China's Appliance Giants Are Building AI Into Their Products — So Why Does That AI Keep Recommending Dyson?

Haier has connected its smart home ecosystem directly to DeepSeek. Midea has fused its in-house "Meiyan" model with DeepSeek, iFlytek Spark, and Doubao inside its flagship air purifiers. Hisense and TCL have announced their own DeepSeek integrations. On paper, China's domestic appliance industry has done everything a brand is told to do to win the AI era: get inside the model.

So we asked the models a simple, high-intent question — the kind a premium shopper actually types — across four of them: DeepSeek, Kimi (Moonshot), Doubao (via its ARK API), and Qwen (via DashScope). The prompt was identical each time: "I want to buy a high-end vacuum cleaner and a high-end air purifier, budget is not a constraint — which brands are worth recommending?"

This isn't a niche audience question. Doubao alone crossed roughly 240–345 million monthly active users by mid-2026 depending on the measurement window, DeepSeek sits in the 127–180 million range, and Qwen has climbed past 150–166 million — meaning tens of millions of Chinese consumers are routing exactly this kind of high-consideration purchase question through these four assistants every month, often instead of a traditional search engine. A brand that doesn't show up here doesn't just lose a citation; for a growing share of shoppers, it loses the moment the decision gets made.

None of the four models recommended Haier. None recommended Midea. Xiaomi came up nowhere. Instead, all four converged on the same small cluster of foreign names: Dyson for vacuums in every single response, and IQAir and Blueair for air purifiers in three out of four. Miele appeared in three of four as a second-tier vacuum pick. Qwen and Doubao each surfaced one domestic name — Qwen's Doubao-generated answer mentioned 莱克 (Leizhi) and 352, both legitimate Chinese brands, but only as third-tier additions behind the Western names, not as the leading recommendation.

The embedding paradox

This is a small but pointed illustration of a gap that matters a lot more broadly than vacuum cleaners: putting a model inside your product is not the same thing as being the brand that model recommends.

Midea, Haier, Hisense, and TCL's AI integrations are a hardware and software feature — DeepSeek or Doubao runs inside an air conditioner or purifier to power voice control, adaptive settings, or predictive maintenance. That is a genuinely useful product capability, and it generates real press coverage in Chinese trade media. But it does nothing to change what the underlying model says when a consumer, in a completely separate conversation, asks it for a purchase recommendation. Embedding a model in your product doesn't feed the model new opinions about your product — those opinions are still shaped by whatever the model was trained and fine-tuned on, which for a category like "premium home appliances" appears to be dominated by long-standing international review content, spec-comparison sites, and English- and Chinese-language buying guides that have covered Dyson's cyclone technology or IQAir's HyperHEPA filtration for over a decade.

What the four models actually said

The consistency is the finding. Across four different labs, four different training pipelines, and four different fine-tuning approaches, the "default premium answer" for this category is the same handful of Western and European names. That's a strong signal that these citations are anchored in a shared pool of source material — the kind of long-tail technical and comparative content (CADR ratings, HEPA grades, motor wattage, filtration test results) that dominates search results and gets scraped into pretraining corpora and RAG indexes alike.

It's also worth flagging a smaller but telling detail: Doubao's answer described 斐纳 (a brand marketed in China as German) using country-of-origin language pulled straight from that brand's own marketing, without any indication the model had verified the claim. AI answers reproduce brand positioning uncritically — good news if you already have a clean, well-documented technical story, and a real risk if a competitor's unverified origin story is quietly winning the recommendation instead of yours.

Why RAG doesn't close the gap either

It would be reasonable to expect DeepSeek — which leans on retrieval-augmented generation and tends to favor sourced, citation-style content — to behave differently from a more purely pretrained model like Doubao. It didn't. Both landed on Dyson and IQAir/Blueair as the anchor answer. That suggests the gap isn't just a training-data artifact that a retrieval layer could patch; it's that the retrievable corpus itself — the reviews, lab tests, and comparison articles that a RAG system pulls from at inference time — is still dominated by the same international names. Midea and Haier's model integrations changed what happens inside their own appliances. They haven't yet changed what gets written about their appliances in the review ecosystem those models query.

Not every category looks like this

It's worth being precise about where this pattern does and doesn't hold, because it isn't a blanket "foreign brands always win" story. hubGEO's tracking of consumer electronics has found the opposite dynamic in smartphones and telecom hardware, where Huawei and Xiaomi hold a clear domestic-market citation advantage over Apple and Samsung — categories where local ecosystem lock-in, carrier relationships, and sheer volume of Chinese-language product discussion outweigh Western brand heritage. Home appliances sit at the other end of that spectrum: a category where the technical review culture — CADR values, HEPA class certifications, decibel ratings — was built up over a decade-plus around European and Japanese engineering brands, in both English- and Chinese-language sources, before Chinese domestic manufacturers caught up on comparable lab-verified content. The lesson isn't "foreign always wins" or "domestic always wins" — it's that AI citation share tracks whichever ecosystem produced the deeper, more consistent technical documentation trail for that specific category, regardless of which company happens to have a model-integration press release.

What this means beyond appliances

The home appliance case generalizes to any category where a domestic manufacturer's AI strategy is "put the model in the box" rather than "get cited by the model." hubGEO's ongoing brand-score tracking across categories has repeatedly found the same pattern: strong retail presence or domestic AI partnerships do not reliably translate into strong AI citation presence, because those are two different systems. One is a product roadmap; the other is a content and reference-data problem.

For international brands, that's the useful part of this story. The fact that Haier, Midea, and Xiaomi have invested heavily in DeepSeek and Doubao integration hasn't bought them the recommendation slot in the very models they've integrated with. The AI recommendation layer is still, for now, running on the same kind of technical authority and citation depth that has always won category leadership — durable spec comparisons, third-party lab results, long-form reviews, and consistent claims that don't change brand to brand.

Takeaway for Brand Marketers

If you sell a premium physical product into China and you're worried that domestic competitors' "AI-inside" partnerships are going to out-cite you, this data says: not yet, and not automatically. What wins the recommendation slot is still the boring, durable work — published technical benchmarks (filtration grade, CADR/CCM values, motor specs), third-party comparison content that models can retrieve and trust, and consistent factual claims about your product across every platform where you have a presence. Don't confuse a competitor's hardware-embedded model partnership with a genuine GEO threat — check what the models actually say when asked to recommend, not just who they've partnered with. And if you already hold a strong citation position like Dyson or IQAir do here, the priority is defending it: keep your technical documentation current, since that's the layer these models are actually drawing from.

Related: see how audio and electronics brands and consumer electronics fare on the same six-model tracking, or browse live scores on the brands page.