How Does Xiaohongshu Decide What to Recommend? A Practical Guide to the Algorithm

Xiaohongshu’s recommendation system is not simply assigning a label to an account, nor is it using one fixed formula to decide whether a post will go viral. Historical technical disclosures describe a multi-stage process that retrieves, ranks, and re-ranks candidate posts while learning from recent user behaviour. For marketers and creators, the practical priority is not “account-warming hacks.” It is helping the system understand what a post is about and giving the right audience a genuine reason to click, stay, save, comment, or continue exploring.

A post generally passes through several stages before it appears in a user’s feed. Retrieval produces multiple pools of potentially relevant content. A lightweight ranking stage narrows the pool. More detailed ranking models estimate likely behaviours such as clicks, dwell time, likes, saves, and comments. A final re-ranking layer may also balance freshness, diversity, and safety before deciding what the user actually sees. The models and objectives can change over time, so the idea that the platform relies on a single account label is an oversimplification.

Historical disclosures describe several forms of collaborative-filtering retrieval. Item-to-item signals can connect posts frequently consumed by the same users. User-to-user signals can identify people with similar behaviour, while author-to-author signals can reveal relationships between creator audiences. If many users who engage with sunscreen reviews also save content about makeup for oily skin, the system may learn an association between those topics. A single like does not determine distribution on its own, however. It is interpreted alongside prior behaviour, content features, timing, and feedback from other users.

Earlier public material from Xiaohongshu described recommendation pipelines that relied partly on T+1 or daily batch processing. From 2018 onward, the company discussed moving parts of that pipeline to real-time stream processing. Later presentations described a broader shift from daily and hourly updates toward minute-level recommendation infrastructure. Historical engineering reports mention interaction-label windows of roughly 30 minutes, while selected model parameters, collaborative-filtering jobs, and indexing components were later updated within minutes. These figures describe specific systems at specific points in time—not Xiaohongshu’s complete recommendation architecture in 2026.

A new post has little behavioural history, so the system cannot know with certainty who will value it. Public technical presentations have described a two-stage cold-start process: content signals—such as text, visuals, video, and creator information—help find an initial audience, and early feedback helps refine later distribution. A more accurate mental model is: initial matching → feedback collection → prediction update → distribution adjustment. This does not prove that every post enters a fixed-size “traffic pool,” nor does strong early engagement guarantee large-scale reach.

Behavioural signals help a recommendation system infer both interest and content value. A search can indicate active demand. A click suggests that the cover and headline earned attention. Dwell time and continued reading may show whether the content fulfilled that expectation. Likes, saves, comments, and follows can represent different levels of approval or future intent. Xiaohongshu has not published a permanent universal weighting formula, and the meaning of each action may vary by format, user stage, and recommendation context. Treat claims such as “a save is always worth more than a like” as unverified rules of thumb, not platform policy.

First, define a specific audience, situation, and problem so both users and systems can understand the topic quickly. Second, make the cover and headline accurately promise the value delivered in the post; clickbait often creates shallow clicks and fast exits. Third, provide useful information early. Fourth, use steps, checklists, comparisons, examples, and save-worthy details to deepen consumption. Fifth, evaluate combinations of search terms, clicks, dwell time, saves, comments, and subsequent visits rather than judging a post by likes alone. Sixth, keep testing topics and formats, but change only a few variables at a time so the result remains interpretable.

The algorithm is not only looking for an abstract “perfect post.” It is continuously estimating which people may find a post relevant, in which context, and at what moment. The practical chain to optimize is: precise topic → qualified click from the cover and headline → content that retains the intended reader → a reason to save, respond, or explore further → enough useful feedback for the system to find a better-matched audience. This is a more defensible operating model than obsessing over posting-time folklore, arbitrary account-warming periods, or invisible labels.

This article is based on historical technical material published by Xiaohongshu and related conference presentations. Its purpose is to explain general recommendation-system mechanisms for marketers and creators. The platform does not disclose its complete ranking formula, and its models, features, weights, and governance policies continue to evolve. This article therefore does not represent Xiaohongshu’s full 2026 architecture and should not be used to promise traffic or bypass platform rules.