Xiaohongshu's second great voyage, this time heading towards AI

The Second Great Voyage of Xiaohongshu—This Time Heading Toward AI

By Dongcha Beating

Source:

Reprint: Mars Finance

By Sleepy

By the end of 2022, shortly after ChatGPT was launched, Mao Wenchao borrowed an employee’s phone. In the chat box, he typed one question: Would Xiaohongshu be overturned?

Reports say that ever since then, he had the team report AI progress every two weeks. Two weeks at a time—because the machine still couldn’t give him answers that would put his mind at ease.

In August 2023, in an internal letter, he wrote that while chatting with foreign friends, he found that many questions people were asking on ChatGPT were based on life experience—how to choose products, how to use them, and how to avoid pitfalls. This overlapped with Xiaohongshu’s business.

But he added that this was because overseas lacked the accumulation of such experience, while Xiaohongshu had it. For the time being, this moat was still difficult for AI to shake.

The term “moat” used to be something entrepreneurs said to investors—but this time it sounded like he was saying it to his anxious self.

That year, Xiaohongshu had just turned ten. Monthly active users surpassed 300 million; it turned profitable for the first time. Revenue reached 3.7 billion dollars, net profit was 500 million, and it was expected that profit in the following year would double again, exceeding 1 billion.

In business history, companies die in two ways: dying from poverty, or dying from wealth. Countless die from poverty—there’s not much to say. Those that die from wealth always make the news: when Kodak died, there was money in the accounts; when Nokia died, it was still the industry leader.

Having a lot of money and having a long life are two different things. Plenty doesn’t eliminate fear—it only turns fear into a series of concrete actions.

By 2026, these actions started to cluster densely.

On June 8, Xiaohongshu launched RED Skill. Under notes, a component can be attached; copy it to Agent and it can be used.

Earlier, on April 30, the AI First-Level department Dots was established. Models, infrastructure, and engineering products were all included at once, reporting directly to the new president, Conan.

Even earlier, it acquired the development company behind the AI search product Diandian—and also obtained a payment license.

On the list of strategic investments, names began to appear: MiniMax, Moon’s Dark Side, and a whole string of AI hardware companies.

Over the past thirteen years, the consumption experience, lifestyle habits, and daily judgments left behind by hundreds of millions of users in notes—this is its real foundation. When AI arrived, it wanted to reprocess these judgments: first into answers, then into tools, and finally into a business. If it didn’t want to wait for itself to be overturned, it had to start moving first.

But can experience withstand processing? To answer that question, you have to go back to 2013—the era of China’s own great voyage.

A Great Voyage of 70 Million People

In June 2013, Qu Fang resigned from her job at a foreign company and co-founded Xiaohongshu in Shanghai with Mao Wenchao. Their first product was not an app, but a PDF: “Xiaohongshu Outbound Shopping Guide.”

That year, the number of Chinese outbound travel trips exceeded 70 million—equivalent to all of France’s population taking a trip abroad together once.

Europe’s great voyages brought back spices, gold, and colonies. China’s great voyages brought back cosmetics, electric rice cookers, and travel guides. Even though the objects were small, the shared desire was the same: to bring the good things from far away back home.

Suddenly, the world of goods beyond the borders opened up. The shelves in duty-free shops were packed with tourists holding their phones, and no one told them what was worth buying. Information gaps were like mineral deposits—whoever first gathered the experience of those early arrivals could become the mine owner.

That PDF was uploaded to a website and downloaded 500,000 times in less than a month. A few months later, it grew into an app; a few years after that, it entered the phones of hundreds of millions.

When Chinese people encounter something, they don’t ask for manuals. They ask people.

Fei Xiaotong wrote in “From the Soil to the City” that in rural society, trust doesn’t rely on contracts—it relies on familiarity. Those who learn a craft learn from their masters. New daughters-in-law ask their mothers-in-law. First-time city newcomers seek out people from their hometown. For thousands of years, experience was passed down this way—slowly, but enough.

“Enough” depends on two premises: people live close to each other, and life moves slowly. Those two premises were gradually lost over the past few decades. Hundreds of millions left their hometowns and moved into buildings where they didn’t even know what surname people in the adjacent unit had. What you could buy expanded from a few hundred items on the shelves of supply and marketing co-ops to hundreds of millions of items on e-commerce pages. It’s hard to ask an elder who has never used a robot vacuum what model to buy—because the experienced people still hadn’t arrived yet.

The internet claimed it would solve this problem, but in the process it made the problem bigger. The internet was invented to obtain information, but in the end, information became so abundant that no one dared to trust any of it. Most online information comes from sellers. Their job isn’t to help you judge—it’s to persuade you to pay. Judgments can only come from people who don’t profit from you.

Xiaohongshu gathered the “I’ve tried” experiences scattered among hundreds of millions of strangers. A girl from Guangzhou wrote about how a certain foundation cakes on oily skin. A young man from Shenyang wrote down eleven pitfalls he hit during a renovation. A mother wrote about hesitating for dozens of days between two types of baby food.

Most of these writers were unknown and not experts. Their writing wasn’t particularly rigorous, and it might even be mixed with brand deals and misjudgments—but those words had warmth.

Encyclopedias pursue definitions; advertisements pursue persuasion. These notes pursue nothing; they are simply testimonies—flawed testimonies. The most credible testimonies in court are precisely these. Overly perfect testimonies feel like memorized scripts. Later, the industry gave this phenomenon a name: planting.

By the end of 2024, this app’s daily search volume was close to 600 million times. People here search far less for knowledge; most of what they search for is daily life—renovation, skincare, and travel guides. Search engines give you data; Xiaohongshu gives you other people’s experience. Of course, there are ads in it too, and they may not always provide the most precise answers—but people still want to look, because many problems in life don’t have standard answers in the first place.

Behind 600 million searches is 600 million hesitations: people holding their phones late at night, unable to decide. That is Xiaohongshu’s entire asset.

Then, AI arrived.

Patience Hits the End

Thirty years of the internet is a history of human patience declining.

In the portal era, information was organized into directories, and people had to look for it themselves. In the search era, it became links, and people had to click themselves. In the information stream era, you don’t even need to search anymore—algorithms feed you. Every change shortens patience by another step. In the AI generation, information is turned directly into answers, and people’s patience reaches its limit.

This isn’t the users’ fault. People’s love for convenience has no bounds—wheels, elevators, remote controls, all invented in this way. Once someone gets used to AI chat boxes, it’s hard to return to the days when they scroll through posts and do their own filtering.

Xiaohongshu’s difficulty is that the part most valuable to it—precisely those experiences that are hard to compress into a single answer—is also the hardest for AI to reduce to one.

In the past, people browsed 20 notes here. They compared, hesitated, and finally made their own decision. The process was slow because you could see three people saying it was good, two regretting it, and one warning you that it works well but is delicate. Someone might write that a hotel’s soundproofing is poor, but the breakfast is great. That line is useful because it comes from a specific person; you can probably guess what they care about, then decide whether their experience is relevant to you.

AI is like a pre-made meal factory: it delivers the flavors of life as standardized recipes. It’s convenient, but the hesitation, failures, and prerequisite conditions—the most valuable parts of experience—are exactly what gets removed.

Experience always grows out of specific people. Skin type, which city they live in, how much budget they have—these determine whether a recommendation actually works. Machine-provided answers lack those prerequisites; they sound like slogans. But slogans can’t help you choose a foundation.

Xiaohongshu understands this precarious situation. Patience can’t be held back. When that day comes, its 600 million searches will become training data for someone else’s models, and it will become a mine—an open-pit mine—where anyone passing by can dig up a scoop.

So it must act on its own. They didn’t wait too long. Since 2023, it has developed its own model, “Xiaodigua,” launched the AI painting tool Trik, and rolled out a dialogue product beta test called “Da Vinci.” Most of these products didn’t create major waves, but that doesn’t mean they were wasted. They were like rounds of probing. Xiaohongshu needed to first find out what AI could truly do for it.

The real direction was discovered by Diandian. It builds life search by combining in-site notes with information across the entire internet. Users can ask in text, images, and voice. Later, Xiaohongshu simply acquired the company behind Diandian. Diandian wasn’t a breakout hit, but the mission of a scout isn’t to storm a fortress.

It discovered something: in the past, search started from keywords, and users entered a doorstep number. Now, questions start from a situation; users enter a whole set of troubles. People no longer just search “a Okinawa trip for families with children.” They ask: “How should I plan a five-day trip to Okinawa with a three-year-old? My budget is 15,000. I want to stay close to the sea—how should I arrange everything?”

To solve these troubles, Xiaohongshu has gradually released research on multimodal retrieval and search understanding, open-sourced an image editing model called FireRed, and open-sourced the search agent framework REDSearcher. It has no intention of competing with big tech for the seats of general-purpose models. While others are racing for parameters and leaderboards, what Xiaohongshu wants to do is to read the real human experiences scattered across images, videos, and comments, break them apart, and recombine them into specific, usable advice. With the establishment of Dots this year, this line has moved from edge experiments into core business.

Making answers by putting together 20 notes is something Xiaohongshu wants to do for users. But a single answer can only solve one problem. What it truly wants is to turn experience into a capability that can be reused repeatedly.

Notes Grow Hands and Feet

RED Skill is exactly doing this—turning experience from content into tools.

After the launch, Xiaohongshu quickly rolled out support activities and curated rankings. 300,000 people began writing AI Skills. “Guicang’s” PPT generation tool, which had previously received more than 10,000 stars on GitHub, was uploaded to Xiaohongshu and installed by thousands within days.

Going even further back, last year’s independent developer competition received 1,355 projects. This spring’s first hackathon—48 hours of closed development, a 500,000 yuan prize pool—saw 60% of participants born in the 2000s (00后), with the youngest only 12 years old. On-site notes about “Build in Public” had already exceeded 1.1 million.

Although these numbers are still not enough to prove that the ecosystem has fully taken shape, they are enough to show what Xiaohongshu wants to do.

In the past, when developers wanted to give products an initial boost, they mostly went to GitHub or Product Hunt. Those places have many peers and investors, but not necessarily many ordinary users. People may give you stars or estimate your valuation, but not necessarily place orders.

Xiaohongshu is targeting this gap. Developers write progress here; users raise needs in the comment section; creators write their usage experiences into notes; and the platform then gathers early attention through rankings. Writing an AI tool is only the beginning. It still needs to be tried, debated, and translated into something ordinary people can understand and use.

When it comes to making tools, Xiaohongshu may not be the strongest. But when it comes to getting tools into everyday life, it’s very familiar.

Over the past 13 years, Xiaohongshu’s creators have been more like storytellers: writing vividly, recommendations that feel trustworthy, and influence that accumulates little by little. Users are willing to listen because they trust the person. In the AI era, creators are starting to become craftsmen. “Fame” doesn’t just look at how they write anymore. It begins to depend on how many people install their tools, how often they’re used, and how many real tasks they actually help users get done.

For those who write notes, experience used to be something only visible. Now it can also be called upon. If it can be called upon, there is the possibility of pricing.

Before Search Terms Appear

In December 2024, Dai Lidan, a partner at Today Capital, joined Xiaohongshu and became Chief Strategy Officer, setting up a strategic investment team. A graduate of Peking University’s computer science department, she worked on Baidu Image Search and Baidu Maps. Later she went to Harvard to get an MBA, returned to China, and joined Today Capital. She has worked across technology, products, and capital.

Before she arrived, Xiaohongshu invested mostly in consumer brands—M Stand Coffee, Moody contact lenses, as well as food, trendy toys, and maternal and infant products. Those were businesses rooted in young people’s lifestyles, which she understood well. After she joined, financial investment and strategic investment were separated, and the strategic team shifted toward hard technology and AI. MiniMax’s shareholder list includes Xiaohongshu, and it also participated in Moon’s Dark Side’s funding round that exceeded 1 billion dollars.

What it bet on wasn’t only AI on screens.

In the Nanshan Science and Technology Park area in Shenzhen—centered around DJI’s headquarters—there was a cluster of AI hardware people. In the second half of 2025, Xiaohongshu participated in investments in nearly 10 startups there. It moved quickly, sometimes finalizing a deal within one or two days, and it was also willing to secure a share by paying higher valuations.

Two of these investments were made through its subsidiary “Shu Neng Sheng Qiao.” One went to Yunwang Innovation. This company turned traditional foam rollers into AI massage robots capable of sensing where the body hurts and autonomously adjusting the pressure and path. The other went to Skyris, making companion robots that float in midair on helium and interact with people using wing mechanics, LED eyes, and voice.

People in the industry often call Xiaohongshu “the entry point for life decisions.” Those eight characters look great on PPTs—but the good-looking words are three feet away from reality.

Decisions are already late-stage. Once someone starts searching for how to use a foam roller, the demand has already been expressed. Before it becomes a search term, the need often has no name yet—it might just be a shoulder that’s been hurting nonstop, or someone sitting at home for three hours.

In the past, Xiaohongshu stayed downstream, waiting for people to turn life experience into notes. Now, it wants to move upstream, proactively finding those needs that haven’t yet turned into search terms.

In 2024, Xiaohongshu’s parent company also funded the Sichuan Sha Jiang Venture Capital fund as an LP. Sichuan Sha Jiang was one of its early investors. In 2014, at a startup contest, it discovered the company, and made an investment the following year. Ten years later, the invested company became an investor itself. Xiaohongshu used a fund share to gain a long-term channel into early-stage projects.

Of course, investing early doesn’t necessarily mean being right. AI hardware hasn’t proven that it can be mass-commercialized yet. Mass production, supply chains, and after-sales support—each is a difficult job, and none of it is a business Xiaohongshu is familiar with. Even more complicated is data. When does your shoulder hurt, the device knows; why does it hurt, the platform also wants to know. Understand too little and the product won’t work well. Understand too much and there are privacy risks.

But it still has to invest. What it truly worries about isn’t today, but that deep-night indecisive person who, instead of opening Xiaohongshu to browse notes, will hand the problem directly to another AI.

When Advertising Lives Inside Answers

Xiaohongshu’s story is inseparable from commercialization.

On this platform, experience and business have always been intertwined. Skincare advice is backed by skincare products. Renovation guides are linked to building materials. Users want to take fewer detours; merchants want to be seen; the platform wants to make money from it. Each of the three wishes alone is reasonable, but put together they require a set of rules.

In November 2025, Xiaohongshu obtained an Oriental Payment license through its subsidiary, completing the final link. AI can recommend products and services for users, but after recommendations, where the order is completed and where the money flows—these determine who ultimately holds the business. Xiaohongshu doesn’t want to only provide advice; it also wants to keep the transaction itself in its own hands.

Xiaohongshu built its commercialization earlier. In December 2024, at the WILL Business Conference, Xiaohongshu released its AIPS audience asset model. It connected Taobao, JD.com, and Vipshop data through the planting alliance, then reconciled it with brand-owned data. At the conference, there were two figures that were particularly interesting. The decision cycle for facial serums was up to 29 days; for maternal and infant nutrition, it was over 70 days.

This is precisely the part of the planting business that can’t be explained clearly. Someone looks at reviews today; ten days later they search for ingredients; twenty days later they place an order on another platform. In between, they’ve watched live streams and asked friends. Finally, who actually brought the money? Merchants want to know, and Xiaohongshu couldn’t clarify before. What AIPS aims to do is to map this vague path clearly.

Xiaohongshu’s true value isn’t traffic. A person scrolling short videos might just be passing time. A person starting to search for serums or baby food is usually close to placing an order.

What’s most valuable is knowing what people are hesitating about. AI can see that hesitation more clearly. In the past, platforms only knew what you had looked at. Now it also knows what you want to solve. What you hand over isn’t just a keyword, but an entire situation—budget, preferences, physical condition—and those worries that people are not so willing to say out loud.

Advertising as a business has always been moving into people’s judgments. At first, it stood by the roadside as a signboard. People could tell at a glance it was an ad; if they didn’t want to see it, they just went around. Later, it blended into articles and became soft ads and product placements. Then it entered the information feed, looking more and more like the content you would naturally see anyway. With every step forward, ads become harder to notice and closer to the decisions people make. In the AI era, it found an even better place—living inside answers.

Machines Learn “I’ve Tried”

In February 2026, according to the national “Administrative Measures for Identifying AI-Generated Synthesized Content,” Xiaohongshu required creators to label AI-generated images and videos and text. Content without labels faced restrictions on distribution. In March, it began cleaning accounts that were being operated entirely by AI—accounts that were fully written and posted by machines—then banned them outright. In April, it for the first time fully disclosed its AI governance principles: it encouraged AI to amplify creativity, but opposed AI fabricating life. Cloning voices, inventing personas, and fabricating experiences—none of it is allowed.

These words sound like declarations, but in fact they’re about self-preservation.

AI is best at learning people. By the end, it even learns “I’ve tried.” This is the fastest thing it learned—and also the least it should have learned. The trust Xiaohongshu has built over thirteen years depends on countless specific “I’ve tried” experiences. Machines can write ten thousand trial notes, but they’ve never truly tried anything. Their skin never gets allergic reactions. Their wallets never hurt.

Once this kind of content becomes abundant, real human experience will lose value along with it. Xiaohongshu might end up turning back into what it once tried to replace—becoming that pile of more beautiful, more “human-like” seller scripts.

Everything that comes next is still not decided. Whether RED Skill can grow a real ecosystem, whether Diandian can enter the main app, whether payments will be integrated into answers—all must be left to time. But the nature of the matter is already clear. Xiaohongshu is acting as a translator: turning real human experience into machine-processable structures, transforming life judgments into tools, and bringing hesitation into business.

Translation values trust, clarity, and elegance; machines have already learned clarity. What Xiaohongshu has to protect is trust.

Borges once wrote about an empire obsessed with precision. There, cartography became more and more sophisticated: a map of each province as large as a city; a map of the entire empire as large as a province. The cartographers still felt it wasn’t enough, so in the end they drew a map the same size as the empire’s territory itself—every city, every road, and every wasteland could be found on the map in the corresponding spot. But once the map became as large as reality, it lost its usefulness. Later generations stopped caring and let it rot in the desert.

AI is now drawing such a map of experience—getting ever more detailed, faster, and easier to forget that the map is not life itself.

In his letter, Mao Wenchao said this moat is difficult for AI to shake for now. He probably also knows the real problem isn’t the moat—it’s the city. Xiaohongshu must build increasingly smart machines, or else the experience accumulated over thirteen years will soon be organized, called upon, and re-priced by others. But once the machine’s voice drowns out human voices, the city becomes empty. The moat protects an empty city. No matter how wide it is, it has no use.

It needs to repair the machine into the city. And it must ensure that what remains in the city is not only the machine, but also those indecisive people from deep into the night—and those who are willing to say “I’ve tried” to them.

This is its true moat, and also all of its anxiety right now.

Epilogue

Before this article was finalized, Bloomberg reported that Xiaohongshu plans to secretly submit an IPO application in Hong Kong by the end of this month. Its valuation had once reached 31 billion USD, and it was expected that its profit for all of 2025 would be about 3 billion USD.

From a PDF to the Hong Kong Stock Exchange—13 years. It has turned the hesitations in the lives of hundreds of millions into something that can make money. Now, the capital market is up to reprice it.

Stock prices will always rise and fall. But those deep-night people who stare at their phones unable to decide, and those willing to tell strangers “I’ve tried,” will not disappear from the story just because of stock price movements. Money makes a company run fast; running long is another matter.

As for what comes next, leave it to time.