X Open-Sourced More of Its For You Algorithm. Here’s What Creators Can Actually Learn From It
X’s August release exposes substantially more ranking, filtering and visibility code and pilots an “Under the Hood” label report. The most viral interpretation of the numbers, however, is exactly the one X says is wrong.
PLATFORMS
Muhammad Ali Akbar
8/24/2026
For creators, the biggest X algorithm update of August 2026 is not a secret formula for going viral.
It is something more useful: a clearer map of the system that decides which posts are candidates for the For You feed, how X predicts a viewer’s likely actions, and how separate visibility systems can prevent a post from being recommended at all.
On 13 August, X expanded its public algorithm repository with key configuration parameters, visibility-filtering code, the systems that produce many visibility labels, Phoenix model training and serving code, and additional retrieval components. X also paired the release with a pilot “Under the Hood” tool designed to show users aggregate statistics about visibility-impacting labels applied to their account and posts.
The release is unusually detailed for a major social platform. But it also produced a predictable wave of oversimplified posts claiming that one action is “worth” a fixed number of likes. X added a clarification the next day saying that interpretation is wrong.
The important correction
The published action weights multiply X’s predicted probability that a particular viewer will take an action. They do not multiply the raw number of likes, replies, reports or shares a post already received. X explicitly warns that a statement such as “one report cancels out 468 likes” is an incorrect reading of the system.
The For You feed starts by finding candidates, not by counting likes
X’s public documentation says a For You request pulls posts from two broad sources. “In-network” candidates come from accounts the viewer follows. “Out-of-network” candidates are discovered using machine-learning retrieval and community-similarity systems, including Phoenix retrieval and SimClusters. Those candidates are then filtered, scored and reranked for the individual viewer.
This matters because there is no single global score attached to your post that every user sees. The system is viewer-specific. Phoenix reads the viewer’s recent engagement history and predicts how likely that viewer is to perform many possible actions on each candidate post.
The predicted actions include obvious engagement signals such as likes, replies, reposts and quotes, but also link clicks, profile clicks, video opens, dwell-related measures, following the author and negative actions such as “not interested,” muting, blocking and reporting. X then combines those predictions using weights into a ranking score.
Why the viral “engagement exchange rate” is wrong
Suppose a published parameter gives reports a much larger negative weight than likes receive as a positive weight. It is tempting to divide one number by the other and conclude that a single report erases hundreds of likes. That is not what the formula is doing.
X’s repository describes the ranking score as a weighted sum of predicted action probabilities. Conceptually, the system asks questions such as: How likely is this viewer to like this post? How likely are they to reply? How likely are they to report it? Each predicted probability is scaled by a weight, and the resulting values are combined. The inputs are predictions for this viewer-post pair, not a scoreboard of raw engagement already collected.
That distinction is critical for creators. It means the correct lesson is not “chase replies because each reply equals N likes.” The lesson is that the system is designed to value a portfolio of predicted behaviors, including attention and negative feedback, and that those predictions depend heavily on the viewer’s own history and context.
Three adjustments happen after the initial weighted score
X’s README describes additional adjustments after the main RankingScorer. First, author diversity reduces repeated posts from the same author so one account does not dominate a feed. Second, an out-of-network discount reduces the score of posts from accounts the viewer does not follow, as well as certain replies and reposts. Third, a new-author boost can lift posts from authors whose impressions are below a threshold toward a target position.
Those mechanics reinforce a point that creators often miss when they obsess over one engagement metric: feed ranking is a pipeline. A candidate has to be retrieved, survive filters, score well relative to other candidates, avoid excessive repetition and pass visibility checks. A high-performing signal at one stage does not guarantee distribution at the next.
Freshness is built into the published candidate pipeline
One concrete parameter is especially relevant to publishing cadence. The public repository describes a pre-scoring AgeFilter that removes posts older than 48 hours from this For You candidate flow. That does not mean every surface on X erases a post after two days, but it does show that recency is explicitly enforced in the published For You pipeline.
For creators, the practical implication is not to spam more frequently. It is to understand that the main recommendation window is time-sensitive: strong posts need to earn useful viewer signals while they are still eligible for fresh recommendation inventory.
Ranking and visibility are two different systems
This may be the most important conceptual insight in the entire release. X separates ranking from visibility filtering. Ranking decides the order of eligible posts. Visibility filtering can decide whether a post is shown normally, placed behind an interstitial or dropped for a particular viewer or recommendation context.
The open-source labeling path includes systems that classify posts and accounts for issues such as spam, adult content, violence and other policy or integrity signals. The visibility layer can also incorporate the viewer’s own blocks, mutes, age, settings, country and relationship to the author. In some cases a post can be excluded from recommendations to non-followers while remaining available to followers.
That is a more precise way to discuss what users often call a “shadowban.” A creator can experience reduced recommendation reach because a label or rule changes eligibility without the post being deleted from the platform. But the term “shadowban” remains too broad to explain why reach changed in any individual case.
What “Under the Hood” can tell creators
X says its new Under the Hood transparency tool is intended to show aggregate statistics about visibility-impacting labels on a user’s account and posts. TechCrunch reported that the initial pilot allows qualifying users to download a JSON file covering labels applied over the previous calendar month, with availability beginning among a test group rather than every account at once.
According to the launch reporting, the pilot initially targets accounts at least a year old, and the download option is for users who have posted at least 10 times during the previous month. Those eligibility details may change as the rollout expands, so the Creator Internet should date-stamp this section and re-check it when updating the article.
The useful workflow is diagnostic rather than conspiratorial: if the report shows a visibilityimpacting label, a creator can compare that label with the public code and documentation to understand the described effect. If no such label appears, that does not prove the algorithm “likes” the account; ordinary ranking competition, topic relevance, predicted engagement, freshness and audience fit can still explain weak distribution.
What X still does not publish
Open source does not mean the entire production system is exposed. X says some material is withheld to reduce gaming, including Grox prompt files and some Botmaker rules. The company also regularly runs experiments on portions of traffic, while repository defaults are intended to represent primary production values rather than every experimental condition at every moment.
The repository itself therefore warns against treating any screenshot of a parameter file as eternal truth. Creators should expect weights, thresholds, filters and models to change. A durable strategy should be based on the system’s design principles, not a spreadsheet of today’s constants.
Seven practical lessons creators can use without gaming the algorithm
Optimize for a real viewer, not a universal formula. X is predicting actions for individual viewerpost pairs, so audience fit matters more than a mythical platform-wide score.
Create posts worth spending time on. The published model considers attention and dwell-related predictions alongside explicit engagement.
Avoid manufacturing negative feedback. “Not interested,” mutes, blocks and reports are modeled as negative actions, and some safety/integrity labels can also affect recommendation eligibility.
Do not flood the feed with near-identical posts. X describes repeated-author diversity adjustments and duplicate/conversation filters, so quantity alone is not a reliable distribution strategy.
Use the first 48 hours intelligently. The published For You pipeline contains a 48-hour age filter for candidates; publish when your actual audience can respond, then continue the conversation rather than endlessly reposting the same item.
Treat reach drops as a diagnosis problem. Check content quality, topic-audience fit, recency and account/post labels before declaring a shadowban.
Archive observations over time. If Under the Hood becomes available to your account, save dated exports and compare them with reach changes instead of relying on screenshots or anecdotes from other users.
What creators should not do
Do not buy or coordinate fake engagement to “hit the weights.” The model includes integrity and negative-feedback systems, and the open-source release is specifically incomplete where publishing more detail could facilitate gaming.
Do not turn published weights into a likes/replies/report currency converter. X explicitly says that math is wrong.
Do not assume a post that remains visible to followers is equally eligible for recommendations to non-followers. The visibility rules can treat those contexts differently.
Do not cite the repository without a date. The code and defaults can change, sometimes quickly.
The bigger story is transparency, not a growth hack
The creator-economy value of this release is that X has made recommendation analysis more falsifiable. Instead of only guessing which signals might matter, researchers and creators can inspect a meaningful portion of the ranking and visibility stack, see how X describes the relationship between retrieval, scoring and filtering, and—in the pilot—compare account-level label outcomes with the code.
That still does not turn a complex recommender system into a recipe. In fact, the August 14 clarification is a warning about exactly that temptation. The more useful insight is structural: X is trying to predict valuable future behavior for each viewer while separately controlling whether content is eligible to be shown. Good creator strategy therefore remains rooted in relevance, originality, viewer satisfaction and avoiding reasons for people or safety systems to reject the post.
Bottom line
X’s August 2026 open-source expansion gives creators an unusually detailed look at the For You feed. It confirms that distribution is built from multiple stages—retrieval, filtering, multi-action prediction, weighted ranking, diversity adjustments and separate visibility rules—and that a new transparency tool can expose some visibility-impacting labels to eligible users.
What it does not provide is a magic equation for viral reach. If someone tells you a reply is worth a fixed number of likes or one report automatically cancels a fixed amount of engagement, X’s own documentation says they are reading the weights incorrectly.
© 2026 The Creator Internet. All rights reserved.