The short answer: When Chinese AI assistants (Doubao, DeepSeek and Tongyi Qianwen) recommend brands, they rely far more on what others say about a brand, such as Douyin videos, trade media and Baidu Baike, than on what the brand publishes itself. They also name brands early, while customers are still asking how to choose. Korean brands selling in China should start by making their creator briefs and product pages readable by AI.
The evidence comes from a report published on July 31, 2026 by Chinese research firm Growthbox (增长黑盒) with Emergine, an AI-visibility monitoring platform. They ran about 5,000 real AI search conversations, with live web search enabled, through three Chinese models (Doubao, DeepSeek, Tongyi Qianwen) and two English models (ChatGPT, Gemini), then coded every answer for the brands it named and the pages it cited.
1. Chinese AI asks around instead of checking the source
The report calls overseas models “verifiers” (查证型) and Chinese models “inquirers” (打听型). Overseas models check specifications on company pages; Chinese models gather what trade media, encyclopedias and short videos say.
| Sources cited by AI | Chinese models | Overseas models |
|---|---|---|
| Official company sites | 30.2% | 76.1% |
| Other sources (video, media, encyclopedias, etc.) | 69.8% | 23.9% |
| Share of evidence from official sites when leading brands are mentioned | about 16–41% | about 61–85% |
Among non-company sources, Chinese models most often cited short video and social platforms, then trade media, news and press releases, reviews and rankings, and encyclopedias. The pattern is strongest in consumer categories: for leading consumer-electronics brands, only about 17% of the evidence came from official sites.
2. On Doubao, Douyin is the top cited source
Doubao, China’s most widely used AI assistant, cited Douyin more than any other source in most industries: about 66% of citations in consumer electronics and about 45% in sports and outdoor.
The key point is that AI cites the text it can extract from a video, not the visuals: titles, subtitles, and the specifications, steps and comparisons a creator says aloud. The report describes purely promotional creator videos as carrying “almost zero” usable information for AI.
3. Brands appear as soon as customers ask how to choose
- Even at the research stage, 50.2% of Chinese model answers named a specific brand, or 62.7% when weighted by topic popularity.
- Within that stage, definition questions (“What is X?”) named brands 22.1% of the time, while guidance questions (“How do I choose, use or avoid problems?”) did so 78.7% of the time (popularity-weighted).
- Answers listed an average of 3.08 brands at the research stage and 4.05 at the selection stage. The top brand in each industry appeared in 29.6% of conversations, and the top three together in 73.2%. Visibility in China is more concentrated among leaders than overseas.
4. AI adds a caveat after the praise
In Chinese models, 7.9% of brand assessments added a qualifying “but” statement. Weighted by topic popularity, the share doubled to 17.4%: the more people ask about a topic, the more caveats AI adds. Of these statements, 62.9% pointed to a specific weakness; only 8.9% suggested another brand was better. Caveats were most common in hotels and travel (25.3%), enterprise services, sports and outdoor (15.0%) and education. The report notes that, apart from enterprise services, these are all consumer categories where people buy an experience.
What Korean brands should change first
The steps below adapt the report’s action framework to Korean consumer brands. They are CMB’s recommendations.
| Action | What to do |
|---|---|
| 1. Use one Chinese brand name everywhere | If the Chinese name differs across the Tmall store, Baidu Baike, the WeChat official account and creator posts, AI can treat them as different brands. The report also advises keeping encyclopedia entries consistent with the official site. |
| 2. Put product information in text | Provide ingredients, size, order of use and suitable skin types as text and tables, not only as images. The report notes that AI cannot extract specifications presented as images. |
| 3. Add informative elements to creator briefs | Ask creators to include the product’s official Chinese name, key ingredients, order of use and selection criteria in video titles, subtitles and narration. Mood-only videos are rarely cited. |
| 4. Target questions with conditions | Rather than competing head-on with leaders for “best” questions, aim to be named in guidance questions with conditions such as skin type, season or situation. |
| 5. Check the same questions monthly | Ask Doubao, DeepSeek and Tongyi Qianwen the same 5–10 questions every month and record whether you are named, which brands appear alongside you and what caveats are attached. Answers vary, so do not judge from a single check. |
| 6. Answer repeated weaknesses directly | If AI keeps attaching the same weakness, publish an FAQ or usage guide addressing it on your official channels before producing competitor comparisons. |
Example check questions (written by CMB): 敏感肌夏天用什么防晒比较好？ / 韩国面膜怎么选？ / ○○（brand’s Chinese name）怎么样？ Mix guidance questions with brand-assessment questions.
Caveats
- The ten industries studied (hotels and travel, consumer electronics, logistics, enterprise services, automotive, finance, education, manufacturing, ICT, sports and outdoor) do not include beauty. Applications to beauty are CMB’s interpretation based on the consumer-category results.
- Popularity-weighted figures use the search-engine volume of related keywords as a proxy, because AI platforms do not disclose how often each question is asked.
- The study was produced with an AI-visibility monitoring company, and the report ends by promoting a GEO course. Read the figures as results for this sample and period.
