X Just Published the Scoring Weights Behind Its Algorithm (Here's What Actually Drives Reach)

A like is worth 0.5. Someone quietly copying your post and sending it to one other person is worth 20.

Those are not estimates. They are the default values sitting in a file called param.rs in X's public repository, and until August 13, 2026, nobody outside the company had seen them.

This matters more than the last round of algorithm coverage suggested, and less than the "go viral with this one trick" posts are claiming. Here is what actually shipped, what the numbers say, and what three of the most widely repeated claims about them get wrong.

What Actually Shipped in August 2026

First, a correction that most coverage this week got wrong: X did not just open source its algorithm. That happened on January 20, 2026, when xAI published the x-algorithm repository, a complete rewrite of the 2023 Scala system in Rust and Python.

What January gave the public was the architecture. You could see the pipeline, read the retrieval logic, and identify which user actions the ranking model predicts. What you could not see was what any of those actions were worth.

The August 13 release added three things:

1. The scoring weights. The actual numeric values used to blend predicted actions into a single score for each post.
2. The visibility filtering system. The code that decides whether a post is allowed to appear at all, which is a separate system from ranking.
3. A label transparency tool called Under the Hood. It surfaces aggregate statistics about the visibility-limiting labels applied to your own account and posts, piloting with accounts at least a year old.

The distinction between architecture and weights is the whole story. Knowing that a system predicts the probability of a reply tells you very little. Knowing that a reply is weighted at 5.0 while a like is weighted at 0.5 tells you that a decade of engagement advice was aimed at the wrong action.

How the For You Feed Assembles a Timeline

Before the numbers are useful, the pipeline needs one paragraph of context.

Posts arrive from two sources. In-network candidates come from thunder, which holds recent posts from accounts you follow in memory. Out-of-network candidates come from phoenix retrieval and simclusters, which surface posts from accounts you do not follow.

Both sets are ranked by the same model. Phoenix reads the viewer's recent engagement history and predicts, for each candidate post, the probability that this specific viewer takes each possible action. A scorer multiplies each probability by a fixed weight and sums them. Three adjustments follow. Then the list is sorted and cut.

Separately, and only after ranking is finished, a different service decides whether each post can be displayed at all.

That separation between ranking and visibility is the single most important structural fact in the repository, and we will come back to it.

The Scoring Weights

These are the default values in the published code. X notes that live production values can differ during experiments, but these are the real starting points.

Positive actions:

1. Share via copy link: 20.0 - the highest weighted action in the entire system
2. Reply: 5.0 - rising to 20.0 on original posts from a mutual follow
3. Quote: 5.0
4. Share via DM: 5.0
5. Follow author: 4.0
6. Share: 2.0
7. Repost: 1.0
8. Favorite (like): 0.5
9. Post click: 0.4
10. Open link: 0.2
11. Photo expand, video open, video quality view, quoted post click: 0.05 each
12. Post unexplored: 0.02, applied to in-network posts only by default
13. Continuous dwell time: 0.004
14. Dwell and profile click: 0.0, currently switched off

Negative actions:

1. Report: -234.0
2. Mute author: -58.8
3. Not interested: -43.2
4. Block author: -31.2
5. Not dwelled: -0.02

The Caveat Almost Every Writeup Is Missing

There is a comment sitting directly above these values in the source explaining that the weights encode two things at once: how much an action is valued in ranking, and how often that action typically occurs across the network.

Negative feedback is rare. So a report is not weighted at -234 because one report is 468 times worse than one like. It is weighted that way because the predicted probability of a report is normally close to zero, and the coefficient has to be large for a tiny probability to influence the final score at all.

The same logic applies in reverse to copy-link shares. These are scaling decisions calibrated against base rates, not a moral hierarchy of user actions. Anyone presenting the negative weights as a punishment ranking has misread the file.

What the Weights Reward

Forwards Beat Likes by 40x

The highest value signal X can observe is one person copying a post and sending it to one specific other person. Not broad reach. Not a viral spike. One forward.

Writing for likes and writing for forwards are different jobs. A post optimized for likes seeks agreement from a crowd, which favors the broadly agreeable and the emotionally simple. A post optimized for forwards has to be useful enough that a reader thinks of a specific colleague while reading it.

For AI companies, this is unusually good news. A benchmark result, a well-explained architectural tradeoff, or an honest cost breakdown of running inference at scale is exactly the kind of content one engineer sends to another. It is also the kind of content that historically underperformed on like counts, which made it look like a failure under the old scoreboard.

Replies Are Worth 10x Likes

At 5.0 against 0.5, a post that earns a considered reply is worth an order of magnitude more than a post that earns a passive tap.

This is not an argument for engagement bait. Low-effort questions invite the "not interested" signal, weighted at -43.2, from viewers who recognize the pattern. It is an argument for leaving something genuinely arguable in the post. A specific claim someone can dispute with their own experience will outperform a safe observation everyone nods at.

Mutual Follows Are a Structural Multiplier

Buried in the scorer is a condition requiring three things simultaneously: the post is not a reply, it is not a repost, and the author is a mutual follow of the viewer. When all three hold, the reply weight for that viewer jumps from 5.0 to 20.0.

Original posts from mutual follows are the single most favored content type in the ranking system. For a founder building an audience, this reframes following back from a courtesy into a compounding mechanic. It also means the boost never applies to replies or reposts, only to original posts.

Your Replies and Reposts Are Discounted Even to Your Own Followers

Out-of-network posts get multiplied by 0.75, which most coverage noted. What almost nobody caught is a flag called EnableOonRescoreForInNetworkRepliesRetweets, which defaults to true and applies that same 0.75 discount to replies and reposts from accounts the viewer already follows.

Your own replies and reposts are structurally discounted relative to your own original posts, even when shown to people who chose to follow you. Commenting on other accounts builds relationships and puts you in front of new audiences. It will not do what publishing something of your own does.

Converting a Stranger Into a Follower Is Worth 8 Likes

Follow-author sits at 4.0. A post that makes someone want the next one is worth eight times a post that gets tapped and forgotten.

For AI founders, this favors serialized thinking over one-off takes. If a post reads like an excerpt from an ongoing body of work rather than an isolated observation, it earns the action the system pays the most attention to after forwards and replies.

Volume Decays, It Does Not Cap

Author diversity applies a multiplier calculated as (1 - floor) * decay^k + floor, with a decay of 0.5 and a floor of 0.25, where k is the rank of that post among your own posts in the same timeline slate.

In practice: your first post scores at full value, the second at roughly 0.625, the third at 0.44, and everything after that trends toward a floor of 0.25. Posting more is not free. It is also not fatal. A fourth post still carries a quarter of its score, which is meaningfully different from being suppressed.

A Post Has 48 Hours and Then It Is Gone

AgeFilter removes anything older than 48 hours before scoring even runs. This is a hard cutoff, not a decay curve. There is no long tail on X. A post gets two days to exist and then it stops being a candidate entirely.

For coordinated launches, this changes the shape of a campaign. Creator content published across a two week window produces no cumulative timeline effect. The same content concentrated into a 48 hour window competes as a single unified moment.

Three Things Everyone Is Getting Wrong

1. The Link Penalty Is Not in the Code

The most repeated claim about X ranking is that external links take a 30 to 50 percent reach penalty. That penalty does not exist in this codebase.

The relevant parameter is OpenLinkWeight and its value is positive, at 0.2. The only URL-related suppression in the repository is a drop rule for URLs flagged as malicious, and it applies only to out-of-network distribution.

Links genuinely do underperform, but the mechanism is different and the difference matters. A link click contributes 0.2 while a copy-link share contributes 20 and a reply contributes 5. Clicking out also ends engagement with the post. You are trading a high value action for a cheap one, not absorbing a punishment.

Putting the link in a reply remains the right call. The reasoning everyone gives for it is fiction.

2. The Negative Weights Are Not a Punishment Ranking

Covered above, but worth restating because it is being repeated everywhere: the negative coefficients encode rarity as well as severity. Reading -234 as "reports are catastrophic" and -31.2 as "blocks are four times less bad" misunderstands what the number is doing.

3. Shadowbanning Is Real, But It Is Not About Individual Posts

Here is the mechanic that explains more unexplained reach collapse than every weight in the file combined.

Visibility filtering runs two separate rule sets. The base set applies to everyone. A second set applies only when a post is being recommended to someone who does not follow the author, and those rules can only drop a post, never promote it. They include spam detection at high recall, do-not-amplify labels, abusive content at high recall, impersonation, and compromised account flags.

The consequence: a post can be fully visible to your existing followers while being invisible to everyone else. Nothing in your analytics will explain it, because the cause is a label attached to your account rather than anything about that specific post.

This is what people have called shadowbanning for six years. The code confirms the mechanism exists, confirms it is label-driven, and, with Under the Hood, now gives account owners a way to check whether it applies to them.

What Is Not in the Repository

Credibility requires naming the gaps.

X withheld the Grox prompts, which are the LLM instructions used to classify post text and media. It also withheld some rule definitions and the trained model weights.

So the arithmetic wrapped around the model is now fully public. The model's taste is not. Any claim that the algorithm rewards a particular writing style, sentence length, or topic remains unverifiable speculation. What can be verified is which user actions the system pays for, and by how much.

What This Means for AI Companies Running Creator Campaigns

The weights validate something we have argued for years at Clickstrike: follower count is close to the worst available proxy for whether a creator partnership will work.

A creator with 500,000 followers whose audience is largely casual scrollers produces likes, and likes are weighted at 0.5. A creator with 40,000 followers whose audience is mostly ML engineers and technical decision makers produces replies, quotes, and private forwards, which the system weights at 5.0, 5.0, and 20.0. The second creator is not marginally better under this scoring model. They are an order of magnitude better, and the code now says so explicitly.

This is why our vetting process rejects roughly 70 percent of creator applicants on audience composition and domain expertise rather than reach. When we ran a developer tool campaign through creators whose regular content covered Docker tutorials, Kubernetes walkthroughs, and REST API builds, the outcome was over 1 million YouTube views and 23,700 social engagements, and the client cited that channel as consistently delivering the lowest cost per registration of anything in their portfolio. The full breakdown is in our case studies.

Three practical implications for AI companies planning X activity:

1. Judge creator partnerships on reply and quote volume, not like counts. Both are worth 10x a like in the ranking system, and both signal that the audience is technical enough to have an opinion.
2. Compress launch activity into a 48 hour window. The hard age cutoff means content spread thin across two weeks never competes as a coordinated moment.
3. Prioritize creators whose audiences overlap with each other and with you. The mutual follow boost is structural, and a network of technical accounts that genuinely follow each other compounds in a way a list of unconnected large accounts does not.

For a fuller treatment of how to select and manage those partnerships, see our guide to AI influencer marketing.

Frequently Asked Questions

Did X just open source its algorithm? No. The algorithm was open sourced in January 2026. The August 13, 2026 update added the scoring weights, the visibility filtering code, and the Under the Hood label transparency tool.

What is the most valuable action on X according to the code? Sharing via copy link, weighted at 20.0. A like is weighted at 0.5.

Do external links reduce reach on X? Not through a penalty. The link click parameter is positive at 0.2. Links underperform because clicking out contributes far less than a reply or a forward, not because posts containing them are suppressed.

How long does a post stay in the For You feed? 48 hours. Posts older than that are filtered out before scoring.

Can I see if my account has been limited? The Under the Hood tool shows aggregate statistics on visibility-limiting labels attached to your account and posts. It is piloting with accounts at least a year old before wider rollout.

How often does the repository change? X committed to updating it roughly every four weeks with developer notes explaining what changed.

The Part Worth Sitting With

The weights themselves are interesting. What they reveal about the last decade is more interesting.

An entire industry built its measurement around the cheapest signal in the system. Like counts became the default proxy for content performance, creator quality, and campaign success, and the file establishing that likes are worth 1/40th of a private forward has been sitting in a public repository since January.

For AI companies, whose best content tends to be technical, specific, and forwarded between engineers rather than broadly liked, this is the rare case where the incentive structure was already pointing at the right thing. The scoreboard was just measuring something else.

If you want help building creator programs and earned media coverage that produce the signals this system actually pays for, Clickstrike works exclusively with AI and technology companies. We have run 1,200+ campaigns through a network of 500+ vetted technical creators and secured 8,250+ media placements for AI products.