How the Xiaohongshu algorithm works in 2026

Your note stalled at a few hundred views. Somebody told you it fell out of the first traffic pool, and somebody else quoted a scoring formula with weights in it. One of those things is documented. The other is not. Here is the split, before anything else.

The question you are asking	What the platform has actually published	Source
Is distribution staged?	Yes. Recall, coarse ranking, fine ranking, named by the platform’s own engineers	Xiaohongshu tech team, March 2023
How many people see it first?	Nothing. No pool size is published	Not published
What is a save worth against a like?	Nothing. No weighting is published	Not published
How fast does the system react?	Interaction labels used to be collected about 30 minutes after a note was shown; the pipeline now updates in minutes	Xiaohongshu tech team, March 2023
Do new notes get their own shot?	Notes published in the last day are close to half of all homepage feed exposure	Xiaohongshu tech team, March 2023
How big is search?	800 million searches a day; 77% of daily users search to solve a problem	Platform figure, May 2026; Huxiu, May 2026

Two of those six rows say the same thing: nobody publishes it. That is the real state of public knowledge about this platform, and it is where most guides start filling in numbers of their own. Every source below was checked twice in September 2026.

What a traffic pool actually is

“Traffic pool” (流量池) is operator shorthand, not a term Xiaohongshu (小红书) publishes. What the platform does publish is the shape of the pipeline your note enters after review.

In the first half of 2021, the main recall, coarse-ranking and fine-ranking modules behind the Xiaohongshu home feed were all updating on a daily cycle. The team rebuilt recall channels, index updates and model training to update by the minute. Source: Xiaohongshu technology and intelligent distribution department (小红书技术部), March 2023. https://www.6aiq.com/article/1679451572481

Read that backwards and the model falls out. Recall gathers candidate notes that might suit a given user. Coarse ranking cuts that set down cheaply. Fine ranking orders what survives. Every slot in every feed is filled that way, for every user, every time they pull to refresh.

So your note is never “released” to a fixed audience of 200. It is entered into a competition it either keeps winning or stops winning. The pool language describes the result, which is that reach grows in steps rather than smoothly, because each step needs the note to keep beating whatever else recall dragged in.

And the competition is large.

Xiaohongshu users publish more than 9 million notes and leave more than 70 million comments a day. Source: Xiaohongshu (小红书) WILL 2026 conference, reported by 100EC (网经社), December 2025. https://www.100ec.cn/detail--6655530.html

Nobody publishes the rung sizes. Not the platform. Not its engineers, and no regulator has ever asked them to. Every ladder you have seen (200, then 2,000, then 20,000) traces back to an operator post quoting another operator post. We left it out.

The four things the Xiaohongshu algorithm reads first

Four signals decide whether a note escapes its first audience. The platform names all four. It publishes a weight for none of them.

Signal	What the system reads	What you control	Published weight
Subject	Image, video, text and tags, read together	Cover, first line, topic tags	None
Match	Which users recall pulls your note for	The vocabulary your buyer actually uses	None
Early response	Clicks, reading, likes, saves, comments, follows	Hook, first screen, the question you ask	None
Standing	Whether the note and the account are clean	Filing paid work, honest claims	Not a weight, a gate

Subject is not the caption. The platform reads the picture too.

Multimodal understanding of content runs through the whole of Xiaohongshu’s search, recommendation and transaction systems, with applications including short-video understanding, content quality evaluation and multimodal retrieval. Source: QbitAI (量子位), April 2022. https://www.qbitai.com/2022/04/34112.html

That is a 2022 statement and it has only become more true. A cover image that says nothing in Chinese is a cover image the system cannot classify.

Match is the one nobody checks. If recall never pulls your note for the people who would like it, nothing downstream matters. This is why a beautiful note about “sustainable outerwear” dies while a plainer one about hardshell jackets (冲锋衣) travels.

Early response is what the ranking models learn from.

When a user interacts with a note, by liking or saving it, that tells the system the user is interested. When the interest is new for that user, a faster system learns it sooner and can surface related notes while the user keeps scrolling. Source: Xiaohongshu technology and intelligent distribution department (小红书技术部), March 2023. https://www.6aiq.com/article/1679451572481

Standing is not a score at all. It is a gate, and the section on holds further down deals with it.

Now the formula. A scoring equation circulates widely in Chinese operator posts and in English guides, giving likes and saves one point, comments and shares four, and a new follower eight. It appears on no platform page, in no engineering talk and in no filing. The Chinese post most often credited with it introduces the formula as the version that circulates most widely, which is not a source. We do not use it, and we would not price a campaign on it.

Why the first hour matters more than the first day

Because Xiaohongshu (小红书) is built around fresh notes, and because the system used to wait half an hour to find out whether yours was any good.

Notes published in the last day have long made up a high share of homepage exposure on Xiaohongshu, and during this period that share grew quickly to close to half. Source: Xiaohongshu technology and intelligent distribution department (小红书技术部), March 2023. https://www.6aiq.com/article/1679451572481

Half the shelf, reserved for yesterday. That is the whole reason a dead note feels so final: tomorrow the shelf refills.

The traditional approach waits about 30 minutes after an item is shown before collecting the interaction labels it will train on. Source: Xiaohongshu technology and intelligent distribution department (小红书技术部), March 2023. https://www.6aiq.com/article/1679451572481

The team’s whole project was cutting that lag, and the payoff was specific.

Moving ranking and recall from daily to minute-level updates delivered over 10% more average time per user in the home feed, over 15% more interactions, and close to 50% better new-note efficiency. Source: Xiaohongshu technology and intelligent distribution department (小红书技术部), March 2023. https://www.6aiq.com/article/1679451572481

What that buys you in practice: the window where your own effort changes the outcome is short, and it is at the start. Answer comments while they arrive. Post when the people you want are holding their phones, not when your head office is at its desk. Do not publish six notes in a row and then nothing for nine days, because each one lands in a market that has already restocked.

Search and feed are two different machines

That is everything about the feed. Now the half of the platform that most foreign teams never budget for.

Foreign brands tune the Xiaohongshu (小红书) feed and then wonder why the numbers collapse after week one. They are tuning the half of the platform that forgets.

77% of Xiaohongshu’s daily active users use search to solve a problem, and 75% browse the recommendation feed. Source: Huxiu (虎嗅), May 2026. https://www.huxiu.com/article/4861801.html

Those add to more than 100 because most people do both. Search is not the small path.

Xiaohongshu passed 400 million monthly active users and 800 million searches a day. Source: Xiaohongshu (小红书) report, reported by Beijing Business Today (北京商报) via Sina Finance, May 2026. https://finance.sina.com.cn/jjxw/2026-05-27/doc-inhziqxq9291575.shtml

That number has roughly tripled in three years.

Xiaohongshu handled about 300 million searches a day in mid-2023 and about 600 million a day in the fourth quarter of 2024. Source: Zhou Tian Finance (周天财经), reported by 199IT, December 2024. https://www.199it.com/archives/1731824.html

And a large share of it is shopping.

39 million users a day show product purchase-seeking behavior on Xiaohongshu, 140 million instances a day. Source: 36Kr (36氪), reporting Xiaohongshu’s GROW merchant conference, April 2026. https://www.36kr.com/newsflashes/3758099821871879

	Recommendation feed	Search
What starts it	The system picks you	The user types a need
Peak window	Hours	Months, sometimes years
What decays	Fast, once the note is no longer new	Slowly, while the note still answers the query
What you tune	Cover, hook, first screen	The words your buyer types
What it is good for	Discovery and reach	Demand that already exists
How it fails	The note simply stalls	The note never gets pulled for the query

The practical consequence: a note can flop in the feed and still earn for a year, if it is the best answer to a question people keep asking. Our guide to Xiaohongshu marketing for foreign brands covers the account side of this.

Tags, titles and covers: what the ranking actually reads

Everything at once, which is the part people get wrong. The multimodal work quoted above means the cover image, the video frames, the title, the body text and the topic tags are read as one object. A gorgeous shot with no Chinese text on it and a title in English is not a subtle note. It is an unlabeled one.

Four rules follow, and not one of them is a secret.

Write the cover, do not just shoot it. Put the promise in Chinese characters on the image. It is the only text guaranteed to be read before anyone clicks.

Put the query in the title, not the brand line. If your buyer would type it, it belongs in the first twenty characters.

Tags are the third one. They tell recall which neighborhood the note lives in, so treat them as topics rather than confetti. Twenty unrelated tags make that job harder, not easier, and a lot of foreign accounts are running twenty.

Say the thing on the first screen. Reading depth is an early signal, and an introduction that clears its throat for four lines has already spent it.

No platform source publishes a keyword density, a title length or a tag count. Anyone who quotes one is quoting themselves.

The ways a note gets held back without a warning

A note can be live on Xiaohongshu (小红书), visible on your own profile, and still be going nowhere, because enforcement here is mostly about display rather than deletion. The scale is published.

Between March and the end of August 2025, Xiaohongshu banned over 12 million fake accounts, actioned 13.76 million false marketing notes and cleared over 360 million fake comments. Source: China Daily (中国日报网), January 2026. https://cn.chinadaily.com.cn/a/202601/20/WS696eef27a310942cc499bf9f.html

Thirteen million notes in six months. Commercial behavior is explicitly in scope now.

Xiaohongshu launched Community Convention 2.0 on 19 January 2026, adding a section on orderly commerce (有序经营) to the community rules. Source: China Daily (中国日报网), January 2026. https://cn.chinadaily.com.cn/a/202601/20/WS696eef27a310942cc499bf9f.html

The most common own goal for a foreign brand is paying a creator off platform. Filed collaborations run through Pugongying (蒲公英), the platform’s own creator marketplace, and the fee is public.

Pugongying charges 10% of the deal in its standard mode and 20% in the premium mode that adds platform promotion. Source: Niaoge Biji (鸟哥笔记), October 2022. https://www.niaogebiji.com/article-482538-1.html

Chinese advertising rules point the same way.

Content that promotes a product through knowledge sharing, experience sharing or a product review and attaches a purchase link is advertising, and the publisher has to label it conspicuously as an advertisement. Source: State Administration for Market Regulation (国家市场监督管理总局), Internet Advertising Measures, Article 9, in force May 2023. https://www.gov.cn/gongbao/2023/issue_10506/202306/content_6885261.html

A brand that routes payment around the marketplace is not saving 10%. It is buying an unmarked ad. The brief is unenforceable too, and the creator is the one carrying the risk. On a platform that actioned 13.76 million marketing notes in half a year, seeding (种草) filed properly is the cheaper option.

Nothing in the published rules promises you a notification. The platform reports what it actioned in aggregate, six months at a time. Your first signal is the analytics panel, which is why the next section is the one that matters most. The in-app checks that confirm a hold, and the six other things that stall an account, are in seven reasons a Xiaohongshu account stops growing.

How to read your own note analytics

Your account is the only benchmark you can verify. Every market average you have been shown was computed by somebody selling something.

The creator dashboard gives you views, likes, saves, comments, shares, new followers and where the traffic came from. Four ratios turn that into a diagnosis, and all four are arithmetic you can do yourself.

Views against impressions tells you whether the cover and title work. A note the system showed to plenty of people and nobody opened has a packaging problem, not a distribution problem.

Saves against likes tell you what kind of note you actually wrote. A like is politeness. A save means somebody intends to come back to it, and coming back is the behavior this whole platform is built on, so a note with plenty of likes and almost no saves was enjoyed rather than used.

Follows against views is the account’s own report card. A note that went wide and converted nobody usually went wide of the people you sell to, which is a targeting result, not a content result, and it will repeat next month unless the subject signal changes.

Search share over time is the one foreign teams never look at. Pull it at day three, day thirty and day ninety on the same note. If the search share is climbing while total views fall, the note is working exactly as it should.

Track those four on every note for eight weeks and you will know more about your own distribution than any published guide can tell you, including this one.

Tracing one note, from post to plateau

This traces the mechanics, not one of our accounts. We do not publish client figures without written sign-off, so there are no view counts here. The numbers that matter are the ones already sitting in your own dashboard.

Stage	What is happening	What you see	What it means if it stalls here
Review	The note is checked before it enters distribution	Note live, views near zero	Normal. No published review time exists
First audience	Recall pulls your note for a small matched set	Views move, impressions low	The subject signal is weak. Fix the cover and the title
Early response	Interaction labels feed the ranking models	Views rise, saves flat	People clicked and left. The first screen is not delivering
Wider feed	The note keeps winning slots against fresh notes	Views climb for hours, not days	Expected. Fresh notes take back the shelf
Search tail	The note answers a query people keep typing	Views low but steady for weeks	The title does not carry the words buyers type

In a weekly report the second and third stalls look identical, and they need opposite fixes. One is a packaging problem, the other is a content problem, and telling them apart is most of the job.

For the account setup that sits under all of this, see our guide to Xiaohongshu business account setup. For what the work costs once you are running, see our Xiaohongshu marketing cost guide. The RedNote agency page lists what we run for clients on the platform.

Frequently asked questions

Why did my Xiaohongshu note get no views?

Usually the subject signal, not a penalty. If the cover carries no Chinese text and the title uses brand language instead of the words people search, recall struggles to place the note and nothing downstream happens. Check whether the note is being shown at all. Low impressions means a matching problem. High impressions with no clicks means a packaging problem.

How long does it take for a note to get distribution?

Faster than it used to. Xiaohongshu’s engineers rebuilt ranking and recall from daily updates to minute-level ones, and notes published in the last day account for close to half of home feed exposure. The platform publishes no review time, so treat the first hour as the window that matters and judge a note over its first day, not its first ten minutes.

Does posting time matter on Xiaohongshu?

Yes, for the feed, and much less for search. Early interaction is what the ranking models learn from, so posting when your buyers are awake in China gives the system real signals to read. A note that earns search traffic keeps earning it for months regardless of when it went up. No platform source publishes a best hour to post.

Stop guessing at your own distribution

Send us the account. We will read the last ninety days of note analytics, tell you which of the four signals is failing, and show you the search queries you are not being pulled for.