How Recommendation Systems Actually Work

People talk about "the algorithm" as if it were a judge handing down verdicts. It is closer to a sorting problem run several hundred million times a day, and the platforms have published enough about how they solve it that you do not have to guess.

Here is the shape of the thing, in the platforms' own words.

Every feed is a funnel, not a feed

Meta's engineering team published a walkthrough of the Instagram Explore ranking system in August 2023. The structure they describe is the standard industrial design, and it is worth reading closely because Instagram, Facebook, TikTok, YouTube, LinkedIn and Pinterest all run some version of it.

The system has four stages: retrieval, first-stage ranking, second-stage ranking, and final reranking. Retrieval exists because ranking everything is impossible. As Meta puts it, "In a world with infinite computational power and no latency requirements we could rank all possible content. But, given real-world requirements and constraints, most large-scale recommender systems employ a multi-stage funnel approach." The retrieval sources each "select hundreds of relevant items from a media pool of billions of items."

Then two ranking passes. A lightweight model scores thousands of candidates. A heavy model, which Meta calls a multi-task multi-label neural network, re-scores roughly the top 100. Google's YouTube recommendation paper from 2016 describes the same two-step: candidate generation winnows "the enormous YouTube corpus" down to "hundreds of videos that may be relevant to the user," and a separate ranking model puts those in order.

The practical consequence is the part most advice skips. Your post does not get evaluated on its merits. It gets pulled into a candidate pool for a specific person, competes against everything else pulled in for that same person, and either survives the cut or does not. Nobody is deciding whether your post is good. A model is estimating whether one particular human will respond to it, in about a hundred milliseconds.

The score is a prediction, and the prediction is about the viewer

Meta published the actual shape of the final score for Explore. It is a weighted sum of predicted probabilities that they call the value model:

Expected Value = Wclick * P(click) + Wlike P(like) - Wseeless P(see less) + etc.

Read that formula slowly, because it dismantles a lot of folklore. The inputs are probabilities of things that have not happened yet. The weights are business decisions, not laws of nature. Meta says plainly that "tuning the weights of the VM allows us to explore different tradeoffs between online engagement metrics," and that there are "literally hundreds of tunable parameters that control the behavior of the system."

Note also the minus sign. Predicted negative reactions subtract.

TikTok described the same weighting logic in 2020: "A strong indicator of interest, such as whether a user finishes watching a longer video from beginning to end, would receive greater weight than a weak indicator, such as whether the video's viewer and creator are both in the same country."

So when people say the algorithm "rewards" comments, what is actually happening is that a model predicts the probability of a comment, and that prediction carries a weight in a sum. Engagement is not payment for good behavior. It is evidence used to guess at the next person's behavior.

Platforms have spent a decade fixing the metrics they optimize

The most useful thing in this literature is watching engineering teams discover that their obvious metric was wrong.

Google's 2016 paper: "Ranking by click-through rate often promotes deceptive videos that the user does not complete ('clickbait') whereas watch time better captures engagement." So YouTube ranks on expected watch time per impression instead of clicks.

Then watch time turned out to be insufficient too. In September 2021, YouTube's VP of Engineering Cristos Goodrow wrote that the system learns from "over 80 billion pieces of information we call signals," and that satisfaction is measured through surveys where viewers rate a video from one to five stars: "Only videos that you rate highly with four or five stars are counted as valued watchtime." He also reported a result that should end the "outrage wins" argument on its own: when YouTube demoted salacious and tabloid-type content, "watchtime actually increased by 0.5% percent over the course of 2.5 months, relative to when we didn't place any limits."

LinkedIn Engineering published the same kind of correction in May 2020. Clicks and likes, they wrote, "can be rare, especially for passive consumers of the feed," and "clicks are noisy indicators of engagement," since a member may click an article and bounce in seconds. So they modelled dwell time instead, and added a P(skip) prediction: the probability that a member's dwell time on your post falls below a threshold. Posts get their score reduced in proportion to how likely you are to scroll past. Adding dwell-time features raised the model's area under the ROC curve "by as much as 10% over multiple trainings," and the A/B tests showed "a large decrease in the number of skipped updates."

The pattern across all of them: the systems keep moving away from cheap, gameable signals and toward slow, expensive ones. Time spent. Completion. Whether you would say you were glad you saw it.

What the platforms say they do not use

This is where you can save real money and anxiety.

TikTok, in the 2020 explainer that remains its clearest published statement: "neither follower count nor whether the account has had previous high-performing videos are direct factors in the recommendation system." Larger accounts get more views because more people follow them, not because the recommender favors them.

LinkedIn's help documentation, last updated in 2026, lists the three signal families the feed evaluates. Identity: profile details, background, location. Content: how often a post is viewed or engaged with, "what the post is about and whether it provides knowledge or advice," recency, language, "how constructive or professional the conversation is." Activity: what you have reacted to, who you interact with frequently, "content you spend the most time viewing." And one line that answers a question small businesses actually ask: "Feed distribution isn't influenced by payments from third parties to LinkedIn, except for promoted (paid) content, which is clearly labeled."

You cannot buy your way into organic rank. You also cannot be punished for not buying.

The four things this changes about how you post

Your post is competing for one person's attention, not for a leaderboard slot. A post that 40 of the right people find useful is doing its job. Optimizing for the number of strangers who see it is optimizing for the wrong term in the equation.

The first seconds are the whole ballgame on ranked surfaces. LinkedIn built a model specifically to predict whether you will skip. YouTube ranks on expected watch time. TikTok weights completion heavily. Whatever you are writing, the opening is not a warmup, it is the measurement.

Diversity rules cap you. TikTok says its feed "generally won't show two videos in a row made with the same sound or by the same creator," and Meta's final reranking stage applies rules like "do not show items from the same authors in a sequence." Posting five times in a day does not give you five times the shelf space with the same person.

New content starts cold, everywhere. TikTok notes that "videos that have just been uploaded or are under review" may be ineligible for recommendation. Every one of these systems needs early behavioral evidence before it can predict anything, and that evidence comes from the people who already know you. That is an argument for writing to them rather than to the strangers you hope the system finds for you.

When this was last checked

The sources above are current as of August 2026, and the Instagram Explore and dwell-time posts are older engineering write-ups that describe architecture rather than this quarter's settings. Architecture is the durable part. Specific weights, thresholds and product surfaces change constantly and are not published. If a piece of advice depends on a number the platform has never printed, treat it as a rumor.