China's Anthropomorphic AI Crackdown: Why Doubao and Qwen's Agent Rollbacks Change Brand Discovery in 2026

On July 15, 2026, China's new rules governing anthropomorphic AI interaction services took effect — and in the weeks before the deadline, two of the country's largest AI platforms moved fast. ByteDance's Doubao and Alibaba's Qwen began disabling humanlike agent features and user-created AI personas, resetting a product category that had been one of the fastest-growing surfaces in Chinese consumer AI. For international brands, the regulatory shift is more than a compliance story. It quietly redraws the map of where and how Chinese consumers will encounter brand recommendations — and it concentrates even more discovery power in the exact query surfaces that Generative Engine Optimization (GEO) targets.

Here is the core finding for brand marketers: the new rules restrict sustained emotional interaction agents, but explicitly exempt knowledge Q&A, customer service bots, workplace assistants, and education tools. In other words, Beijing just pruned the companion-AI branch of the ecosystem while leaving the recommendation-and-answer branch untouched. The queries that decide whether your brand gets cited — "which luxury skincare brand is best for sensitive skin," "recommend a business hotel in Shanghai" — are not only unaffected; they are now a proportionally larger share of what these platforms do.

What the Regulation Actually Covers

The rules apply to AI services that simulate human personality traits, thinking patterns, and communication styles to provide sustained emotional interaction. That definition captures the AI-companion and roleplay-agent features that Doubao in particular had leaned into — user-created personas, long-running emotional chat threads, and character-style agents that blurred the line between tool and relationship.

Crucially, the regulation carves out functional AI: customer service, knowledge Q&A, workplace productivity, education, and scientific research tools remain outside its scope so long as they avoid sustained emotional interaction. Doubao and Qwen responded in early July by disabling or restricting the affected agent features ahead of the deadline, a pattern consistent with how Chinese platforms typically front-run regulatory enforcement.

This split matters enormously for brand visibility because it separates the two roles Chinese AI apps have been playing. One role is the companion — sticky, emotional, high-engagement, but low commercial-intent. The other is the answer engine — the surface where consumers ask comparison questions and where brand citations happen. The regulation constrains the first and leaves the second fully operational.

The Scale at Stake

The platforms involved are not niche. Doubao reached roughly 240 million monthly active users by April 2026, making it China's largest consumer AI app. DeepSeek commands over 180 million MAUs as of Q2 2026, Baidu's Ernie Bot sits at around 210 million, and Qwen has crossed 150 million. For comparison, ChatGPT's global user base was approximately 200 million MAUs in early 2026 — meaning individual Chinese platforms now match or exceed the largest Western LLM's worldwide footprint within a single market.

hubGEO's own tracking across six Chinese models (Doubao, Kimi, DeepSeek, Qwen, Wenxin, Hunyuan) consistently shows that brand visibility scores diverge sharply by platform even for the same brand in the same category. A regulatory event that changes how users engage with one platform — or shifts engagement between platforms — therefore has direct, measurable consequences for which model's brand rankings matter most to your China strategy.

Three Ways the Reset Changes Brand Discovery

1. Query mix shifts toward commercial and informational intent. Companion-style usage generated enormous session time but few brand-relevant queries. With emotional-agent features curtailed, the remaining usage on Doubao and Qwen skews harder toward the utility queries — product research, comparisons, how-to questions, travel planning — where brand citations actually occur. A higher share of platform activity now flows through surfaces where GEO signals determine outcomes. For brands, the addressable query pool did not shrink; it got more concentrated.

2. Brand-built personas and unofficial "brand agents" lose their lane. Some marketers had begun experimenting with character-style brand agents inside Chinese AI apps — persona-driven ambassadors designed to build affinity through ongoing conversation. That playbook is now legally fraught. Any brand activation that resembles sustained emotional interaction risks falling inside the regulated category. The safer, regulation-proof channel is the one GEO has always targeted: being the brand the model itself cites when a user asks a neutral question. Earned citation beats owned persona under the new rules.

3. Platform trust positioning favors the answer-engine framing. Doubao and Qwen are now incentivized to emphasize their utility identity — search, Q&A, productivity — in both product design and public messaging. Expect both platforms to invest further in retrieval quality, source citation, and structured answers, because that is the growth lane regulators have blessed. Platforms that improve retrieval and citation mechanics tend to reward brands with strong, machine-readable evidence layers: authoritative Chinese-language content, consistent entity data, and third-party corroboration on the sources each model actually retrieves from.

Who Gains and Who Loses

DeepSeek and Kimi deserve particular note. Neither built its user base on companion features: DeepSeek's identity is technical reasoning and Kimi's is long-document research. The regulation effectively validates their positioning, and enterprise or high-consideration B2B queries — due diligence, supplier evaluation, whitepaper analysis — continue uninterrupted. Brands selling into considered-purchase categories should treat Kimi and DeepSeek visibility as unaffected and, if anything, strategically safer ground.

What to Watch Through Q3 2026

Two second-order effects are worth monitoring. The first is enforcement breadth: early July's feature rollbacks were voluntary front-running, and the real test comes when regulators clarify edge cases — for example, whether shopping assistants with friendly, personality-inflected tones fall inside "sustained emotional interaction." A broad reading could force further product changes and another round of user-behavior shifts. The second is where displaced engagement goes. Users who valued companion features will not simply vanish; some engagement will migrate to compliant formats, to smaller offshore apps, or back into the answer-engine surfaces of the same platforms. Analysts advising CMOs on China entry already recommend shifting the majority of China-focused search budgets toward AI visibility and super-app presence over traditional SEM — the July reset strengthens that case, because it removes the one consumer-AI surface where GEO signals did not apply.

The Deeper Signal: China Wants AI to Be Infrastructure, Not Intimacy

Zoom out and the July 15 rules fit a consistent regulatory pattern. Chinese authorities have repeatedly steered consumer internet platforms away from engagement-maximizing mechanics (gaming curfews, algorithm transparency rules, livestream tipping limits) and toward utility framing. Applying that lens to AI, the message is clear: AI as a tool for answers, work, and learning is encouraged; AI as a synthetic relationship is suspect.

For brand marketers, this reduces a real strategic uncertainty. Six months ago, it was plausible that Chinese consumer AI would evolve toward persona-driven, emotionally sticky interfaces where brand discovery happened through character interactions — a world requiring entirely different marketing skills. That future has now been regulated away, at least in China. The discovery model that remains is the one GEO practitioners already understand: a user asks a model a question, the model retrieves and synthesizes evidence, and brands with the strongest machine-legible presence get named.

Takeaway for Brand Marketers

First, reallocate any budget or roadmap earmarked for AI persona experiments in China toward citation-layer fundamentals: authoritative Chinese-language content on the source platforms each model retrieves from, consistent brand entity data, and third-party validation (industry rankings, professional reviews, knowledge-base entries). These assets work on the exempt, growth-blessed side of the regulatory line.

Second, re-baseline your brand's visibility scores now. Regulatory resets change user behavior and platform priorities, and citation patterns can drift as Doubao and Qwen re-tune their products around utility use cases. A score measured in Q2 2026 is not a safe assumption for Q4. Track all six major models rather than extrapolating from one — cross-platform score gaps remain the most common blind spot we see in brand audits.

Third, treat customer-service AI as the compliant brand-owned channel. The regulation explicitly exempts customer service bots. A well-built, genuinely useful brand assistant — order support, product Q&A, store lookup — remains fully legal and is now the clearest sanctioned way to own an AI touchpoint in China, provided it stays functional rather than emotional.

The platforms consumers use to discover brands in China just got structurally simpler: fewer companions, more answer engines. For brands that have invested in being citable, that is a quiet win.

Related: Explore current brand visibility scores across all six Chinese AI models on our brand tracker, or see how the Doubao–DeepSeek duopoly concentrates brand discovery.