The Camera Query: China's AI Is Answering 100 Million Brand Questions a Day That Contain No Words

Every brand visibility score you have ever seen was built on the same assumption: a person types or speaks a question, and a model answers in text. "Best moisturizer for dry skin." "Reliable German sedan under 400,000 yuan." "Which running shoes for flat feet." Every GEO tool on the market — ours included — samples prompts like these across Doubao, Kimi, DeepSeek, Qwen, Wenxin and Hunyuan, counts mentions, and returns a score from 0 to 100.

In April 2026, Taobao's visual search tool Pailitao handled close to 100 million searches in a single day, surfacing roughly 1.5 billion product images in response. Its user base has grown nearly 150% in four years. None of those queries contained a brand name. Most contained no text at all.

That is the gap this piece is about. A fast-growing share of Chinese commercial intent now enters the AI layer through a camera rather than a keyboard — and the retrieval pipeline behind a camera query is not the one your visibility score measures.

Two pipelines, one score

When a Chinese consumer types "推荐几个抗老精华" into Doubao, the model runs the process GEO has spent two years learning to influence: query understanding, retrieval from an indexed corpus (Douyin, Toutiao, SMZDM, Baike, brand-owned pages), synthesis, and a shortlist of named brands. Content you publish can enter that corpus. Authority citations can lift your odds. This is the game everyone is playing.

When the same consumer points a phone at a serum bottle on a friend's bathroom shelf, almost none of that applies. The pipeline is:

1. Image embedding — the photo becomes a vector.
2. Nearest-neighbour match against a catalogue index of SKU images.
3. Entity resolution — matched listings resolve to a product, then to a brand.
4. Ranking of visually similar alternatives, priced and in stock.
5. Optionally, a language model writes a sentence or two on top.

Steps 1 through 4 never touch a single word you wrote. Your position is determined by whether your SKU images are densely represented in the commerce index, how distinctive your packaging is under bad lighting, and how your listings are titled and categorised — not by your content strategy.

The commercial stakes are not theoretical. Taobao has already layered an "AI find a lower price" function onto image search, which means a camera pointed at your product returns your product and the cheapest visually similar thing in the catalogue. In a text query, a competitor has to earn a mention. In a camera query, a competitor is placed next to you automatically, by geometry.

The glasses make it ambient

Visual query volume would matter even if it stayed on phones. It is not staying on phones.

IDC expects China's domestic smart glasses shipments to exceed 4.915 million units in 2026, the point it describes as the market's large-scale inflection. Chinese manufacturers are forecast to ship 22.67 million units globally, up 56.3% year on year and roughly 45% of the world market. Omdia's separate count put global AI glasses shipments at 8.7 million units, with mainland China the fastest-growing market. In January 2026, AI glasses entered China's national subsidy programme for the first time — a 15% discount capped at 500 yuan — and monthly sales stepped up from there.

What these devices do, commercially, is collapse the distance between seeing a product and querying it. Alibaba's Quark AI Glasses, from 1,899 yuan for the entry model to 3,799 yuan for the Snapdragon AR1-based S1, do exactly what the hardware implies: the wearer looks at a product, and the device returns its Taobao price. At MWC 2026 Alibaba announced its next glasses run on the Qwen assistant, with international availability starting in selected Asian and European markets.

Note the ownership structure. The dominant Chinese visual-commerce surface is built by Alibaba, indexed against Taobao and Tmall, and answered by Qwen. The dominant Chinese text-answer surface is built by ByteDance, indexed heavily against Douyin and Toutiao, and answered by Doubao — which reached roughly 382 million MAU and, by iResearch's 2026 count, handles over 1.5 billion natural-language queries a day, and whose 2.0 release specifically upgraded fine-grained visual perception and spatial reasoning.

Two companies. Two indexes. Two different definitions of what makes your brand findable. Most international brands are measuring one and funding one.

What a camera query rewards

Text GEO and visual GEO reward almost opposite assets:

That last row deserves attention. In a text answer, a parallel-import seller almost never appears — models cite brand and media sources. In a visual match, an unauthorised listing carrying a photo of your product is indistinguishable from an authorised one at the embedding level. The camera query can resolve your brand to somebody else's storefront, at somebody else's price, with no content problem to diagnose and no mention anywhere for your dashboard to record.

For categories where packaging is the product's public face — beauty, spirits, FMCG, eyewear, sneakers, small appliances — this is not a marginal channel. It is a parallel discovery surface with its own winners, and the brands winning it are the ones whose Chinese marketplace presence is dense, well-photographed and tightly controlled.

Why your score can't see it

A visibility score is only as honest as its sampling frame. Ours, and every competitor's, samples text prompts. That produces three specific blind spots.

It cannot detect misidentification. A brand can score 92 on text queries and still be visually confused with a domestic imitator in a quarter of camera matches. Nothing in the text sample would move.

It cannot detect substitution. With "find a lower price" behaviour built in, the relevant metric is not were you returned but what was returned beside you, and at what price gap. Text scoring has no equivalent measurement.

It cannot attribute channel damage. When camera queries resolve to grey-market listings, the loss shows up as margin erosion and inconsistent pricing — problems that land on the distribution team, not the marketing team, and never get connected back to AI discovery at all.

We have written before about measurement frames that quietly stop matching reality: model version churn, memory layers, router apps, preinstalled phone assistants. The camera query belongs on that list, and it is the largest of them by raw volume. One surface, one company, one day: 100 million queries.

Takeaway for brand marketers

Do not rebuild your GEO programme around this. Text answers still carry the highest-consideration queries, and the Doubao–DeepSeek text duopoly still deserves the bulk of the budget. But treat visual discovery as a second, separately-owned surface, with four concrete actions.

1. Audit your own camera result. Photograph your top ten SKUs in ordinary conditions — shelf, home, bad light, partial angle — and run them through Taobao image search. Record three things: is the top match your product, is the top match an authorised seller, and what is the price gap to the cheapest visually similar alternative. This takes an afternoon, and most brands have never done it.

2. Treat SKU imagery as index coverage, not creative. Multiple angles, clean backgrounds, consistent packaging shots across every authorised listing. Sparse or inconsistent imagery is the visual equivalent of a thin content footprint.

3. Put marketplace listing hygiene on the GEO owner's desk. Category tags, product titles and attribute completeness are now discovery signals, not just conversion signals. In most organisations these sit with e-commerce operations and never reach the team accountable for AI visibility.

4. Measure the price gap, not just the mention rate. In visual commerce the competitive threat is not absence — it is adjacency. Track how often, and by how much, a cheaper lookalike gets placed beside you.

The brands that scored 100 in our text benchmarks did so by being densely, consistently present in the sources Chinese models read. The same principle applies here. The sources are just images — and almost nobody is checking.

Related: see our brand visibility scores across China's six major AI models, and our earlier analysis of preinstalled phone assistants, the other measurement blind spot hiding in plain sight.