At the end of April, when Xiaohongshu issued an internal memo to all staff announcing the establishment of its AI division, Dots, three years had elapsed since Mao Wenchao first posed the question to GPT: “Will Xiaohongshu be disrupted?”

Three years is ample time for a company to transition from a wait-and-see approach to making significant investments, evolving from tentative exploration to full-fledged transformation. During this period, Kimi’s valuation surged from $4.3 billion to over $20 billion; Kuaishou’s Kling AI pursued independent financing at a $20 billion valuation, backed by $500 million in annualized revenue; ByteDance’s Doubao reached 345 million monthly active users (MAU), cementing its position at the forefront of AI product rankings.

Over the same span, Xiaohongshu assembled a hundred-strong AI team, launched a standalone app, released a series of open-source models that made little impact in the tech community, and only in spring 2026 announced the elevation of AI from product experimentation to an organizational priority by establishing a dedicated division.

Of course, this timeline alone does not necessarily label Xiaohongshu as “slow.” Within a product structure inherently tense between authentic community content and AI-generated material, Xiaohongshu has yet to find a compelling way to intervene. While other companies have forged clear product paths, Xiaohongshu is still grappling with a more fundamental question: What role should AI play in a community where “authentic human presence” is the core competitive edge?

Once AI intervenes on a large scale in content production or answer distribution, users will inevitably begin to wonder: Is the experience on the other side of the screen a genuine human life record, or a machine-curated result? This touches on the heart of Xiaohongshu’s most valuable asset—trust. Brands advertise here because users believe recommendations stem from authentic experiences; users search here because they trust they’ll find real pitfalls others have encountered and choices they’ve made. As reported by Tiger Sniff, in 2025, 73% of Xiaohongshu’s monthly active users engaged in searches, a mindset built on that same foundation of trust.

Xiaohongshu holds one of the densest repositories of real-life experiences, but during AI’s two hottest years, it failed to swiftly transform this content into a new entry point. Now that it’s finally accelerating, it still dares not run too fast—because a community built on authentic life experiences inherently clashes with large-scale machine involvement in content production and answer distribution.

In other words, “How deeply can AI intervene without undermining the authenticity of the community?”

AI Fails to Grow a New Story

DianDian, initially envisioned as Xiaohongshu’s “new entry point,” began with a self-developed foundation model, then pivoted through external model integration, humanized persona reshaping, and multimodal redesigns. After several rounds of adjustments, DianDian still failed to become a standalone success. However, after returning to the main app, it moved closer to Xiaohongshu’s truly valuable search scenarios.

As of publication, DianDian ranks 163rd in App Store download charts with only 49 user ratings, while leading AI app Doubao has 2.23 million ratings. For an AI product to deliver these numbers after roughly a year and a half is hardly a resounding success.

DianDian’s issue isn’t merely product maturity. It has always competed in two directions simultaneously: as a standalone AI app, facing direct pressure from Doubao, Kimi, and DeepSeek; as a community tool, it highly overlaps with Xiaohongshu’s existing “Ask” feature. Failing to find its place in either arena, it ultimately retreated to the main app, relying on community traffic to sustain its presence.

The same holds true at the model level. Xiaohongshu Hi Lab (now upgraded to Dots) sequentially released dots.llm1 (text LLM), dots.vlm1 (visual-language model), dots.ocr (document parsing model), and a toolchain covering multimodal reasoning, image editing, and deep search. Yet these open-source models failed to gain noticeable traction in the tech community or integrate into Xiaohongshu’s core products: DianDian uses DeepSeek, while “Ask” uses Tongyi Qianwen.

The persistent disconnect between product and self-developed models means Xiaohongshu’s technical investments at the model layer have yet to translate into product differentiation that users perceive.

In a 36Kr report, a former Xiaohongshu AI employee said, “Because it’s not listed, Xiaohongshu needs AI stories even more.” But at this stage, Xiaohongshu’s AI seems more about filling capability gaps in its existing community chain than growing into an independent platform narrative. Functionalizing AI doesn’t necessarily drive platform-level valuation.

Salesforce offers an external reference. It hasn’t lacked AI moves—from Einstein to Agentforce, AI capabilities are embedded across CRM, sales, customer service, and marketing. In September 2024, Salesforce launched Agentforce and made it generally available in October, receiving short-term market approval: its stock price rose nearly 18% from late October to late November 2024.

But this AI narrative didn’t sustainably rewrite its valuation. Salesforce’s 2025 revenue reached $37.895 billion, up 9% YoY; its 2026 revenue guidance is $40.5–40.9 billion, up 7–8% YoY—growth hasn’t significantly accelerated due to AI. In 2025, Salesforce’s stock price fell ~20%. If AI can’t prove new monetization paths, new user scales, or new platform entry points at Xiaohongshu, the market’s room for revaluation may remain limited.

That doesn’t mean Xiaohongshu’s path is wrong. DianDian’s return to the main app reconnects AI capabilities with Xiaohongshu’s most valuable assets: community traffic, authentic content, and search scenarios. It can validate AI’s value in real business contexts: whether it boosts search frequency, improves ad conversion, or helps merchants understand user preferences. These validations matter more to Xiaohongshu’s actual commercial needs than isolated AI app downloads.

For Xiaohongshu itself, this route also offers a relatively hidden benefit: it doesn’t need to bear the high-intensity pressures of continuous model training, computing power procurement, and user growth like pure AI companies. An internal technician told 36Kr, “We have enough cards for simpler tasks, but for foundational models, we might fall short.” This statement itself reflects Xiaohongshu’s AI strategy: pragmatic, bounded, and subordinate to the main business’s resource allocation logic.

AI Transforms Search Without Rewriting the Business

“Ask” is Xiaohongshu’s closest-to-success AI product so far, and its success simultaneously defines the boundary of how far Xiaohongshu’s AI can go.

According to 36Kr, in 2025, “Ask” improved community user retention by ~2–3% and reached ten million-scale DAU, with user query volumes increasing by millions. In a mature product with MAU consistently exceeding 300 million, a 2–3 percentage-point retention boost is no small feat. A former Xiaohongshu employee who worked on the product described the mechanism: users may spend less time browsing individual answers, but overall usage frequency increases—“users who asked once a week now ask daily.”

This outcome completely contradicted management’s early fears. They thought if users could get quick AI answers, they’d have no reason to browse notes. It turns out “quick answers” cultivated higher-frequency question-asking, and these questions still relied on Xiaohongshu’s community content for support and extension.

“Ask” proves AI’s effectiveness, but first and foremost as “community + AI,” not “AI standing alone.” The product never existed independently of the community; its answer-generation capability directly depends on the quality and density of community content, and its value is only validated within community search scenarios. Xiaohongshu’s most certain source of commercial value from AI remains community traffic and content assets, not model capability itself.

Over time, Xiaohongshu had to pursue AI search because users’ question-asking entry points were migrating outward. NBER and OpenAI research shows that by July 2025, ChatGPT had been adopted by ~10% of the global adult population, with over 70% of use for non-work purposes.

CNNIC data shows that as of December 2025, China’s generative AI user base reached 602 million, with 80.9% using AI to “answer questions” and 30% using it as a “life assistant.” These two scenarios overlap highly with Xiaohongshu’s core user behaviors. Users who once searched Xiaohongshu for “3-day family travel itineraries” or “oil-skin foundation recommendations” can now ask Doubao or DeepSeek directly for consolidated answers.

For Xiaohongshu, the risk is whether users’ “first questions” start migrating away. Rather than letting external AI summarize its community experience, Xiaohongshu aims to become the distribution layer—summarizing answers in-app and redirecting them to notes, products, and creators. This is defensive logic, not offensive.

This raises a second-order problem: the closer AI gets to transaction scenarios, the more it amplifies Xiaohongshu’s long-standing structural issues.

Xiaohongshu’s ad monetization logic is built on community credibility. When users search for product recommendations via “Ask,” it generates a ranking based on authentic community discussions: which brand is mentioned most, which product is recommended most frequently. This feels credible to users precisely because it comes from real community voices. Brands advertise here fundamentally because users trust the content isn’t pure soft advertising.

The problem is that authentic community rankings don’t always align with commercial clients' expectations. For users, the more AI rankings reflect authentic discussions, the more trustworthy the answers; for advertisers, authentic rankings sometimes mean loss of control. Especially when a heavy-spending brand ranks poorly or receives negative feedback in AI answers, advertisers’ perception of the platform’s commercial value changes.

Doubao can push paid subscriptions partly because its monetization doesn’t rely on advertisers’ expected control over content outcomes; Xiaohongshu’s situation is the opposite. An inherent conflict exists between AI accuracy and commercial clients' interest expectations, making monetization paths more complex than Doubao’s.

Xiaohongshu recognizes this. According to 36Kr, the commercialization team once wanted the AI search team to avoid proactively generating shopping recommendation lists, but the feature was ultimately retained—showing Xiaohongshu didn’t fully sacrifice AI search authenticity under commercial pressure.

In the traditional search era, users had to flip through notes, compare comments, and form judgments from images and tones—a process that was itself content consumption and an ad display scenario. AI search tends to compress multiple notes into one answer, delivering users straight to conclusions and skipping the original content browsing chain. Efficiency gains and content consumption compression are two sides of the same coin.

For Xiaohongshu, this means the better AI search performs, the fewer ad display opportunities may exist per session—a risk that must be offset over time by increased usage frequency. Current data only proves frequency is rising, not that ad revenue is growing in sync.

Between Community, Monetization, and AI

Xiaohongshu’s AI constraints come not just from models and computing power—community atmosphere, commercial clients, and creator ecosystems must all be accommodated simultaneously.

The newly established AI division, Dots, reports directly to new president Ke Nan, who also oversees community, e-commerce, monetization, and the entire tech system. This design prevents AI from operating in isolation, ensuring technical investments directly serve the main business. But it also means Xiaohongshu’s AI can’t just focus on model capability and product growth—it must also consider community atmosphere, commercial clients, and creator ecosystems.

According to 36Kr, “Ask” initially covered only 1–2% of search requests. The team wanted to expand to 10%, but management, worried about impact, capped it at 3–4%. No one could prove before then that expanding to 10% wouldn’t alter the community atmosphere. On a platform that relies on community trust, this uncertainty was enough to make decision-makers hit the brakes.

A similar decision arose in discussions about integrating stronger models. In early 2025, the “Ask” team debated whether to integrate DeepSeek R1 like Tencent Yuanbao. The opposition wasn’t about data security or model quality but R1’s longer thinking chains, which increased answer generation time. In community search scenarios, users won’t wait—delayed responses might be perceived as product failures. This decision was later highly praised by participants, but it also showed Xiaohongshu AI’s optimization priority: speed and experience always trump answer richness.

In talent accumulation, Xiaohongshu aims to make AI an organizational capability. Since March, its internal tech team has frequently organized AI sharing sessions, involving non-AI roles in learning AI tools—some employees even need to build workflows with them. In its 2026 campus recruitment, Xiaohongshu also clearly leaned toward AI-related roles.

Since former CTO Qie Xiaohu’s departure around 2020, Xiaohongshu hasn’t filled the CTO position; its tech architecture is mainly overseen by two tech VPs. In contrast, ByteDance’s Doubao team exceeds 1,000 people with a systematic foundation in large model research; Kimi’s founding team continuously produces technically influential work in pre-training, inference optimization, and Agent architectures.

Meanwhile, current president Ke Nan has a background in consulting and finance, while Chief Product Officer Deng Chao studied architecture. At least from its leadership composition, Xiaohongshu’s core decision-makers aren’t technically oriented, making its AI strategy seem more product-focused than foundational model-driven.

However, given that the monetization of AI applications remains rife with uncertainty, adopting a conservative approach can also be seen as a form of risk management. ByteDance's Doubao has introduced paid subscription services, yet it grapples with challenges such as the fact that the average annual renewal rates for domestic consumer-end tools seldom surpass 30%, along with the structural predicament where "an increase in paying users leads to higher computing costs." According to reports, Kling attempted to secure independent financing at a valuation of $20 billion. However, in the realm of video generation, ByteDance's Seedance 2.0 already holds a dominant position, and the competition for computing power is far more fierce than any single challenge Xiaohongshu encounters.

Xiaohongshu is not oblivious to AI—it simply finds it more challenging than most companies to overlook the costs associated with integrating AI into its community.

Amidst the rapid advancement of AI, Xiaohongshu is not hastily spinning grand technological narratives. Instead, it prefers to first solidify its strongholds in areas such as travel, fashion, and lifestyle, where it boasts truly irreplaceable content density and user trust. AI's role is to transform these assets into more precise search results and efficient transaction channels.

But "stability" comes at a cost. By 2026, the industry's focus had shifted from chat products to agents, and the fear of missing out (FOMO) sentiment was surging across the internet. In this dynamic environment, Xiaohongshu's cautious approach will continue to face scrutiny: Where is its next growth trajectory? Can AI serve as the catalyst to break through its advertising and e-commerce growth bottlenecks, or will it merely remain an efficiency tool to streamline existing processes?

This question remained unanswered three years ago; three years later, Dots was established, and Xiaohongshu has simply positioned it more prominently.

*The question image and illustrations in the text are sourced from the internet.