The empty shelf: maternity, supplements, pet food on China's biggest AI assistant

Second note in our industry series. 210 frozen questions across three more industries, July 2026 — one engine (Doubao), seven buyer-intent layers, one API run per question. A wide-space snapshot: not market share, not consumer-app behavior, not a time-stability estimate. Assertions inside cited sources remain unverified unless separately reviewed. Raw tables: geochecktool.com/research.

The finding this time is an absence: in dietary supplements, 61.4% of answers named no identifiable candidate at all. The shelf is largely empty — and one of the names that keeps appearing is the regulator.

Dietary supplements: the vacuum shelf

70 frozen questions; search in 69/70; 108 engine-displayed queries; 441 cited source positions. An identifiable candidate appeared in only 27/70 answers (38.6%) — the lowest of the six industries we've measured. Explicit exclusions: just 1/70.

The most-adopted names top out at 3/70 — and one of the three is not a brand: the State Administration for Market Regulation's special-food verification platform (the official registry for supplement approvals), which the engine points users to for checking product claims. The pattern we saw in beauty (the regulator's app inside the answers) repeats here, but with a difference: in beauty the engine names specific SKUs alongside the regulator; in supplements it mostly names no identifiable candidate. For international supplement brands entering China, no exact-name candidate appeared in more than 3/70 answers in this snapshot. That is low measured occupancy, not proof that the category is competitor-free.

Maternity & baby: the fragmented shelf

70 frozen questions; search in 67/70; 119 queries; 399 cited positions. Candidates in 40/70 answers; 7/70 answers (10.0%) explicitly excluded a named entity — the second-highest veto rate we've measured, behind only luxury (20.0%).

No name reached even 5% of answers: the top adopted name appears 3/70, followed by a wide tier at 2/70 that includes international brands — Mead Johnson's 蓝臻 line, Kabrita, Neocate, Pigeon, Dr. Brown's, Philips AVENT, Hegen, COMOTOMO. International names are present but nobody owns the shelf; answers scatter across dozens of products, and the veto layer is active. In a fear-heavy category (infant feeding), that combination — fragmentation plus active exclusion — is where record-layer strength decides who survives the answer.

Pet food: the certified shelf

70 frozen questions; search in 70/70 — the only industry so far at 100%; 131 queries; 475 cited positions. Candidates in 50/70 answers; exclusions 6/70.

The top of the leaderboard is split between a domestic brand (5/70) and SGS — the certification and inspection company — also at 5/70. The verification layer keeps buying its way into answers across industries: CCIC in luxury, the regulator's platforms in beauty and supplements, SGS in pet food. Below the top, the shelf mixes domestic names with international ones (Acana, Farmina, Purina Pro Plan) in the 3–4/70 band.

The pattern across six industries

Industry	Candidate rate	Explicit exclusions	Who keeps showing up
Luxury	62/70	14/70 (20.0%)	Resale & authentication platforms
Pet food	50/70	6/70	A certifier (SGS) tied for #1
Beauty	53/70	6/70	SKUs + the regulator's app
New cars	44/70	4/70	Content platforms
Maternity & baby	40/70	7/70 (10.0%)	Fragmented; internationals at 2/70 each
Supplements	27/70 (38.6%)	1/70	Mostly nobody — plus the regulator

Two structural constants are emerging: the verification layer (regulators, certifiers, authentication platforms) appears inside answers in five of six industries; and shelf occupancy varies enormously — from luxury's crowded, veto-heavy shelf to the supplement vacuum. Where measured occupancy is low, a brand that builds a retrievable, genuine record layer enters a less occupied answer set in this dated snapshot. That is a measurable, time-limited condition — snapshots are dated for a reason.

Method: identical to note one — frozen 70-question banks per industry (≥50% built from real public buyer questions), seven intent layers, one API run per question, July 2026; engine-displayed queries and cited source positions logged verbatim; per-answer candidate extraction Tier-2 coded with pre-flight double-coding calibration. API and consumer-app are separate surfaces, 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.