For years, creators and brands on X operated without a clear map of how the platform decided what content deserved attention. The For You feed felt unpredictable. A carefully written post could stall at a few hundred views while a simpler post from a larger account traveled much farther. People filled the information gap with theories about shadowbans, secret boosts, and hidden ranking factors that no one could verify.
Most users optimized for the only metric that was easy to see: likes. The assumption was that more likes equaled more reach. In reality, likes were only one weak signal among many, but without transparency there was no way to know which actions the system actually valued most. This led to widespread frustration and inefficient content strategies across the platform.
The lack of visibility also affected trust. When reach fluctuated without explanation, creators and marketing teams often assumed the worst. Some believed the system was deliberately suppressing certain voices. Others simply accepted that growth on X was largely a matter of luck or existing follower size. The black box nature of the ranking system made consistent progress difficult.
Every Day: The Cost of Guesswork
The same pattern repeated daily for both individual creators and brands. Thoughtful threads and original insights underperformed. Engagement tactics that once appeared effective began delivering diminishing returns. People experimented with posting frequency, timing, hashtags, and formats without knowing whether those changes actually aligned with how ranking worked.
Brands faced additional pressure. Campaigns that looked strong on paper often produced disappointing organic reach. Marketing teams measured success through impressions and likes while the underlying ranking logic remained opaque. This created a cycle of reactive adjustments rather than strategic improvement.
Over time, the absence of clear rules encouraged short-term tactics. Some accounts leaned into engagement bait. Others posted more frequently in the hope of staying visible. Many simply reduced their activity because the effort no longer felt proportional to the results. The daily cost of operating without reliable information was both creative energy and missed growth opportunities.
One Day: The Ranking Code Became Public
In August 2026, X released a major expansion of the open-source repository at github.com/xai-org/x-algorithm. The update significantly increased the amount of publicly available code related to the For You timeline, including ranking parameters, visibility filtering systems, and configuration weights. At the same time, the company launched a pilot version of the “Under the Hood” tool that allows eligible users to see labels applied to their accounts and posts.
This release was different from earlier transparency efforts. It was not limited to high-level descriptions or research papers. The repository contains production-oriented code covering candidate sourcing, scoring, ranking, and filtering. At the center of the ranking process is Phoenix, a Grok-based transformer model designed specifically for recommendation.
The availability of this code changed the conversation. For the first time, outsiders could examine how posts are scored, which actions carry the most weight, and how ranking decisions are separated from visibility filtering. While full model weights remain private and some infrastructure details are abstracted, the core architecture and scoring logic are now open for inspection.
How X Ranking Actually Works
X does not rank posts using a single vague relevance score. Phoenix predicts the probability of many different user actions on each candidate post. These predicted probabilities are then combined with explicit, tunable weights to produce a final ranking score. The calculation is handled by the RankingScorer component within the home-mixer system.
The fundamental formula is straightforward: Final Score equals the sum of each weight multiplied by the predicted probability of the corresponding action. Positive actions receive positive weights. Negative actions receive negative weights. This multi-action approach allows the system to distinguish between different forms of engagement rather than treating all interaction as equal.
Publicly visible default weights reveal clear priorities. Share via Copy Link carries a weight of 20.0. Reply sits at 5.0 and can increase substantially when mutual follows are involved. Share via DM and Quote also carry weights of 5.0. A follow that originates from the post is valued at 4.0. Retweet is weighted at 1.0, while a standard Favorite (Like) is only 0.5. Lower-weighted signals include clicks, open-link actions, photo expands, and video opens. Negative signals such as “Not Interested,” mute, block, and especially report pull the score down sharply.
Two architectural decisions are particularly important. First, the ranking transformer uses candidate isolation. During inference, candidates cannot attend to each other. Each post is scored only in relation to the viewer’s history and features. This design keeps scores consistent and independent of batch composition. Second, ranking and visibility are deliberately separated. Ranking determines order. Visibility filtering, driven by labels from spam detection, safety models, and enforcement systems, decides whether a post is allowed to appear at all, shown behind an interstitial, or dropped entirely.
What the Weights Really Reward
Once the weights are visible, the growth implications become clear. The ranking system places significantly higher value on conversation and sharing than on passive approval. A post that generates thoughtful replies, quotes, private shares, and follows will generally outrank a post that only accumulates likes, even if the like count is higher.
Early engagement velocity also matters. The system uses signals from the first period after publishing to decide whether to expand distribution further. Posts that fail to generate meaningful interaction early often remain limited in reach. Author diversity penalties further discourage rapid successive posting from the same account, applying a decaying multiplier to later posts within a short window.
Negative feedback carries outsized cost. Strong negative signals such as reports or widespread mutes can suppress not only the current post but also reduce the reach of subsequent content from the same account. The open weights make this trade-off explicit: the system is designed to amplify content that people actively engage with and to limit content that generates irritation or rejection.
What Creators Should Do to Grow on X
Creators who want stronger distribution should design posts specifically for the highest-weighted actions rather than optimizing primarily for likes. The ranking logic rewards content that people want to reply to, quote, share privately, or follow the author from. Writing with these outcomes in mind produces better alignment with how Phoenix scores posts.
Practical application starts with the structure and tone of the post itself. Content that ends with a genuine question, presents a grounded but incomplete idea, or offers a specific observation that invites addition tends to generate more replies. Staying active in the replies during the first 30 to 60 minutes and responding substantively further strengthens the signal, because author engagement compounds the conversation value the model detects.
Creators should also manage volume and relationships carefully. Posting with intentional spacing reduces the impact of author-diversity penalties. Building genuine mutual follows inside a niche amplifies the effective weight of replies. Keeping external links out of the main post when possible and placing them in a reply helps avoid unnecessary deboosts. Avoiding content that reliably triggers mutes, blocks, or reports protects long-term account reach. Over time, consistent focus on high-value actions produces more reliable visibility than chasing short-term vanity metrics.
What Brands Should Do to Capture Attention on X
Brands operate under the same ranking physics as individual creators, but the goals and constraints are different. Organic attention on X is earned by content that the model predicts people will reply to, quote, and share. Brands that treat the platform primarily as a broadcast channel for polished announcements often underperform relative to those that treat it as a conversation surface.
Effective brand strategy begins with content that invites participation rather than passive consumption. Original observations, customer stories, product insights framed as questions, or industry takes that leave room for response tend to generate higher-weighted engagement. Measuring success through replies, quotes, profile visits that convert to follows, and private shares provides a clearer picture of ranking alignment than impressions or likes alone.
Brands also benefit from operational discipline. Empowering employees and community members to participate as authentic voices expands the surface area for genuine conversation. Using the Under the Hood transparency tools helps identify whether any visibility-limiting labels have been applied so root causes can be addressed. Avoiding engagement bait that triggers “Not Interested” or reports protects long-term distribution. Focusing on content that keeps users on-platform and encourages high-value actions such as copy-link shares and thoughtful replies produces more sustainable attention than tactics optimized only for short-term visibility.
Conclusion
The open-source release of X’s For You ranking code removes a significant amount of previous guesswork. Phoenix’s multi-action prediction approach, the explicit scoring weights, candidate isolation, and the separation of ranking from visibility filtering are now publicly inspectable. Full model weights remain private and some infrastructure details are abstracted, yet the core logic is visible and actionable.
Creators and brands who study how ranking actually works gain a practical advantage. Those who align their content with the highest-weighted signals — replies, quotes, meaningful shares, and follows — while avoiding strong negative feedback will find the For You feed more navigable. The algorithm still decides what surfaces, but the map is no longer completely hidden.
What was once a black box has become a system whose priorities can be examined and responded to. The accounts and teams that read the ranking logic carefully and consistently produce content people actually want to engage with will be better positioned to build visibility that compounds over time.
