For years, AI in China was a policy story. Five-year plans, research papers, facial recognition pilots. That period is over. Since 2023, artificial intelligence in China has become a consumer habit, and the numbers behind it are now large enough to change how a foreign brand gets found, gets read and gets answered in this market.
This article was updated in 2026. Most of what was written about Chinese AI before the large language model wave is now wrong, so I rebuilt the data and kept only what still holds.
Written by Olivier Verot, founder of GMA. I have been running digital campaigns from Shanghai since 2012. Since 2024 my teams have been rebuilding client content so it gets picked up by DeepSeek, Doubao and Baidu’s AI answers, and I have watched what works and what does not on live accounts.
China’s AI Industry in 2026: the Real Numbers
Start with adoption, because that is the part most foreign executives underestimate. According to the 57th statistical report from CNNIC, published in February 2026, China counted 602 million generative AI users as of December 2025, up 141.7% in one year. Penetration reached 42.8% of the population, against 17.6% a year earlier. China’s total internet user base is now 1.125 billion. Those figures are published by the State Council and relayed by Global Times.
In roughly twenty-four months, a country went from almost no chatbot usage to four adults in ten. No Western market moved that fast.
On the industry side, the Minister of Industry and Information Technology told the March 2026 session of the National People’s Congress that China’s core AI industry passed 1.2 trillion yuan (around 170 billion USD), with more than 6,200 AI companies registered. The 2024 figure was around 900 billion yuan, so the growth is real, not a rounding effect.
Who Actually Uses What
QuestMobile tracks monthly active users on AI-native apps. Their Q1 2026 report puts the whole category at 440 million monthly actives in March 2026. Here is the ranking.
| App | Owner | MAU, March 2026 | What it is used for |
|---|---|---|---|
| Doubao 豆包 | ByteDance | 345 million | General assistant, voice, tightly linked to Douyin content |
| Qwen / Qianwen 千问 | Alibaba | 166 million | Assistant plus shopping tasks inside Taobao |
| DeepSeek | DeepSeek (Hangzhou) | 127 million | Reasoning, research, technical questions, high stickiness |
| Yuanbao 元宝 | Tencent | ~50 million (June 2026) | Assistant inside the WeChat environment |
| Ernie 文心一言 | Baidu | Not disclosed separately | Powers Baidu’s AI answers on the search results page |
Two things matter in that table. Doubao wins on distribution because ByteDance pushes it inside apps people already open. DeepSeek wins on intensity: QuestMobile measured 41.7 sessions per user per month, second only to Doubao’s 54.8. People do not open DeepSeek to play. They open it to decide something.
The Hardware Constraint
US export controls on advanced accelerators have not stopped Chinese model development, but they shaped it. Two visible consequences.
- Chinese labs optimised hard for training and inference efficiency. DeepSeek built its reputation on doing more with less compute, and that pressure is why Chinese models are cheap to call by API.
- Domestic silicon took the volume. Huawei is raising Ascend output substantially in 2026, and Beijing is pushing state-linked data centres toward domestic chips. Nvidia’s H200 shipments to China have been stuck in regulatory limbo on both sides, as Tom’s Hardware reported in 2026.
The practical effect for a marketer is boring and useful: Chinese model access is cheap. Running a Chinese-language chatbot on Qwen or DeepSeek costs a fraction of what an equivalent Western stack costs, which is why small brands here automate customer service that European brands still do by hand.
The Rules You Have to Live With
China regulates generative AI more concretely than most people abroad assume. Three points.
- Filing is mandatory. Any public-facing generative AI service in China must be filed with the Cyberspace Administration. As of 30 June 2026, CAC public notices showed 988 filed services and 598 registered AI applications or features. If you plan to embed a chatbot in your China website or mini-program, this is a compliance line item, not a detail.
- AI content must be labelled. The Measures for Labelling AI-Generated Content, plus the mandatory national standard GB 45438-2025, took effect on 1 September 2025. Text, images, audio and video generated by AI need a visible label and hidden metadata. Bird & Bird published a clear comparison with the EU AI Act.
- Platforms enforce it. Douyin, Xiaohongshu and WeChat check and tag uploads. If you generate product visuals with AI and post them unlabelled, the risk is not a fine first, it is reduced distribution.
What This Changes for a Foreign Brand Selling in China
This is the part that makes the AI story relevant to a marketing budget rather than to a tech newsletter. Four concrete areas.
1. Getting Cited in the Models (GEO)
When a Chinese buyer asks Doubao or DeepSeek “which European brands make industrial dosing pumps”, the model returns a short list of names. You are on it or you are not. There is no page two to climb to.
The mechanism is simple once you see it. Chinese models pull from Chinese-language sources they can crawl and from search partners: Baidu and Sogou pages, Zhihu threads, Baijiahao and WeChat articles, industry portals, Xiaohongshu notes. Your English website in Frankfurt is invisible to that pipeline. What gets you cited is a body of Chinese-language content, published on domains the models actually index, that states plainly what you make, for whom, and with what specification. Comparison pages and question-shaped pages work better than brochure pages, because they match the phrasing of the query. We covered the full method in why your website is invisible to DeepSeek.
One warning. Everyone is doing this now, so the easy gains from 2024 are gone. I wrote about that saturation in everyone in China uses AI now.
2. Automated Customer Service
Chinese buyers expect a reply in minutes, at 11pm, in Chinese. That expectation broke a lot of European brands before AI made it affordable. Tencent integrated DeepSeek into WeChat, which put a capable assistant one tap away from the average user, and it changed the baseline of what people consider a normal response time. See what the WeChat and DeepSeek integration means for brands and the follow-up on WeChat’s AI agent rollout.
What works: a bot handling the first layer, that is price ranges, lead times, certifications, MOQ, shipping, then handing over to a human the moment the question gets specific. What fails: a bot pretending to be a salesperson. Chinese users spot it in two exchanges and leave.
3. Chinese Content Production
Machine translation into Chinese is now good enough for a first draft and still bad enough to embarrass you if you publish it raw. The failures are not grammar, they are register and category vocabulary. A model will translate your product line correctly and still use a word nobody in that Chinese industry uses.
The workable split: AI drafts the volume, a native editor fixes the terminology and adds the local proof points. That cuts content cost by roughly half without turning your brand into noise. Do the same for Xiaohongshu, where search behaviour is now closer to a search engine than a feed, as explained in our guide on Xiaohongshu as a search engine.
4. The Limits
Three, honestly stated. Chinese models hallucinate brand facts, including prices and certifications, and you cannot correct them directly, you can only publish better sources. AI citations are unstable and change with model updates, so treat them as a channel to monitor monthly, not a rank to lock in. And AI does nothing for a brand with no e-reputation: if nobody in China has written about you, the model has nothing to quote.
A Client Example: Felipe, Portuguese Industrial Equipment
Felipe runs a mid-size Portuguese manufacturer of food-processing equipment. He came to us in early 2025 with a specific complaint: his Chinese distributor said Chinese buyers “had never heard of them”, while the same buyers named two German competitors without hesitating. His China website was getting around 400 visits a month and produced 3 qualified leads in a year.
He had already tried two things. A Baidu SEM campaign, which brought clicks at a high cost and no conversion because the landing page was a translated PDF-style brochure. And a Chinese agency that pushed 40 short Baijiahao articles in three months, all generic, all near-duplicates. Neither moved anything.
What worked was narrower. We built 22 Chinese pages answering the exact questions buyers type: equipment sizing per production volume, EU hygiene certification and what it means for a Chinese plant, spare parts lead time to China, comparison with the two German brands he kept losing to. Published on his China site plus Baijiahao and Zhihu. Then a WeChat article series and a small set of Xiaohongshu notes from two food-industry accounts.
Why it worked: the models needed something specific and Chinese to quote, and the comparison pages matched the shape of the question buyers were asking. By March 2026, the brand name appeared in DeepSeek and Doubao answers for four of his eleven target queries, up from zero. Monthly qualified inbound went from roughly 0.3 to 5 per month. Two turned into distributor contracts. That took thirteen months, not one quarter.
Where AI Already Sits in Chinese Retail
Taobao, Tmall, JD and Douyin have used AI recommendation for years. It filters the catalogue for each shopper based on browsing, purchase history and demographics. That is not new and it is not where the change is. The change is that the recommendation layer now talks: Alibaba’s assistant answers product questions inside the shopping flow, and Douyin’s search returns generated summaries above the video results.
Virtual try-on remains a real use case, especially in cosmetics. Online retailers use augmented reality so a customer can test a shade or a frame before buying, which cuts returns and raises time on page.
Healthcare is the other sector where Chinese AI deployment is ahead of Europe, mostly in imaging triage and elder care systems, following the 2017 Next Generation Artificial Intelligence Development Plan (新一代人工智能发展规划) that made medical AI an explicit target.
How China Got Here
None of this happened by accident. China has built a national industrial policy agenda around technology since 2006, crossing a first threshold in 2010 with the Strategic Emerging Industries program. Three features of that policy explain the current position.
- A deliberate bet on artificial intelligence, endorsed at the top of government in the 2016 Innovation-Driven Development Strategy and formalised in the 2017 Next Generation Artificial Intelligence Development Plan.
- Reliance on private companies rather than state-owned ones. Beijing funds and steers, but ByteDance, Alibaba, Tencent and DeepSeek are the ones shipping models. That is why adoption moved at consumer speed.
- Cities as test beds. Smart city programmes gave Chinese firms deployment volume, and deployment volume gave them data.
The 2015 “Made in China 2025” plan set the direction: move from low-end manufacturing to high-end production and cut dependence on foreign suppliers. Ten years on, the chip situation shows how literally that goal was taken.
The Main Chinese AI Companies
- ByteDance, owner of Douyin and TikTok, now the largest AI assistant in China with Doubao. Machine learning has been its core product since 2012, and it distributes models through apps people already use.
- Alibaba, whose Qwen family is the most widely used open-weight model line in the world and the second assistant in China by MAU.
- DeepSeek, founded in Hangzhou, the model that made Chinese AI credible internationally on cost and reasoning quality.
- Baidu, whose Ernie models power the AI answers on China’s largest search engine. For a brand doing Baidu SEO, that is the most direct link between search and AI visibility.
- Tencent, with Yuanbao and DeepSeek integrated into WeChat, which is the distribution channel that matters most for private domain marketing.
- iFlytek (科大讯飞), founded in 1999 in Anhui, still the reference for Chinese speech recognition, translation and education products.
- SenseTime and UBTECH Robotics, both in computer vision and robotics rather than consumer language models.
Things to Know Before Entering the Chinese AI Market
The Perception of AI in China
Chinese consumers are far less anxious about AI than Europeans. They adopt features first and ask questions later, which is exactly what the jump from 17.6% to 42.8% penetration in twelve months describes. That works in your favour: an AI-assisted service on your WeChat account is read as modern, not intrusive.
E-Reputation Still Decides Everything
Chinese buyers are suspicious of brands they have never seen mentioned. They check reviews, forum threads and third-party articles before they trust a website. AI made this sharper, not weaker: the models are trained on the same third-party content the buyer checks manually. A brand with no Chinese footprint is invisible twice, to the human and to the machine. Protecting that footprint matters, which is why we also cover how to protect your brand in China.
Practical Starting Points
- A Chinese-hosted site with proper Baidu indexing. Without it, nothing else compounds.
- Question-shaped Chinese content on Baijiahao, Zhihu and WeChat, published steadily rather than in one batch.
- Xiaohongshu notes from real accounts in your category, because that is where product search now starts for consumer brands.
- A first-layer AI assistant on WeChat, filed and labelled correctly, with human handover.
- Monthly monitoring of what DeepSeek, Doubao and Qwen say about your brand and your category. Track it like you track rankings.
FAQ
How many people in China actually use generative AI?
602 million as of December 2025, according to CNNIC’s 57th statistical report published in February 2026. That is 42.8% of the population, up from 17.6% a year before. On the app side, QuestMobile counted 440 million monthly active users on AI-native apps in March 2026, rising to 499 million in June 2026. So it is not an early-adopter niche, it is mainstream usage across age groups.
Can I use ChatGPT-based tools for my China marketing?
For internal drafting, yes, nothing stops you. For anything public-facing inside China, no. A generative AI service offered to the Chinese public must be filed with the Cyberspace Administration, and foreign models are not. Practically, if you want a chatbot on your China site or WeChat mini-program, you build it on a filed domestic model such as Qwen, DeepSeek or a Tencent Cloud service.
Do I have to label AI-generated visuals on Chinese platforms?
Yes. Since 1 September 2025, the Measures for Labelling AI-Generated Content and the GB 45438-2025 standard require a visible label plus hidden metadata on AI-generated text, images, audio and video. Douyin, Xiaohongshu and WeChat detect and tag it themselves. Posting unlabelled AI visuals usually costs you reach before it costs you a penalty, so treat labelling as a distribution issue.
How long does it take to appear in Chinese AI answers?
Six to twelve months for a brand starting from zero Chinese content, in our experience across B2B accounts. The models need indexed Chinese sources before they can quote you, and those sources need to be published, crawled and corroborated. Anyone promising results in six weeks is selling you article volume. Volume alone does not get cited.
GMA is a Shanghai-based agency. We build the Chinese-language content and the Baidu foundation that make a foreign brand quotable by DeepSeek, Doubao, Qwen and Baidu’s AI answers. We also set up compliant AI customer service on WeChat, with human handover where it counts. If you want to know what Chinese models currently say about your brand, contact us and we will run the check.
