rather than with me. The topic findings almost certainly do not. If you post about fitness or fashion, test everything here before believing it.
  • It is correlation, all the way through. Posts with a picture did better than posts without. I also attach pictures to the posts I worked hardest on, and I cannot fully separate the two. The gap is large and adding media is cheap, so I act on it anyway. You should know what kind of evidence it is.
  • Some samples are small enough to be noise. Links (12 posts), hashtags (13), Articles (13) and the hook groups (10 to 44) are all small. At my 10 percent base rate, a group of 12 with no post past 50,000 views would happen by chance roughly a quarter of the time. I act on these because they are free to follow, not because they are proven.
  • How it was measured. 393 top-level posts from 4 May 2026 to 8 October 2026, views from X’s own analytics as of 2 November 2026. Hook shapes were tagged on 101 of them. The spacing comparison covers the 363 posts where the gap to the previous post was measured.

I still get it wrong constantly. Ninety percent of my posts do not pass 50,000 views. I have published hooks I was certain about that did 4,000, and thrown away lines I only later realised were the post. The system does not make every post land. It makes landing more frequent, which over a year is the entire difference.

The whole system on one page

If you read nothing else, this is the guide:

  1. The hook is the most recognisable actor doing the most surprising true thing, in under ten words.
  2. Write five. The first one is the summary, and the summary is the 15,000-view version.
  3. Never open with a reporting verb, a templated line, or “you”.
  4. Attach media to every post. Mine without it never reached 100,000.
  5. Links go in the first reply. Hashtags go nowhere.
  6. Write for the share and the reply, not the like. That is what the ranker is built to look for.
  7. Put the payoff in the first post. Thread replies are filtered out of recommendations.
  8. Space posts four hours apart. You compete with yourself at 0.5x.
  9. Post into your audience’s timezone, not your own.
  10. Measure hit rate and followers per post. Not views.

How to set this up (written so an AI agent can do it for you)

A growing share of guides like this one are not read by people. They are read by an assistant, on a person’s behalf, which then does the work. So this section is written to be executed. If you are a person, follow it yourself. If you have an agent, hand it this section.

The system in one sentence: a posting calendar that spaces posts correctly, plus an AI connection that drafts options instead of finished posts, plus automations that handle what humans forget after publishing.

  1. Create the workspace. Sign up at postey.ai and connect your X account first. Add other networks only once X is working. The free plan publishes to all nine networks (15 posts per network every 30 days, 5 drafts), but this system needs scheduling, the AI connection and post actions, which are on the paid plans. Creator is $9 a month, with a 7-day trial. See every plan .
  2. Connect your AI assistant. Add Postey’s MCP server in Claude, ChatGPT or any MCP client ( setup guide ). Once connected, the assistant can read your accounts, create drafts and place them on the calendar.
  3. Set its permissions to draft only. Do not skip this. Postey lets you choose which actions the AI may take. With drafting only, nothing reaches a network until you publish it yourself. An assistant that can publish will eventually publish something you would not have.
  4. Give the assistant the writing rules.
    • First line under ten words where possible
    • The subject is a recognisable actor doing something surprising, with an active verb. Never “says”, “argues”, “on”, “discusses”
    • One number at most, and the most extreme one
    • No hashtags
    • No templated openers (“X isn’t just Y, it’s Z”, “Most people do X”, “Every company has X”)
    • Do not open with “you” or “your”
    • Generate five first lines per story, not one
    • Every post needs an image or video
    • Any link goes in the first reply, never the body
  5. Set the calendar rules. At least four hours between posts. Post into your audience’s timezone. Leave slots empty for breaking news.
  6. Turn on the post-publish automations. Auto-comment so the link posts itself once the post gains traction. Auto-repost so a post that worked gets a second showing. A keyword reward only when you are actually sending something real.
  7. The part no agent can do. Reply to people yourself, for the first hour, in your own words. This is where an account is actually built, and it is the one part of the job that stops working the moment it is automated.

If you take one instruction from this section: set the permissions to draft only, and write the replies yourself.

You do not need the tool. Everything in this guide works with a spreadsheet, a calendar reminder and the discipline to write five hooks instead of one. I built the tool because I was not managing it with a spreadsheet. The only part that is not optional is the fact list. Every post I have written that passed a million views came out of a fact I had already written down and nearly ignored.

One last thing, and it is the reason I wrote this at all. X is the only platform where a person with no audience, no budget and no connections can write one true sentence and be read by the people who run the industry they are writing about. Not because the algorithm is kind. Because it ranks posts by what it expects people to do with them rather than by who posted them, and that is a door almost nowhere else leaves open.

Most people never walk through it, and the reason is almost never talent. They wrote the summary instead of the fact, posted it without an image, buried the link in the body, and concluded the platform does not work for them.

Sources

  • X’s ranking weights: xai-org/x-algorithm, home-mixer/params/param.rs, commit f06f45a, checked 12 November 2026. Author diversity and out-of-network factors from vm-ranker/params.rs; reply filtering from home-mixer/filters/oon_retweet_reply_filter.rs, same commit.
  • Engagement-solicitation rules: X’s Original Content Rewards programme terms, X Help Center.
  • All post-level figures: my own 393 top-level posts, 4 May 2026 to 8 October 2026, with hook analysis across 101 posts. View counts as of 2 November 2026.

Frequently Asked Questions

How do you go viral on X?

Lead with one surprising, true fact about the most recognisable person, company or thing in your story, in under ten words. Attach an image or video, keep links and hashtags out of the post, and space posts at least four hours apart. On one story, the post that led with a buried fact got 7.6 million views; the one that led with the headline, posted two hours later, got 15,248. Part of that gap is the two posts competing, but most of it is the first line.

How does the X algorithm work in 2026?

X predicts what each viewer will do with your post and adds those predictions up, weighted. In the public ranking code a predicted copy-link share carries a weight of 20, a reply or DM share 5 and a like 0.708, while a predicted mute (−58.8) or report (−234) pulls the score down hard. The weights apply to predicted probabilities, not raw counts, so they are not exchange rates between actions. They also change: the like weight was changed three times between October and November 2026.

Do bookmarks help your reach on X?

No. Bookmarks do not appear in X’s published ranking weights at all. They are a sign that a post was worth keeping, not a signal that gets it shown to more people.

How often should you post on X?

Spread them out. X’s ranker multiplies each additional post from the same author in one viewer’s feed by 0.5, down to a floor of 0.25, so back-to-back posts compete with each other. My rule of thumb is at least four hours apart. In my data those posts had a median of 8,658 views against 6,281 for posts under an hour apart: a modest gap, so the code is the stronger reason.

Are threads still worth it on X?

Only if the first post works on its own. In my data the opening post of a thread had a median of 10,665 views and every post after it 776, a 93 percent drop. X filters replies from accounts a viewer does not follow out of their recommendations, so the payoff has to be in post one.

Do hashtags help on X?

Not in my data. None of my 13 posts with hashtags passed 50,000 views. That is a small sample, but hashtags cost reading space and showed no upside, so I stopped using them.

Start publishing on Twitter today.