Most of what circulates about the X algorithm is a list of 2023 multipliers: a reply counts about 27x a like, that sort of thing. Those numbers were real, for the 2023 Scala codebase (twitter/the-algorithm). They no longer describe the live system. In January 2026 xAI published a near-total rewrite as open source, built around a different ranker, so the old multipliers have been superseded rather than corrected.
Everything below is read straight from that published code, not inferred and not leaked. You can open the files and check it yourself.
Is the X algorithm actually open source?
Yes, and more completely than in 2023. The code lives in xAI's open-source x-algorithm repository, the official xAI organization, under an Apache-2.0 license, written in Rust and JAX, first published on January 20, 2026 with releases on a monthly cadence since.
The ranking model at the center of it is a transformer called Phoenix, derived from Grok-1. It predicts how likely you are to take each possible action on a given post, and those predictions get combined into a single score. Which means the weights that matter are not secret and not estimated. They are constants in a file.
What does the X algorithm reward most?
Here are the positive action weights as of the November 2026 release, taken directly from home-mixer/params/param.rs, ranked high to low. The weight is what the ranker multiplies a given action by when it scores a post for you.
| Action | Weight | What it tells the ranker |
|---|---|---|
| Copy-link share | 20.0 | Someone grabbed the link to send it somewhere else. The single strongest signal. |
| Reply | 5.0 | Effectively about 20.0 on an original post among mutual-follow authors. Real conversation. |
| Quote | 5.0 | Worth enough to react to in public with your own take. |
| DM-share | 5.0 | Privately sent to a specific person. A high-intent recommendation. |
| Follow the author | 4.0 | The post converted a viewer into a subscriber. |
| Generic share | 2.0 | Shared through the share sheet without copying the link. |
| Repost | 1.0 | A low-effort endorsement. |
| Like | 0.66 | Cheap approval. Raised from 0.5 in November 2026, still about an eighth of a reply. |
| Click and dwell | 0.4 | Opened and lingered. The weakest positive worth counting. |
One caveat on the exact figures: the repository is a periodic snapshot, rewritten roughly monthly, and a few files are withheld (Grok prompts and parts of the safety rules). Several of these values are config-driven and can shift between releases, so treat the numbers as the shape of the system rather than permanent constants. We log every release that moves a weight in the X algorithm changelog.
Read the order, not just the numbers. A copy-link share is weighted about thirty times a like. The actions the code prizes are the ones that move a post off X and in front of a specific person: copying the link, sending it in a DM, replying to it. Passive approval barely registers. If a post is not worth privately sending to someone, the ranker treats it as close to invisible.
Why is a report weighted -234 in the X algorithm?
The negative weights look alarming on their own:
| Action | Weight | What it tells the ranker |
|---|---|---|
| Report | -234 | The strongest negative signal in the file. |
| Mute | -58.8 | Make this author stop reaching me. |
| "Not interested" | -52.272 | Explicit rejection of the post. Strengthened from -47.52 in November 2026. |
| Block | -31.2 | Hard cut, but personalized. |
Here is the nuance almost nobody explains, and it is the whole key to reading these numbers: the weights multiply the predicted probability of an action, not a raw count. The point is stated twice in the params file.
A report is roughly 1000x less likely to happen than a like. So a -234 weight is not "one report cancels 234 likes." It scales a tiny probability back up into a score that matters. The large number exists precisely because the baseline chance of a report is so small. Compare it to a like, which is common and carries a weight of only 0.66.
Reports and blocks are also personalized. The model predicts how likely you are to report a post based on users similar to you. So coordinated mass-reporting mostly suppresses a post among people who resemble the reporters, not across the whole platform. It is a dampener on the wrong audience, not a global kill switch. (There is a separate layer for that, covered below.)
Do replies and reposts get less reach than original posts on X?
Yes, and the code is explicit about it. Several structural levers sit on top of the per-action weights and decide how far a post can travel at all.
- Original top-level posts beat replies and reposts. An out-of-network discount hits replies and reposts even when they go to your own followers. The feed is built to surface fresh authored posts, not threads of reactions.
- There is no long tail. Posts older than about 48 hours are hard-filtered out of ranking. A post does not quietly keep earning reach for a week. It has roughly two days.
- The cold-start boost is narrow. It fires only in a tight window: under 2 hours old, author under 50k followers, under 200 impressions. That is the help a new or small account gets, and it closes fast.
- Spacing matters. An author-diversity decay suppresses your second and third post inside a single feed render, so the same viewer does not see three of your posts in a row. Posting a burst back-to-back works against you. Space them out.
This is the opposite of a completion-rate platform. Unlike TikTok, which ranks on video completion, X rewards an authored post that earns a reply and a private send inside its first couple of hours.
Should you put links in a post on X?
Keep the whole payload inside the post. External links are neutral-to-weak in the ranking code, and worse, they pull clicks off X and onto a page the ranker cannot see, which starves the strong on-platform signals (the reply, the quote, the copy-link share) that actually move a post.
X is text-and-image-first. It is not a watched-to-the-end video platform, and it is not a save-and-revisit feed the way Instagram's single strongest signal, sends per reach, shapes its own ranking. On X the win condition is a post complete enough to stand on its own, provoke a reply, and be worth forwarding in a DM.
The visibility layer that sits above the score
One more thing the code makes clear: the per-action score is not the final word. A separate visibility-filtering layer runs spam and credibility classifiers that can drop a post regardless of how well it scores. A post can earn a strong rank and still be held back if it trips those filters. Ranking decides order among eligible posts. This layer decides eligibility.
Where this fits with the other platforms
X is the only major platform whose ranking is literally readable, which makes it the clearest window into how all of these systems think. The signals differ by platform, but the pattern rhymes: each one rewards a specific high-intent action and discounts passive approval. For the full map, see how social media algorithms work in 2026.
Sources
Everything above is confirmed from published source code, not leaks or estimates. Primary source: xAI's open-source X ranking code, specifically the action weights in home-mixer/params/param.rs and the scoring logic in xai-value-model/scoring.rs. Secondary coverage and analysis appeared at CryptoBriefing, The Deep View, Towards AI, and PPC Land.
X ranks on predicted actions, not counts: a copy-link share is weighted about thirty times a like, the -234 report penalty is huge only because reports are rare, and an original post that earns a reply and a private send in its first two hours is the entire game.
We read the code so your posts don't have to guess
Protegee runs a standing research function that tracks how every platform ranks content, from X's open-sourced weights to the signals TikTok, Instagram, Facebook, and YouTube publish, and builds those learnings into every post it generates. The durable rule never changes: earn the first second, hold attention to the end, and make it worth privately sending.
Sign up