Who China's biggest AI assistant actually names: beauty, luxury, cars

A 210-question API snapshot across three industries, July 2026. One engine (Doubao), frozen questions spanning seven buyer-intent layers, one run each — a wide-space snapshot, not market share, not consumer-app behavior, and not a time-stability estimate. Product, price and safety assertions in cited sources remain unverified unless separately reviewed. Raw per-question tables and method notes: geochecktool.com/research.

The one-sentence finding: in each industry, a different kind of entity owns the answer box — and in none of them is it simply "the biggest brand."

Beauty: specific SKUs — and the regulator's app

70 frozen questions; search triggered in 69/70; 142 engine-generated search queries; 474 cited source positions; 413 distinct exact-name candidate strings (not entity-resolved brands). At least one identifiable candidate in 53/70 answers; 6/70 answers explicitly excluded a named entity.

The most-adopted names are individual products, not houses: 薇诺娜舒敏保湿特护霜 (Winona, 5/70) and 芙丽芳丝净润洗面霜 (freeplus, 5/70) lead; Curél, Avène, Eucerin, CeraVe SKUs follow. And one non-brand keeps appearing: 化妆品监管APP — the national cosmetics-regulator app — was adopted in 4/70 answers, typically as the way to verify product registration. In beauty, the engine's answer shelf is SKU-level and audit-minded: it names the exact cream, then tells the user how to check the license.

Luxury: resale and authentication platforms outrank the maisons

70 frozen questions; search in 68/70; 123 queries; 414 cited positions; 577 distinct exact-name candidate strings (not entity-resolved brands) — the widest shelf of the three by that measure. Candidates in 62/70 answers; 14/70 answers (20.0%) explicitly excluded a named entity — the highest veto rate of the three industries.

The most-adopted name in luxury questions is not a maison. It is 红布林 (Plum, a resale platform, 9/70), followed by 中检 (China Certification & Inspection Group, 7/70) and 优奢易拍 (an authentication app, 6/70). Hermès (6/70), Chanel (6/70) and LV (5/70) sit level with or below the verification layer; Farfetch, Prada, 转转 and 胖虎, each at 4/70, round out the top ten. Luxury buyers ask fear-shaped questions — real vs fake, where to authenticate, whether to buy second-hand — and the answer box belongs to whoever operates the trust infrastructure, not to the brand being bought.

New cars: one content platform leads a tied second tier

70 frozen questions; search in 69/70; 121 queries; 377 cited positions; 275 distinct exact-name candidate strings (not entity-resolved brands). Candidates in 44/70; explicit exclusions in 4/70 — the calmest veto profile of the three.

The top adopted name is 汽车之家 (Autohome, 5/70). At 4/70, 懂车帝 (Dongchedi) ties the BYD Seagull model and the brand-level names Li Auto and Tesla. A long tail of specific trims follows (BYD Qin L DM-i, Toyota Frontlander, XPeng MONA M03, Li L6, AITO M6, 3/70 each). In this snapshot, one comparison platform leads, but the next tier mixes another platform, a model and two brand-level names.

What this means if you run a brand in China

Three industries, three different answer-box structures:

Industry	Who owns the answer box	Veto intensity (explicit exclusions)
Beauty	Specific SKUs + the regulator's own app	6/70
Luxury	Resale & authentication platforms	14/70
New cars	Car-content platforms, then models	4/70

If your brand's China AI presence is managed the way you manage search — at the brand level, on the web — you are playing the wrong unit on the wrong shelf. The engine answers with SKUs, verification layers and comparison platforms. Whether your SKUs, your verification trail and your platform presence survive those layers is a measurable question — measured on the real consumer surface, question by question, with screenshots.

Method: frozen 70-question banks per industry built from real public buyer questions (≥50% real-question main track), seven intent layers, one API run per question, July 2026; engine-displayed search queries and cited source positions logged verbatim; per-answer candidate extraction Tier-2 coded (ten-sample pre-flight agreement: candidate micro-F1 0.9714, explicit-exclusion micro-F1 0.8000, comparable-order exact rate 1.0000). API and consumer-app are different surfaces; consumer-app anchor evidence is collected separately and never pooled. One engine, one week. We sell measurement; no named entity paid for inclusion; measurement itself cannot be purchased.

Wang Bo — GeoCheckTool. The China AI measurement firm.