Luxury brand AI visibility in China: the 5 sources DeepSeek and Qwen cite (and what changes for ultra-luxury).

Mainland-Chinese consumers increasingly ask DeepSeek, Qwen, Doubao (豆包), Baidu ERNIE (文心) or Tencent Yuanbao (元宝) directly: "Which luxury watch should I buy under ¥50,000?" The AI answers with a list. Some brands appear; others don't. Brands that surface enter the consumer's consideration set; brands that don't are harder to discover at this layer. The question every luxury brand should be asking: which websites does the AI reach for when it makes that recommendation?

SMZDM dominates Chinese AI luxury recommendations — at the aspirational tier

SMZDM (什么值得买, "What's Worth Buying") was a top-3 cited source in every single test we ran for aspirational-luxury queries. Mention rate: 80%–99% across watches, luggage, and handbags on both DeepSeek and Qwen.

SMZDM is a Mainland-Chinese consumer review and deal-aggregation platform. It was named as a top-3 source in:

• DeepSeek answers about luxury watches (89%)
• DeepSeek answers about luxury luggage (96%)
• DeepSeek answers about luxury handbags (85%)
• Qwen answers about luxury watches (99%)
• Qwen answers about luxury luggage (93%)
• Qwen answers about luxury handbags (80%)

When each panel is split into two independent runs (for the reliability check), SMZDM held top-3 in every single one: 12 of 12. Zhihu (知乎), Xiaohongshu (小红书), and YouTube were also universal but typically ranked lower. For aspirational-luxury positioning, SMZDM is the highest-priority editorial channel to test first. For ultra-luxury, the source map changes — see the next section.

Ultra-luxury tier (Hermès / Chanel exotic / Birkin) breaks the SMZDM pattern

For 30,000+ RMB (3万+) handbags, SMZDM's mention rate collapses from 100% to 33% on DeepSeek and from 100% to 71% on Qwen. The engines themselves shift weight to The Purse Forum, Vogue Business / WWD, and auction-house archives.

The 80–99% SMZDM mention rate above is a category-level average. When we re-cut the same handbag panel by price tier — entry/aspirational (≤10,000 RMB / 1万) versus ultra-luxury (30,000+ RMB / 3万+, the Birkin/Kelly/exotic-Chanel zone) — SMZDM's role changes substantially. This matters for any ultra-luxury brand reading this, because over-association with a deal-and-review aggregator is itself a brand-safety question for that tier:

What the engines themselves say in their source-attribution turn is striking. A representative DeepSeek answer to "How to pick 100,000+ RMB luxury bags? (10万+奢侈包怎么挑？)" downgrades 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" and Sotheby's / Christie's auction archives to "medium weight, for rare-piece valuation reference". The engines themselves are signalling that the source mix at the Birkin tier is structurally different from the source mix at the entry tier.

Brand-safety implication for ultra-luxury positioning. SMZDM is, in its own positioning, "what's worth buying" — a deal-and-comparison aggregator. For an aspirational-luxury brand that already lives in a "considered purchase / good buy" frame, that's an asset. For Hermès, Chanel exotic, Bottega Cabat, Goyard custom — categories whose value proposition rests on scarcity, waitlists, and craftsmanship rather than comparability — heavy SMZDM presence sits awkwardly against the desired positioning. The data suggests this isn't just a positioning preference; it is also where the engines themselves shift weight away from SMZDM.

Should ultra-luxury brands like Hermès optimise for SMZDM? No — deprioritise SMZDM as the lead channel at this tier. The highest-weight intervention hypotheses are (a) presence in The Purse Forum long-form authentication / collector threads, (b) earned coverage in Vogue Business and WWD, and (c) being the subject of auction-house lot descriptions and provenance archives (Sotheby's, Christie's, Phillips, plus Mainland equivalents Yongle (永乐) and China Guardian (中国嘉德)). Xiaohongshu remains constant across tiers and stays in the plan.

The Core 5 sources Chinese AI cites for luxury — across watches, luggage, handbags

Beyond SMZDM, three more sources were cited in every test cell. A fifth — Reddit — was cited in 11 of 12 (missing only from Luggage × Qwen × Run-2). Together, the Core 5:

Two more sources — Bilibili (哔哩哔哩) and Wikipedia — surfaced strongly on DeepSeek but inconsistently on Qwen. They sit one tier down: a high-priority addition for any DeepSeek-targeted programme, conditional for Qwen-targeted.

Caveat on YouTube and Reddit: both are blocked from typical Mainland-China consumer access without a VPN. Their appearance here likely reflects each engine's training-corpus / global-web priors rather than typical Mainland consumer behaviour. Treat them as English-language content layers the engines reach for when the answer is global in scope (Western luxury brands), not as direct Mainland consumer media plans.

How the luxury Core 5 differs from the pan-niche Universal Five

Readers who've seen our companion piece on three Chinese AIs will notice a different five-source list there: Zhihu, SMZDM, Xiaohongshu, Bilibili, Dianping. That list is correct for that study — 10 generalist Mainland-CN niches across three engines including Doubao. The luxury Core 5 above is correct for luxury watches / luggage / handbags across DS + Qwen. They differ for two reasons that matter for a luxury media plan:

1. Luxury is a global-corpus topic. The brands recommended (LV, Hermès, Rolex, Rimowa) sit in Western training corpora, so YouTube and Reddit surface meaningfully. In generalist Mainland niches, the brands sit in Mainland corpora and Bilibili / Dianping take those slots instead.
2. This luxury panel did not include Doubao. Doubao is the largest Mainland consumer-AI surface and pulls the source mix Mainland-ward.

DeepSeek concentrates, Qwen distributes — two engines, two media-plan shapes

DeepSeek and Qwen don't behave the same way when surfacing sources.

DeepSeek concentrates on a smaller, deeper set of self-attributed Chinese community sources. Qwen distributes across a much wider tail that includes more Western publications and brand-owned content. A DeepSeek-targeted GEO programme is fundamentally different from a Qwen-targeted one. On DeepSeek, hitting the top-15 captures most citation traffic. On Qwen, you need broader coverage and stronger top-N positioning, because the long tail is much larger.

Lifestyle gradient — Xiaohongshu rises with category, Wikipedia weakens

The more "lifestyle-driven" the category, the heavier the weight of Xiaohongshu and the lighter the weight of Wikipedia.

For craft and heritage categories (watches, fine spirits, luggage), Wikipedia entity coverage carries weight in this study's data. For fashion-lifestyle categories (handbags), Xiaohongshu KOL relationships are the heavier lever. Don't pick the wrong tier for your category.

When vertical specialists like Watch Home (腕表之家) dominate, brand-owned luxury sites disappear

In the watches category, a single Mainland-Chinese specialist publication — Watch Home (腕表之家) — was cited in nearly 100% of responses on both engines. Beside it, the Western magazine Hodinkee appears in roughly half.

When a specialist source like this exists, brand-owned websites don't enter the top-15. Rolex.com, Omega's site, and other brand homes did not appear in our top-15 luxury-watch source list for either engine.

In luggage and handbags, where no equivalent specialist dominates, brand-owned sites surfaced meaningfully:

• Rimowa's official site: 29% on DeepSeek-Luggage
• Tumi's official site: 19%
• Brand China sites: 9% on Qwen-Handbags

Also a category surprise in handbags: resale and authentication platforms entered the top-10 (The RealReal, Vestiaire Collective, The Purse Forum, and Mainland resale platform Plum 红布林). None appeared for watches or luggage.

In specialist-dominated categories, don't expect your own zh-CN site to win source attribution against vertical media. Keep your zh-CN site technically and semantically strong as a baseline; put incremental editorial budget into the specialist ecosystem first. In categories without a dominant specialist (luggage, handbags), brand-owned sites are a higher-priority lever.

The engines agree on luxury brands but disagree on sources — by ~3×

The two engines overlap on roughly 70% of recommended luxury brands. When we ran the same comparison for sources, agreement dropped to about 27%.

In plain English: the two engines tell consumers similar things, but they cite very different sources behind those recommendations. The divergence widens as the category becomes more lifestyle-driven. A brand audit on DeepSeek tells you very little about your Qwen visibility on the source side. Per-engine source-side measurement is essential. Brand-side measurement is partly substitutable across engines; source-side is not.

Chinese sources, Western brands — the substrate is local, the subject is global

The Chinese AI engines we tested cited Chinese-language sources heavily, with 46% to 63% of all source mentions coming from Mainland-Chinese platforms. But the brands they recommended were overwhelmingly Western:

• Watches: every brand in our top-10 is Swiss, German, or Japanese. Zero Chinese brands in either engine's top-10.
• Luggage: 1 Chinese brand (Xiaomi, 小米) at #7 on DeepSeek; otherwise Western (Rimowa, Tumi, Samsonite, Bric's, Globe-Trotter).
• Handbags: zero Chinese brands in either engine's top-10. Louis Vuitton, Chanel, Hermès, Dior, Gucci, Prada dominate.

In short: these engines output Western brand recommendations alongside Chinese-language sources. The output substrate is local; the subject is global. A brand that publishes only in English or only on its global website is invisible to whatever Chinese-language source layer is doing the work. Showing up in Chinese, on Chinese platforms, is a high-priority test for any Western luxury brand entering this layer.

What this means for your luxury brand AI visibility strategy

Each row below is a highest-priority intervention hypothesis to test, not a guaranteed-causal recommendation. Pair any of them with a before/after measurement at T+12 weeks.

Methodology

• Sample. 1,620 brand-recommendation responses across luxury watches, luggage and handbags. 45 natural Mainland-Chinese consumer prompts per category, each run 6 times, on each engine. Total 6 test cells; 12 sub-cells when split into two independent runs for reliability testing.
• Engines. DeepSeek (deepseek-chat); Qwen (qwen-plus, DashScope international). Pure API calls (no live web retrieval, no chat-with-Search toggle). Doubao deferred to a follow-up study.
• What we measured. The websites the AI says it relied on after answering. A stable, replicable descriptive signal, not direct proof of internal retrieval. Source names are normalised to platform level (e.g., "Xiaohongshu" rather than a specific post).
• What we did not measure. Sales, conversion, attributable revenue. Findings are for Mainland-Chinese consumer audiences only; chat-with-Search-ON browsing surfaces are a separate study.
• Reliability. Source-side findings replicated across two independent runs of the same panel. Mean Spearman ρ = 0.97 for sources, 0.90 for brands; mean Jaccard top-15 = 0.74 for sources, 0.87 for brands. 11 of 12 cells passed the conventional "strong" threshold.
• Tier-cut caveat. The handbag panel was the only category with enough native price-tier prompts to support the SMZDM-tier-collapse cut. Watches and luggage have prompt-level price signals too but a smaller share of "ultra"-tier prompts in v1; replicating this tier-cut for fine watches (¥200K+) and bespoke luggage is the most useful follow-on to confirm the pattern is category-general.

Full prompt panels, normalisation rules and reproducible scripts available on request.