algorithm
How the X algorithm treats crypto accounts
How the X algorithm ranks crypto accounts: the 2023 Twitter weights, the 2026 xAI code and its weights, and what they mean for replies, links and ragebait.
The X algorithm scores a crypto account's posts with the same formula as everyone else's. For each viewer, a model predicts how likely they are to like, reply, share, mute or report, and then ranks posts by a weighted sum of those predictions. Conversation earns distribution. Anything that makes your audience mute or report you costs it.
Most advice about the Twitter algorithm in crypto circles still quotes numbers from 2023. X has since replaced that system and, in August 2026, published the weights of the new one. This article covers both releases, what the public code can and cannot tell you, and what to change in how your project posts.
The 2023 heavy ranker weights
In 2023 Twitter open-sourced its recommendation code, including the "heavy ranker" that ordered the home timeline. Its README, dated April 2023, listed the weight on each predicted engagement:
| Predicted action | Weight |
|---|---|
| Like | 0.5 |
| Retweet | 1 |
| Reply | 13.5 |
| Good click (opening the conversation and engaging) | 11 |
| Good profile click | 12 |
| Reply that the post's author engages with | 75 |
| Negative feedback (show less often, block, mute) | -74 |
| Report | -369 |
The same document notes that the weights had been "periodically adjusted to optimize for platform metrics", so even in 2023 these were a snapshot.
These numbers shaped a generation of CT advice. A reply was weighted 27 times a like. A reply the author then engaged with was weighted 150 times a like. That is where "always answer your replies" and "never get reported" became house rules for serious accounts, ours included.
One caveat applied then and applies now. The weights multiply the model's predicted probability of each action for a given viewer. They are not a price list for counts. The claim "one report cancels 369 likes" was never an accurate reading.
What changed under xAI
In January 2026, X's engineering team announced that it had open-sourced a new algorithm, built on the same transformer architecture as xAI's Grok model. The code lives at xai-org/x-algorithm. According to its README, the For You feed now works like this:
- Candidates come from two places. Thunder holds recent posts from accounts you follow. Phoenix retrieval and SimClusters find posts from accounts you do not follow.
- Ranking is done by Phoenix, a transformer that reads your recent engagement history and predicts, for each post, the probability of many actions: like, reply, repost, quote, several kinds of share, clicks, dwell time, following the author, and negative actions like "not interested", mute, block and report.
- The score is an explicit weighted sum of those predictions.
- Adjustments follow. Each additional post from the same author is multiplied by a decaying factor. Posts from accounts you do not follow are discounted. Posts from authors with few impressions get a boost toward a target position.
The update dated August 13, 2026 added the weights themselves to param.rs. The README says scripts keep those defaults in line with the main production values, while experiments run on slices of traffic. When we checked on September 15, 2026, the file held these values:
| Predicted action | Weight |
|---|---|
| Like | 0.5 |
| Reply | 5.0 |
| Reply, on an original post from an author the viewer mutually follows | 5.0 + 15.0 boost |
| Repost | 1.0 |
| Quote | 5.0 |
| Share | 2.0 |
| Share via DM | 5.0 |
| Share via copy link | 20.0 |
| Follow author | 4.0 |
| Click | 0.4 |
| Open link | 0.2 |
| Profile click | 0.0 |
| Not interested | -43.2 |
| Mute author | -58.8 |
| Block author | -31.2 |
| Report | -234.0 |
A few things stand out. The 2023 "reply engaged by author" line has no direct counterpart. Plain replies carry less weight than in 2023, but a +15 boost goes to original posts from people the viewer mutually follows. X documents how that boost was tested and rolled out in July 2026 in BIDIRECTIONAL_BOOST_CHANGE.md. Shares, especially copying a post's link, are weighted heavily. Negative actions are still large.
X also repeats the 2023 caveat in the code, with more detail. A comment in param.rs says the baseline probability of a report is more than 1000 times lower than a like, which is why its weight is large. The predictions are personalized, so reports mostly affect recommendations for viewers similar to the people reporting.
The repo is not complete. The README says that Grox prompts and some botmaker rules (part of the spam and visibility system) are withheld to make gaming harder. Instead, X points to an Under the Hood tool that shows visibility labels applied to your own account and posts. Ranking and visibility are separate layers: a post can score well and still be dropped or put behind an interstitial by visibility filtering.
Two filters matter for anyone planning distribution. Posts older than 48 hours are removed from For You candidates. Replies and reposts from accounts a viewer does not follow are filtered out as well, so a reply you leave under a big account is not recommended in For You to strangers. People see it in the thread itself.
What this means for crypto accounts
The code gives you the mechanics. The practical rules below are our reading of them, combined with how we run accounts day to day. Treat them as working rules to test on your own account, not as guarantees.
Answer the people who reply to you
The 2023 system rewarded author replies directly. The 2026 system predicts replies per viewer from each viewer's history, and it gives extra weight to original posts from mutual follows. Both point the same way. A thread where the project answers people is a thread people come back to, and the accounts you actually talk to are the ones that end up following you back.
In practice, reply to every substantive comment under your posts, ideally within the first 30-60 minutes. Write answers that give the person a reason to respond again.
Good question. The cap exists because the oracle updates every 20 minutes, so larger positions could be priced on stale data.
We'd raise it once the new feed is live. What size were you trying to open?
A plain "thanks!" closes the conversation you just earned.
Reply early, and reply where people read
Because For You drops posts older than 48 hours and filters out replies from accounts a viewer does not follow, a reply's audience is mainly the people reading that thread. So timing matters. Our rule is to reply to posts younger than about 30 minutes, where your reply can still land in the top 20, and to skip threads that already have hundreds of replies. The selection method is covered in detail in the crypto Twitter reply strategy guide.
Put links in the first reply
Here the public code does not back up a popular belief. The published weights contain no negative weight for links: opening a link carries a small positive weight of 0.2. We found no link penalty in the published weights or the scoring code, so we will not claim one.
We still put links in the first reply under the post, for a practical reason. A post is scored on the replies, quotes, shares and dwell time it is predicted to earn. A post whose whole message is "click here" gives people little to do on X, while a post that makes its point on the page and puts the link underneath serves both groups. Test it on your own account if you doubt it.
Cashtags, hashtags and topic reach
Neither cashtags nor hashtags have a weight in the published ranking. Their value is practical. Traders monitor tickers through cashtag search and X Pro columns, so a relevant cashtag puts your post where people who care about that asset are already looking. It is also where we start when sourcing replies for a client.
Hashtags earn less. One relevant hashtag is fine. A string of them reads like spam to humans, and X's authenticity policy lists posting similar content across accounts to the same trending topics or hashtags as prohibited. Do not borrow another project's cashtag to ride its attention either. People who searched for that ticker and found your pitch are exactly the viewers likely to tap "not interested".
Why ragebait costs weeks
Ragebait works on a dashboard for a day: lots of replies, lots of quotes. Then the costs arrive, and several parts of the published system explain why they last.
- Mutes and blocks are hard filters. The README lists a filter that removes posts from accounts the viewer blocks or mutes. Someone who muted you after one bad take will not see your next fifty good ones.
- Negative predictions are personal and learned. Phoenix predicts from each viewer's engagement history. If people in your niche start muting or reporting you, the model learns to predict that response, and by X's own explanation reports mostly affect viewers similar to the ones who reported. For a crypto project, those similar viewers are your target audience.
- The weights are large. Mute at -58.8, not interested at -43.2 and report at -234 sit against a like at 0.5.
Our working rule, from running accounts, is that the dip after a hostile thread lasts weeks, not days. Sharp opinions are fine. The difference shows in the writing:
If you're still farming Orbix points you deserve what's coming.
Orbix points look generous until you price the lockup: 90 days illiquid for a 6% boost.
What are we missing?
The first invites mutes. The second invites argument, and argument is what the reply prediction rewards.
What coordinated activity looks like
X's authenticity policy prohibits coordinating to exchange engagement, paying others to inflate likes, replies, reposts or follows, and running multiple accounts that engage with the same posts. It also names "follow churn", following and unfollowing large numbers of accounts to grow your own count.
The 2026 code adds a detail that should end the habit of raid groups. A comment in param.rs says engagement only counts for ranking on posts served in the Home Timeline, and that opening a post directly, for example from a group chat link, has no ranking impact. The classic Telegram raid, twenty people opening a link and liking within two minutes, carries policy risk and, by X's account, no ranking benefit.
Patterns that look coordinated, and that you should avoid:
- Your whole team liking and replying to every project post within minutes.
- Identical or near-identical replies from several accounts.
- Mass "gm" replies and follow-for-follow.
- Engagement pods with other projects.
- Replies fired off in bursts at perfectly regular intervals.
Healthy looks different: team members reply when they have something to add, at a human pace, in their own words. The accounts we run follow that pace: no GM farming, no follow-for-follow, no engagement pods, and every reply written for its thread.
Where to go from here
The algorithm rewards what a well-run account does anyway: consistent posts worth discussing, early and useful replies, fast answers to your own commenters, and zero hostility. For the full operating system around these rules, read the crypto Twitter marketing playbook. If you are building from a small account, start with how to grow a crypto project on X. Threads, which tend to earn the shares and bookmarks the model values, are covered in writing crypto Twitter threads.
FAQ
Does the X algorithm suppress crypto content?
The published ranking formula is a generic weighted sum with no crypto-specific weight. Spam and visibility rules are a separate layer, and some of those rules are not published. If you suspect a problem, check the Under the Hood tool for labels on your account. Paid promotion of crypto is restricted separately, see disclosure rules for paid posts.
Are the 2023 Twitter algorithm weights still accurate?
No. They describe the 2023 heavy ranker. The system X open-sourced in 2026 uses a different model and publishes different values, for example a reply weight of 5.0 instead of 13.5 and a report weight of -234 instead of -369. The direction is the same: conversation helps, negative feedback hurts.
Do links in the post body get penalized on X?
The published weights show no negative weight for opening a link, and we found no link penalty in the published scoring code. We still put links in the first reply because the post itself is scored on engagement that happens on X.
How quickly should a project reply to comments?
Within the first 30-60 minutes of posting where you can, and to every substantive comment. The 2026 code does not reward author replies directly the way the 2023 weights did, but active threads and mutual relationships are what the current reply predictions and boosts favor.