The finding: language is the switch, not the engine
Asked in English, Chinese engines named Chinese products in 1.3% of recommendations — identical to the Western engines. Asked in Chinese, every engine moved, together.
Two engines are worth singling out. Claude and Kimi named zero Chinese brands when asked in English. Kimi is a Chinese model; asked in English its recommendation list contained no Chinese product at all.
That is not a preference. It is what happens when the corpus an engine retrieves from is layered by language. A brand that exists only in English literature does not exist on the Chinese side of the index, on any engine.
| Engine | Chinese-brand share — asked in English | Asked in Chinese |
|---|---|---|
| 豆包 Doubao | 2.0% | 40.5% |
| Perplexity | 2.5% | 31.7% |
| Claude | 0% | 31.3% |
| DeepSeek | 2.0% | 31.3% |
| Kimi | 0% | 26.7% |
| 通义千问 Qwen | 1.0% | 23.8% |
| Grok | 0% | 22.9% |
| ChatGPT | 2.2% | 18.0% |
The Split Answer Report, July 2026. 48 calls per engine, single sample.
Chinese engines do not show their sources — what that changes
The Western engines returned citation links. The Chinese engines returned none: they give an answer and do not say where it came from.
So the Western playbook — read the citations, go earn a place among them — does not transfer unchanged. On the Chinese side you are working without the feedback signal, which makes the measured mention rate the only real instrument you have.
It also means a Chinese engine's answer cannot be audited the way a Perplexity answer can. We report it as not-applicable rather than as zero, because absent data is not a zero.
Which sources these engines actually draw on
Only the Western engines expose citations, so this is their list across the whole experiment — both languages, not a Chinese-only slice, because that slice is not something the data supports.
Even so it is pointed. In an experiment with twelve English questions and twelve Chinese ones, two Chinese technical communities tied for first place above every vendor site.
| Most-cited domain | Times cited |
|---|---|
| cnblogs.com 博客园 | 27 |
| blog.csdn.net | 27 |
| learn.g2.com | 21 |
| zhuanlan.zhihu.com 知乎专栏 | 16 |
| cloud.tencent.com 腾讯云社区 | 14 |
| reddit.com | 14 |
What these four Chinese sources have in common: anyone can register and publish on them.
What is closed to an overseas company
Worth knowing before it costs you weeks. Baidu's encyclopedia and its commercial categories require a mainland Chinese business licence — a company entry needs the registered mainland corporate name, a brand entry needs a CNIPA trademark. A Canadian or US company satisfies neither, and there is no foreign-language creation path.
Chinese map platforms are the same wall: a POI listing needs a mainland business entity with a physical address.
What is open, without a licence: technical communities and content platforms that accept personal registration. That is where the citations in the table above actually come from, which is convenient — the accessible platforms and the cited platforms are the same list.
Common questions
Do I need a Chinese website, or is a translated page enough?
You need Chinese content that a Chinese reader would recognise as written rather than translated, on a URL an engine can fetch. Machine-translated English structure reads as foreign to both the reader and the corpus around them. A separate indexable Chinese URL with reciprocal hreflang beats a language toggle that only appears after a click — an engine cannot click.
Can I skip the Chinese engines if I only sell in North America?
Depends who is asking, not where you are. If a meaningful share of your buyers ask in Chinese — which in Metro Vancouver is common — then the finding above applies to you directly: they are getting an answer, and it is not naming you.
Are the Chinese engines less reliable to work with?
Not in this run. The four Chinese engines returned a valid answer to all 48 of their calls. The failures were all Western: Grok failed 33.3% of its calls, ChatGPT and Claude 12.5% each. That is one measurement of a few days in July 2026, not a standing property.
Scope and limits
Measured July 2026, eight engines, twelve B2B software questions, single sample, official APIs rather than consumer apps — the apps typically add their own search and recommendation layers on top. Dataset under CC BY 4.0 at /research/split-answer.
Last updated 2026-09-26.
Is your brand in the answer?
Avowd measures how eleven AI engines describe and recommend a brand, in English and Chinese, then does the work to change it. The free check takes about thirty seconds.