The 5 Websites Chinese AI Reads Before Recommending a Luxury Brand

When a Chinese consumer asks DeepSeek or Qwen "which luxury watch should I buy under ¥50,000," the model doesn't invent an answer from brand awareness alone. It reaches for a small set of third-party sources, weighs what they say, and only then produces a recommendation. A new source-attribution study from Eastbound Research, covering 1,620 brand-recommendation responses across watches, luggage, and handbags, mapped exactly which sites those two engines cite — and the pattern is far more concentrated than most brand teams assume.

The headline number: one platform, SMZDM (什么值得买, "What's Worth Buying"), appeared as a top-3 cited source in 12 of 12 test cells for aspirational-luxury queries, with mention rates of 80%–99% across categories on both DeepSeek and Qwen. If a brand has no meaningful presence on SMZDM, it is working from a structural disadvantage before a single consumer ever types a prompt.

The Core Five: what shows up in every luxury query

Beyond SMZDM, four more sources appeared consistently across the watches, luggage, and handbags panels:

Two sources — Bilibili and Wikipedia — surfaced strongly on DeepSeek but inconsistently on Qwen, making them a conditional add depending on which engine a brand is optimizing for.

This is a meaningfully different picture than the "top of Baidu SEO" world most international brand teams still plan around. None of the Core Five is a brand's own .com domain, and only two are the household names (Zhihu, Xiaohongshu) that most China marketing playbooks already mention. SMZDM in particular is rarely a line item in a Western luxury brand's China media plan, yet it is the single most consistent citation across both engines tested.

Ultra-luxury breaks the pattern entirely

The more striking finding sits inside the handbag data. When the same panel is split by price tier — entry/aspirational (≤¥10,000) versus ultra-luxury (¥30,000+, the Birkin/Kelly/exotic-Chanel zone) — SMZDM's dominance collapses:

One DeepSeek answer to a ¥100,000+ handbag query explicitly downgraded SMZDM to "low weight, used only for second-hand market discussion" while elevating The Purse Forum to "high weight, used for rare-skin authentication discussion among global collectors." The model is not just retrieving different pages at the ultra-luxury tier — it appears to be applying a different trust hierarchy altogether, shifting from a deal-comparison frame to a scarcity-and-provenance frame.

For a brand like Hermès or Chanel's exotic-skin line, heavy SMZDM association would actually work against the intended positioning: SMZDM's own brand promise is "what's worth buying," a comparability frame that sits awkwardly next to waitlist-and-craftsmanship messaging. The data suggests the engines have already made that distinction on the brand's behalf.

DeepSeek concentrates, Qwen distributes

The two engines don't just cite different volumes of sources — they behave structurally differently. DeepSeek cites 200–400 unique sources per category, with its top 15 sources capturing 60%–80% of all citations. Qwen cites a much wider pool — 480–940 unique sources — with its top 15 capturing only 40%–65%. Practically, this means a DeepSeek-focused GEO program can concentrate on a shorter source list and expect meaningful coverage, while a Qwen-focused program needs both broader source coverage and stronger top-tier positioning within that longer tail, including more Western publications and brand-owned content, which surface more often on Qwen (Rimowa's own site appeared in 14% of luggage citations, Tumi's in 6%).

The two engines also agree far more on which brands to recommend than on which sources back those recommendations: roughly 70% brand overlap across categories, but only about 27% source overlap. That gap widens as categories get more lifestyle-driven — 36% source overlap for watches versus 20% for handbags. The implication for measurement is direct: auditing a brand's visibility on one Chinese engine tells you very little about its source-level exposure on another. Brand-side and source-side visibility need to be tracked separately, and per-engine.

Where vertical specialists crowd out everything else — including the brand's own site

In the watches category, a single specialist publication, Watch Home (腕表之家), was cited in nearly 100% of responses on both engines, with Western outlet Hodinkee appearing in roughly half. When a specialist this dominant exists, brand-owned websites disappear from the citation set entirely — Rolex.com and other brand homepages did not surface in either engine's top 15 for watches.

Luggage and handbags, which lack an equivalent single dominant specialist, tell a different story: Rimowa's own site reached 29% of DeepSeek luggage citations, and resale/authentication platforms (The RealReal, Vestiaire Collective, The Purse Forum, Mainland platform Plum/红布林) entered the top 10 for handbags — a category signal that didn't appear at all for watches or luggage.

Chinese substrate, Western subject

Across all three categories, 46%–63% of cited sources were Chinese-language platforms — yet the brands those sources led the AI to recommend were overwhelmingly Western. Zero Chinese brands appeared in either engine's top 10 for watches or handbags; only Xiaomi cracked the luggage top 10, at #7 on DeepSeek. In other words, the engines are doing local-language reasoning to arrive at global-brand answers. A Western luxury brand that only publishes in English, or only invests in its global .com, is largely invisible to the Chinese-language source layer doing the actual recommendation work — regardless of how strong that brand's Western digital footprint is.

Takeaway for Brand Marketers

For most international luxury brands entering or maintaining presence in China's AI-search layer, the highest-leverage first move is not a brand-owned content push — it's an editorial relationship with SMZDM, paired with sustained activity on Zhihu and Xiaohongshu, which together form what the source data calls the "Mainland-activation core." Brands selling at aspirational price points should treat SMZDM presence as close to non-negotiable given its 80–99% citation rate.

Ultra-luxury brands should do close to the opposite: deprioritize deal-aggregator channels and instead invest in long-form authentication and collector-community presence (The Purse Forum), earned trade coverage (Vogue Business, WWD), and provenance visibility through auction-house archives — channels that barely register at the entry tier but carry the citation weight once price crosses into Birkin territory.

Two structural points should shape any measurement plan. First, DeepSeek and Qwen require separate source-side audits — brand-level visibility doesn't transfer between them, and source-level visibility transfers even less. Second, in categories with a dominant vertical specialist (watches have Watch Home; other categories may have their own), don't expect a brand's own zh-CN site to compete for citation share against that specialist — treat the brand site as a semantic baseline and put incremental budget into the specialist ecosystem itself.

None of this replaces English-language global SEO or Western PR. It sits alongside it, as a distinct, Chinese-language-first citation layer that determines whether a brand enters a Chinese consumer's AI-generated consideration set at all — which, given how directly consumers are now asking DeepSeek and Qwen for purchase recommendations, is increasingly the layer that matters most before any storefront or search ad ever gets a look.

Why this matters beyond luxury

The methodology behind this finding is worth noting for any brand team running its own GEO measurement, not just luxury players. The study ran 45 natural Chinese-language prompts per category, six times each, split into two independent runs for reliability — and the source-side results replicated with a mean Spearman correlation of 0.97 across runs. That's a much tighter reliability bar than most single-prompt "ask ChatGPT about my brand" checks that pass for AI-visibility audits today. A one-off prompt tells a brand almost nothing; the source hierarchy only becomes visible once a query is repeated across dozens of natural variations and cross-checked between engines.

It's also a reminder that "China AI visibility" is not one number. The same brand can score well on DeepSeek's brand-recommendation layer while being functionally invisible in DeepSeek's source-attribution layer if its own content lives only on channels the model doesn't trust for that category and price tier. Brand tracking and source tracking are two different instruments, and both are necessary to know whether a marketing investment is actually reaching the citation layer these models rely on.

Related: see how individual brands score across China's six major AI models on the GEO Hub brand rankings.