The SMZDM Gatekeeper: Why One Platform Controls 80–99% of Luxury Brand Citations on DeepSeek and Qwen (2026)

When a Chinese consumer asks DeepSeek or Qwen which luxury watch, suitcase, or handbag to buy, both engines reach for the same handful of websites almost every time. According to a Q2 2026 research briefing from Eastbound Research, which tested 1,620 brand-recommendation responses across watches, luggage, and handbags, one platform outranks every other cited source: SMZDM (什么值得买, "What's Worth Buying"), a Chinese deal-and-review aggregator. Its mention rate ranged from 80% to 99% across both engines and all three categories, placing it in the top three cited sources in all 12 test sub-cells run. For brands measuring AI visibility in China by brand score alone, that finding points to a layer underneath the score that matters just as much: which websites the model says it relied on to reach that answer.

What "source attribution" tells you that a brand score doesn't

A brand score answers one question: does the model mention you? Source attribution answers a different one: what does the model say it read before answering? DeepSeek and Qwen both self-report source websites alongside their brand recommendations, and Eastbound's data shows those self-reported sources cluster tightly around a small set of Chinese platforms. Beyond SMZDM, three more sources appeared in every one of the 12 test cells: Zhihu (知乎, mention rate 32%–98%), Xiaohongshu (小红书, 17%–96%), and YouTube (22%–73%). Reddit appeared in 11 of 12. Two more, Bilibili and Wikipedia, surfaced strongly on DeepSeek but inconsistently on Qwen.

For a brand building a China GEO content plan, this is the practical instruction the score alone can't give: if SMZDM, Zhihu, and Xiaohongshu don't carry your brand, you are largely absent from the layer these two models draw on before they recommend a competitor instead.

The pattern breaks at the ultra-luxury tier

The 80–99% SMZDM figure is a category average, and it does not hold at every price point. When Eastbound re-cut the handbag panel by price tier, comparing entry/aspirational purchases (≤¥10,000) against ultra-luxury (¥30,000+, the Birkin/Kelly/exotic-Chanel zone), SMZDM's role changed sharply.

One DeepSeek answer to a prompt about ¥100,000+ handbags reportedly 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." That is the model describing, in its own words, a different source hierarchy for a different buyer. A brand selling into the Birkin tier that pours its content budget into SMZDM is optimizing for the wrong layer.

DeepSeek concentrates its sources; Qwen spreads across a much longer tail

The two engines don't just cite different weights on the same sources — they cite structurally different numbers of sources. DeepSeek pulls from roughly 200–400 unique sources per category, with the top 15 capturing 60%–80% of all citations. Qwen pulls from 480–940 unique sources, and its top 15 captures only 40%–65%. That difference changes what "good enough" coverage looks like on each engine: on DeepSeek, breaking into the top 15 sources captures most of the citation traffic; on Qwen, a brand needs both broader coverage and stronger top-tier positioning, because the long tail carries more weight.

The two engines also diverge more on sources than on the brands they recommend. Across watches, luggage, and handbags, DeepSeek and Qwen agreed on roughly 70% of the brands they surfaced — but only 27% of the sources they cited behind those recommendations, a gap that widens as the category becomes more lifestyle-driven (handbags hit a 3.4x divergence ratio between brand overlap and source overlap). A brand audit run only on DeepSeek says very little about a brand's Qwen source-side visibility, even when the two engines are recommending the same brands.

Where brand-owned websites disappear — and where they matter

In the watches category, a single Chinese vertical specialist, Watch Home (腕表之家), appeared in close to 100% of responses on both engines, with Hodinkee behind it. When a specialist like that dominates a category, brand-owned websites tend not to make the top 15 at all — Rolex.com and other official watch brand sites didn't surface in Eastbound's data for either engine. In categories without an equivalent specialist, brand-owned sites carry more weight: Rimowa's official site reached 29% mention on DeepSeek-Luggage, and Tumi's reached 19%. The practical read for content teams: check whether your category has a dominant Chinese vertical specialist before betting budget on your own zh-CN site's source-side visibility. Where one exists, editorial placement in that specialist's ecosystem outranks brand-site optimization; where none exists, the brand's own site is a higher-priority lever.

The lifestyle gradient: Xiaohongshu rises, Wikipedia fades

Source weight also shifts with how lifestyle-driven a category is. Comparing technical (watches), utility (luggage), and lifestyle (handbags) categories, Xiaohongshu's mention rate climbs steadily while Wikipedia's falls.

For craft and heritage categories — watches, fine spirits, high-end luggage — Wikipedia entity coverage still carries measurable weight in both engines. For fashion-lifestyle categories like handbags, Xiaohongshu KOL relationships are the heavier lever, and Wikipedia's weight nearly disappears. A single content strategy applied across a brand's full product range risks over-investing in KOL seeding for a technical sub-category, or under-investing in entity documentation for a lifestyle one.

The substrate is Chinese; the subject is Western

Between 46% and 63% of all source citations across the three categories came from Chinese-language platforms — yet the brands DeepSeek and Qwen actually recommended were overwhelmingly Western. Every brand in the top 10 for watches was Swiss, German, or Japanese, with zero Chinese brands appearing in either engine's list. Handbags showed the same pattern: Louis Vuitton, Chanel, Hermès, Dior, Gucci, and Prada dominated, again with no Chinese brand in the top 10. Luggage was the lone exception, with Xiaomi appearing at #7 on DeepSeek alongside Rimowa, Tumi, Samsonite, Bric's, and Globe-Trotter.

The implication for Western brands is direct: these engines are answering in a Chinese-language source layer about products the training corpus already treats as globally Western. A brand publishing only in English, or only on its global .com domain, is effectively invisible to the layer doing the actual citation work. Publishing in Chinese, on Chinese platforms — not simply translating an English press release — is the baseline test, not an optional localization step.

Takeaway for Brand Marketers

Three actions follow directly from this data. First, treat SMZDM, Zhihu, and Xiaohongshu as the non-negotiable baseline for any Mainland-China AI visibility program — Eastbound's data holds them constant across every category and price tier tested for aspirational-to-mid-luxury goods. Second, if your brand sells into an ultra-luxury tier, redirect editorial effort away from SMZDM toward The Purse Forum-style collector communities, Vogue Business/WWD earned coverage, and auction-house provenance archives — the source mix the models themselves say they reach for at that price point. Third, run source-attribution checks per engine, not just per brand score: a strong DeepSeek presence doesn't predict a strong Qwen one, because the two models draw on largely non-overlapping source pools even when they agree on which brands to recommend. Pair any of these moves with a before/after visibility check at the 12-week mark — source-side shifts take longer to register than brand-mention shifts.

Related: see current brand visibility scores across China's six major AI models on the hubGEO brand tracker.

Source: Eastbound Research, "Luxury Brand AI Visibility in China: The 5 Sources DeepSeek and Qwen Cite" (Q2 2026), based on 1,620 brand-recommendation responses across watches, luggage, and handbags, each prompt run six times per engine.