Inventory, Signals, Predictions, and Relevance Scores: How Meta's Four-Stage Architecture Filters and Ranks Social Feeds

Users frequently claim that social media algorithms are actively listening to their conversations or arbitrarily hiding specific posts to suppress their reach. The evidence contradicts this personalized-sabotage narrative entirely. Meta’s actual ranking architecture operates as a strict, indifferent mathematical pipeline that processes billions of inputs without evaluating their semantic meaning.[1]

Rather than a single omniscient judge, the system that curates the Facebook Feed relies on a decentralized, four-stage computational process. This architecture breaks down into Inventory, Signals, Predictions, and Relevance Scores. It is designed to filter an overwhelming volume of daily content into a digestible, highly personalized stream.[1]

The scale of this sorting operation is difficult to overstate. Over 3.07 billion monthly active users rely on the platform, and each session requires the servers to evaluate a massive pool of potential posts. The system must execute this four-stage pipeline in milliseconds every time a user refreshes their screen.[2]

"The algorithm's role is to sort, prioritize, and deliver third-party content to each user—not to create new content of its own," according to a 2026 legal filing regarding platform algorithms.[5]

It merely reflects human judgments about what kind of content to prioritize based on historical interaction vectors, rather than acting as an independent author. The architecture is entirely dependent on the data it receives from the initial candidate pool.[5]

Gathering the Candidate Pool

The first stage of the architecture is Inventory, which acts as the foundational data retrieval phase. When a user opens the application, the system queries its databases to gather every single piece of content eligible for display. This pool represents the absolute maximum universe of posts that could theoretically appear.[1]

Historically, this inventory consisted almost entirely of connected content. Following a major strategic pivot in 2022, the composition of the candidate pool shifted dramatically. Up to 50 percent of the available inventory now comprises recommended content pulled from accounts the user does not actively follow.[1][3]

A typical user might have anywhere from 1,500 to over 150,000 potential posts waiting in their inventory at any given moment. Displaying this content in a raw, chronological stream would render the platform unusable. The subsequent stages exist entirely to reduce this massive candidate pool into a curated shortlist.[3]

This shift away from the pure social graph fundamentally alters how information propagates. Because the inventory now includes millions of recommended posts, the system requires a more sophisticated method to determine what actually matters to the viewer. That requirement drives the second stage of the pipeline.[2]

Extracting the Data Vectors

Once the inventory is established, the architecture moves to the Signals stage. Here, the system evaluates hundreds of thousands of distinct data points attached to both the content and the user. These signals provide the raw mathematical context required for the prediction models to function accurately.[1]

Signals are broadly categorized into explicit, implicit, and relationship vectors. Explicit signals include active user choices, such as liking a post or leaving a comment. Implicit signals track passive behavior, measuring exactly how long a user hovers over a video or how quickly they scroll past a specific text update.[1]

Relationship signals carry significant weight in the modern architecture. The system monitors how frequently a user interacts with a specific person or page, tracking messaging history, profile visits, and shared comment threads. A post from a frequently messaged friend will generate a much stronger relationship signal than an update from a distant acquaintance.[2]

Environmental and hardware signals also feed into the pipeline. The architecture registers the user's current device type, their connection speed, and the time of day. A high-definition video might generate a strong signal on a fast Wi-Fi connection, but the same video will be downgraded if the user is on a slow cellular network.[4]

The Probability Engine

The third stage, Predictions, is where the actual machine learning models execute their core function. The system takes the thousands of signals gathered in the previous step and feeds them into a massive array of predictive algorithms. These models do not evaluate whether a post is objectively good or bad.[1]

Instead, they compute the exact mathematical probability that the user will take a specific action. Meta runs over 100 different prediction models simultaneously during this phase. One model calculates the likelihood of a user leaving a comment, while another predicts the probability of them sharing the post via a direct message.[2]

"We use ML to predict which content will matter most to each person to support a more engaging and positive experience," Meta's engineering team explained in a technical disclosure.[2]

The models rely heavily on historical data, assuming that a user who frequently watches videos to completion will likely do so again when presented with similar formats. The system continuously updates these probabilities based on real-time feedback from the user's ongoing session.[2]

Crucially, the prediction targets diverge based on the specific surface being evaluated. For the traditional Feed, the models heavily optimize for active conversational friction, predicting whether a post will spark a multi-thread discussion. For short-form video surfaces, the models pivot to optimize for passive retention, prioritizing watch-through and replay rates.[1]

This divergence means that the exact same user behavior yields entirely different algorithmic outcomes depending on the format. A text post that generates a long debate scores exceptionally well in the Feed models, while a video that loops seamlessly without generating a single comment dominates the short-form prediction engines.[2]

Calculating the Final Sort

The final stage of the architecture is the Relevance Score. After the 100-plus prediction models output their individual probabilities, the system must consolidate those numbers into a single, actionable metric. The Relevance Score is the ultimate mathematical verdict on a piece of content's value to that specific user.[1]

Every single post in the initial inventory receives a unique Relevance Score. The system then sorts the entire candidate pool in descending order based on this final number. The post with the highest score appears at the very top of the screen, while lower-scoring items are pushed further down the feed.[1]

This hyper-personalized scoring mechanism explains why organic reach for commercial pages has plummeted to an average of 1.65 percent in recent years. If a brand's post fails to generate the specific signals required to trigger high-probability predictions, it receives a low Relevance Score and is effectively rendered invisible.[2]

The architecture also applies a final layer of diversity rules before rendering the feed. If the top five posts all happen to be videos from the same creator, the system will artificially demote some of them to ensure a mix of text, images, and different authors. This prevents the feed from becoming monotonous.[1]

The Limits of the Architecture

Despite its computational power, the four-stage pipeline possesses inherent limitations. Because the system optimizes strictly for predicted engagement, it struggles to differentiate between genuine interest and outrage-driven friction. A post that angers users into commenting still generates the active signals the prediction models are designed to reward.[5]

To counter this, Meta has introduced direct user surveys to train the models on qualitative value rather than just quantitative engagement. Users are periodically asked whether a specific post was genuinely worth their time. These survey responses act as a corrective weight against the raw engagement predictions.[1]

The architecture also faces challenges with cold-start inventory, which consists of new posts from unknown creators that lack historical interaction signals. To solve this, the system briefly injects these posts into a small number of feeds to gather initial data. If the early signals are strong, the post's Relevance Score cascades upward.[2]

Ultimately, the four-stage architecture represents a triumph of scale over curation. By breaking the ranking problem down into Inventory, Signals, Predictions, and Relevance Scores, Meta has built a system capable of sorting human communication at a planetary level. It is a pipeline built entirely on probability.[1]

Understanding this mechanism strips away the mysticism surrounding social media visibility. The algorithm is not a discerning editor making subjective choices about quality. It is simply a highly efficient sorting machine, relentlessly calculating the odds of human interaction one data point at a time.[7]