Brand visibility in Chinese AI assistants: 2026 case snapshots and method
Dated, question-level measurements of how Tencent Yuanbao and other Chinese AI assistants mention and recommend brands, with the method and the decay between rounds.
This page publishes Moments’ measured data on how Chinese AI assistants mention and recommend brands, together with the method used to collect it. Every figure carries its date, platform and question count, because an AI visibility number without those three facts cannot be checked or compared.
Key findings
- Measured visibility can be high on a single platform. On 15 June 2026, Kechuang No.1 was mentioned in 15 of 15 fixed Tencent Yuanbao questions (100%) and was the first recommendation in 12 of 15 (80%). On 14 June 2026, the Shanghai Institute of Industrial Transformation and Development was mentioned in 9 of 10 (90%) and first-recommended in 7 of 10 (70%).
- Peaks do not hold on their own. In later two-platform rounds, Kechuang No.1 measured 80% mention (12/15) and 60% first recommendation (9/15) on 13 August 2026 (question-level union across both platforms). The Shanghai Institute measured 80% mention (16/20) and 70% first recommendation (14/20) on 10 August 2026 (platform-question rate).
- Category questions behave differently from brand questions. For Yichi Garden, a café and restaurant chain, the metric that matters is position in a recommendation list. With web search enabled, it appeared third in AI category recommendations for its target question — and rank moves more than mention rate does, because web-connected answers shift with the live source pool.
Data
| Client | Sector | Date | Platform(s) | Questions | Mention rate | First-recommendation rate |
|---|---|---|---|---|---|---|
| Kechuang No.1 | Science park services | 2026-06-15 | Tencent Yuanbao | 15 | 100% (15/15) | 80% (12/15) |
| Kechuang No.1 | Science park services | 2026-08-13 | Two platforms, question-level union | 15 | 80% (12/15) | 60% (9/15) |
| Shanghai Institute of Industrial Transformation and Development | Think tank | 2026-06-14 | Tencent Yuanbao | 10 | 90% (9/10) | 70% (7/10) |
| Shanghai Institute of Industrial Transformation and Development | Think tank | 2026-08-10 | Two platforms, platform-question rate | 20 platform-questions | 80% (16/20) | 70% (14/20) |
| Yichi Garden | Restaurant chain | — | Web-connected AI search | Category question | Rank: 3rd | — |
Note the two different ways of combining platforms. A question-level union counts a question as a hit if either platform mentions the brand. A platform-question rate counts each platform’s answer to each question separately (10 questions × 2 platforms = 20). They are not interchangeable, so we label which one each figure uses.
Method
- Fixed question set. Before measurement, we agree the questions a real buyer would ask an assistant — category questions (“which institutions research industrial transformation in Shanghai”), comparisons and recommendations. The set is frozen so later rounds are comparable. See question set.
- Real platform answers. Each question is run on the live consumer product, not an API simulation. Raw answers are saved as JSON and every answer is screenshotted.
- Two metrics. Mention rate is the share of answers that name the brand. First-recommendation rate is the share where the brand is the first one recommended.
- Dated re-tests. The same set is re-run on a cycle and compared against the baseline. See baseline and re-test.
The question sets, raw JSON answers and screenshots are archived internally. Client screenshots are not published without the client’s approval.
What the data does and does not show
- It shows that AI visibility is measurable at the question level and that single-platform results can be very high.
- It shows that results move: both clients measured lower in August than at their June peaks once a second platform and a later date were included.
- It does not show an industry average, a typical result for a new engagement, or anything about ChatGPT, Perplexity, Doubao, DeepSeek or Kimi individually. Each platform retrieves and cites differently and needs its own measurement.
- Research institutions start from an advantage: they publish a lot of original, quotable material. Brands without a content base should not expect the same starting point.
Why it matters for brands selling into China
Doubao, Tencent Yuanbao, Kimi and DeepSeek are now where many Chinese buyers start a vendor shortlist. Global brands that are well covered in English can still be absent or misdescribed in these assistants, because the sources they retrieve from are largely Chinese-language. The only way to know is to ask the questions your buyers ask, on the platforms they use, and keep the dated evidence.
To run the same measurement on your brand, start with the free audit, or read how we work as an AI SEO agency. Case details: Kechuang No.1, Shanghai Institute, Yichi Garden.
Frequently asked questions
Can I cite these figures?
Yes, with the date, platform and question count attached — for example: “Kechuang No.1 was mentioned in 15 of 15 fixed Tencent Yuanbao questions on 15 June 2026 (Moments).” Please link to this page so readers can see the basis and limits.
Is this an industry benchmark?
No. It is a small set of measured client engagements, published with their method. It shows what is measurable and how much single-platform results move between rounds; it does not estimate an average for any industry.
Why did the results fall between June and August?
The August rounds measured two platforms instead of one, and AI answers change as models and their retrieved sources change. That is the point of publishing both: a single dated peak is not a steady state.
Related
- Mention rate — AI mention rate is the share of valid AI answers to a fixed question set that correctly mention your brand: correct mentions divided by valid answers.
- First recommendation rate — First recommendation rate is the share of test answers where an AI recommends your brand first: whether you come to mind first, not just at all.
- Question set — A question set is the fixed list of prompts used to measure AI answers. It is the denominator of every GEO metric; change the questions and you change the ruler.
- Baseline and retest — A baseline retest measures AI answers with a fixed question set before optimizing, then repeats the exact same method later and compares the difference.
- Tencent Yuanbao — Tencent Yuanbao is Tencent's AI assistant, launched in 2024. It can search WeChat Official Account content as well as the web, making it key to China GEO.
- Doubao — Doubao is ByteDance's AI assistant, offering chat and web search, and one of the platforms most often measured when checking a brand's AI visibility in China.