We Asked Four Chinese AI Assistants the Same Question. Not One Brand Made All Four Lists.
We put the same 13 questions in Chinese to Doubao, DeepSeek, Kimi and Yuanbao on 30 October 2026. Across 46 answers, no single brand was recommended by all four.
→ Two of the four actively removed the best-known brands. Doubao wrote that Byredo, Le Labo and Diptyque were "too popular now, so I am not putting them in the main recommendations."
→ For brands with no official China channel, the assistants said so out loud — and then sent buyers to daigou, Farfetch and duty free.
What's AI marketing exactly?
One of our clients is a marketing lead at a company we work with. She is smart, experienced, and controls a real budget. She asked about our GEO service — generative engine optimisation, meaning whether your brand appears inside AI-generated answers — and mentioned that her global team was already "running GEO, spending fifteen dollars a day."
We were confused at first. A few questions later, it was clear what she meant: what their global team had done was buy a placement inside the AI's answer, like a search ad, tested with a minimal budget in another market.
However, a paid placement inside a Chinese AI assistant's generated answer does not exist today. What the Chinese assistants do instead is recommend products out of the commerce ecosystem they already own.
In the West, the AI assistant mostly arrived as a standalone product in search of a business model, so the money is being built around and inside the answer: subscriptions first, then advertising formats layered onto the response. In China, the assistant arrived inside a company that already owned the transaction. Doubao sits in ByteDance, next to Douyin's e-commerce. Yuanbao sits in Tencent, next to WeChat Mini Programs. Qwen sits in Alibaba, next to Taobao.
Another interesting thing: a few weeks ago we were talking with one of our partners — a marketing professional with more than thirty years in personal branding, working mainly in Western markets. The conversation was about where Chinese and Western marketing actually diverge. She was emphatic about Google visibility, and she tells her clients to search their own name and read what comes back. Good advice. So we asked whether GEO was something her market had started to deal with.
She had never heard the term.
Two people, both professionals, both of whom are paid to think about market visibility. One was ready to buy something that does not exist. One had not heard of it at all.
That made us ask: if professionals cannot see inside those answers, what is actually in there?
So we ran a test.
Where do Chinese buyers actually look now?
Searching your own name is a normal professional habit. Most brand marketing teams do the same with their brand: type it into Google, see what comes up, check the first page.
Now move that habit to where your Chinese buyers are.
Chinese consumers used to look on social platforms — Xiaohongshu(RedNote), Douyin, WeChat — or to go straight to a marketplace. Increasingly, they ask an AI assistant instead — which is what most people now mean by AI search in China. This article looks at four consumer AI assistants (chat apps) built on Chinese LLMs: Doubao (ByteDance), DeepSeek, Kimi (Moonshot AI) and Yuanbao (Tencent).
The question is typed as a sentence: "I want to buy a niche perfume, which brands would you recommend?"
And the answer that comes back is not ten blue links you can scan. It is one paragraph, with four or five brand names in it.
If you are not in that paragraph, you are not in the consideration set. There is no second page to be on.
How the answer actually gets built
Worth being precise here. When someone types a question, three things can happen: the model may answer from what it already learned during training. It may run a live web search and build the answer from what it finds. Or it may do both, blending stored knowledge with fresh pages.
Users cannot see which path produced the answer, and the path is not stable. In our test, Doubao reported "searched 4 keywords, referenced 21 sources" on one question and "6 keywords, 27 sources" on another. DeepSeek showed "searched 14 web pages" on one question and 13 on another.
This means two things.
Not every answer involves a live search. So "I was not cited" and "I was not found" are different problems.
The citation list is not the full picture. It is what the system chose to show.
Why we tested Doubao, DeepSeek, Kimi and Yuanbao
| AI assistant | Owner | Content ecosystem | Monthly active users (MAU) |
|---|---|---|---|
| Doubao (豆包) | ByteDance | Douyin, Douyin Baike, Douyin e-commerce | 382m |
| DeepSeek | DeepSeek (independent) | None of its own | 130m |
| Yuanbao (元宝) | Tencent | WeChat Official Accounts | 49m |
| Kimi | Moonshot AI | None of its own | 7m |
| Qwen (千问) — not tested | Alibaba | Taobao/Tmall, Alipay Mini Programs, Amap, Ele.me | 167m |
MAU as of May 2026, from QuestMobile's 2026 first-half report, published 14 July 2026. China's AI-native apps reached 499m monthly active users in total, up 85.4% year on year.
We did not pick the four largest. We picked for ecosystem variety: two assistants owned by companies that also own enormous content and commerce platforms, and two with no content pool of their own.
The honest gap is in the last row. Qwen is Alibaba's assistant and the second largest in the market, and we did not test it. Alibaba owns Taobao, and Taobao links turn up repeatedly in what we did measure — so the one ecosystem we most wanted to see is the one missing. It goes in the next round.
Finding 1: four assistants, four different shortlists
We asked all four: "I want to buy a niche perfume, which brands would you recommend?"
| Assistant | Brands returned, in order |
|---|---|
| Doubao | Parfum d'Empire, Liquides Imaginaires, Nicolai, D.S. & Durga, Imaginary Authors, Goldfield & Banks, Amouage, Xerjoff, Ormonde Jayne, Yuyi |
| DeepSeek | Le Labo, Aesop, Diptyque, Maison Margiela Replica, L'Artisan Parfumeur, Tsu Lange Yor, Matiere Premiere, Stora Skuggan, OCO, Hunayn, AROMAG, Shuzhi Youwei |
| Kimi | Le Labo, Byredo, Diptyque, Aesop, Perfumer H, Liquides Imaginaires, Nonfiction, Tamburins, Santa Maria Novella, Penhaligon's, Lalique, Acqua di Parma, L'Artisan Parfumeur, Serge Lutens |
| Yuanbao | Byredo, Diptyque, Le Labo, Frédéric Malle, Miller Harris, To Summer |
Thirty-odd brands across four answers. The overlap is two: Diptyque and Le Labo, on three lists out of four.
Not one brand appeared on all four lists.
Finding 2: each assistant draws heavily on its own ecosystem
Yuanbao was the only one of the four whose exports carried full URLs, 332 of them. So it is the only one we can count properly.
| Source type | Links | Share |
|---|---|---|
| WeChat Official Accounts (mp.weixin.qq.com) | 39 | 11.7% |
| Sina properties | 40 | 12.0% |
| Taobao | 26 | 7.8% |
| JD | 25 | 7.5% |
| Fragrance verticals (Fragrantica and similar) | 16 | 4.8% |
| Maigoo / CNPP list sites | 12 | 3.6% |
| SMZDM | 8 | 2.4% |
| Encyclopaedia pages | 6 | 1.8% |
| Other portals | 15 | 4.5% |
| Everything else, incl. brand sites and overseas pages | 145 | 43.7% |
| Total | 332 | 100% |
Two things stand out.
Yuanbao is Tencent's assistant, and its single largest named source is WeChat Official Accounts. Tencent's assistant goes back to Tencent's own content pool. Doubao, ByteDance's, leaned on Douyin Baike and a Douyin video.
And then the one worth a second look: Taobao and JD together account for 51 links, many of them affiliate URLs — union-click.jd.com, s.click.taobao.com.
The assistant is citing shopping links as evidence. That is worth knowing before you assume your Western playbook transfers.
One more, in a vertical category, vertical communities beat the big platforms. The citations here were not led by the big general marketplaces and deal sites. They were led by Nosetime and Fragrantica — a specialist perfume community and a specialist fragrance database.
The narrower your category, the more a specialist community matters.
Finding 3: your own website wins one question and loses the other
| Question type | Does the brand's own site get cited? |
|---|---|
| "What is brand X?" | Yes — usually first or second |
| "Which brands would you recommend?" | Almost never |
On the brand questions, official pages dominated: diptyqueparis.com three times, lelabofragrances.com twice plus an Estée Lauder investor page, nishane.com twice, charlottetilbury.com twice plus the Puig group page.
On the category questions, the brand sites were simply absent. Rankings, reviews and community profiles took every slot.
So the two jobs separate:
Getting the AI to describe you correctly is work on your own pages. You control that. Getting the AI to recommend you at all is work on other people's pages. You do not control it, but you can earn it.
Most teams only do the first and then wonder why nothing changed.
Three checks you can run this week
You do not need a vendor for this. You need thirty minutes and a browser.
Use doubao.com, chat.deepseek.com, kimi.com and yuanbao.tencent.com. Turn web search on in each one — DeepSeek has it switched off by default, and if you forget, you are testing the training data instead of the live web. Use a fresh browser profile. Write down the date.
Check 1 — Does it know you?
What kind of brand is XX? (XX 是什么品牌？)
Look for: wrong country, wrong parent company, wrong founding year, wrong category. Watch what it does when it knows very little. In our test, thin information did not produce "I don't know". It produced confident detail.
Check 2 — Does it describe you correctly?
What are XX's products like, and who are they for? (XX 的产品有什么特点，适合什么人？)
Compare the answer with your own positioning line. Then check which pages it cited. If your own site is missing from a question about you, that is the first thing to fix.
Check 3 — Does it recommend you unprompted?
Which brands would you recommend in [category]? ([category] 有什么牌子推荐？)
This is the one that matters commercially, and the one most brands never run, because it does not contain their name. Ask it without your brand in the sentence. Then count where you appear, and who appears above you.
Run all three on all four assistants. Getting four different answers is normal, not an error.
What SEO, GEO and paid media each actually solve
These three terms are often used interchangeably, and should not be.
SEO, search engine optimisation, gets you into a list of links a person then chooses from. In China that still mostly means Baidu, and it still matters.
GEO, generative engine optimisation, is about whether you appear inside an answer that has already done the choosing for the buyer. There is no second page. Either you are in the paragraph or you are not. The work has two halves, and Finding 3 is why: make sure your own pages state plainly who you are, and make sure the sources those assistants actually read are saying the right thing about you.
Paid media buys attention around the answer, not inside it. This is where our client's assumption broke down. On the Chinese AI assistants, there is currently no advertising slot inside the generated response. Doubao sits inside ByteDance's ad ecosystem, Yuanbao inside Tencent's, so there is plenty you can buy around the answer. The paragraph itself is not for sale.
Who this is for
- Foreign brands entering China who have never checked what a Chinese AI assistant says about them, in Chinese.
- Brands with no official China channel yet, who assume the assistants stay silent about where to buy. They do not.
- Anyone being sold "GEO in China" who wants to know what is actually measurable and what is not.
Who this is not for
- Brands looking for a guaranteed ranking. There is no ranking to buy here, and anyone selling one is selling something else.
- Brands whose own pages still contradict each other. Fixing what you say about yourself comes before worrying about what a model says about you.
When you should move on from this
A visibility reading is a diagnostic, not a strategy. Once an assistant describes you correctly and your category pages are being cited, the constraint is no longer measurement — it is whether anyone outside your own domain has a reason to write about you. At that point the work moves to earning third-party coverage in the specific communities your category is read in, which in China is rarely the platform a Western playbook would name first.
Common questions
How do I check what Chinese AI assistants say about my brand?
Ask the same three questions in Chinese on Doubao, DeepSeek, Kimi and Yuanbao, with web search switched on in each one: what kind of brand you are, what your products are like and who they are for, and which brands the assistant recommends in your category with your name left out of the sentence. Use a fresh browser profile and write down the date. Expect four different answers — that is the normal result.
What is GEO in China, and how is it different from SEO?
SEO, search engine optimisation, gets you into a list of links a person then chooses from — in China that still mostly means Baidu. GEO, generative engine optimisation, is about whether you appear inside an answer the assistant has already composed. There is no second page: either you are in the paragraph or you are not.
Which AI assistants do Chinese consumers actually use?
Doubao (ByteDance) is far ahead at 382m monthly active users as of May 2026, followed by Qwen (Alibaba) at 167m and DeepSeek at 130m. Tencent's Yuanbao had 49m and Kimi 7m. Because each one draws on a different content ecosystem, checking only the largest gives you one of several different answers.
Is Baidu SEO still worth doing for foreign brands in China?
Yes. What has changed is that it is no longer the only place a buyer looks, and it is not where the assistant looks.
Why do different Chinese AI assistants recommend different brands?
Because they read different sources. In our test Tencent's Yuanbao cited WeChat Official Accounts more than any other named source, while ByteDance's Doubao leaned on Douyin properties and a specialist perfume community. Each assistant draws heavily on content inside its own corporate ecosystem.
Can I pay to appear inside a Chinese AI's answer?
Not inside the generated answer, as of October 2026. There is no advertising slot in the response itself on the major Chinese assistants. You can buy placement through e-commerce affiliate marketing if you are already the merchant in their ecosystem.
My brand is not sold in China yet. Does this still apply?
It matters even more. In our test, when a brand had no official China channel, the assistants said so and then directed buyers to daigou, cross-border retailers and duty free, with a counterfeit warning attached.
How often should brand visibility in Chinese AI assistants be re-tested?
These systems change their sources quickly. We treat a visibility reading as valid for about one quarter.
If you're working through this now
At Aquarc we run this as a diagnostic first: a fixed set of questions across the four assistants in Chinese, your brand and your three closest competitors, scored on the same three checks above.
You will get the raw answers, the citation sources, and a list of which pages are shaping what buyers are told about you.
China Digital Readiness & Market Fit — a structured read on your offer, your buyer, and how they actually discover and decide in China's digital environment.
If you want to see what the Chinese AI assistants currently say about your brand, request a China visibility review or contact us at bonnie@aquarc.com.