360's Nano AI routes a single query across 80+ models and merges the answers. That adds two ranking events your per-model brand score has never measured.

The Router Layer: When One Chinese AI App Asks 100 Models About Your Brand at Once (2026)

Every brand visibility score you own assumes a one-to-one relationship: one query, one model, one answer. You test Doubao, you get a Doubao score. You test DeepSeek, you get a DeepSeek score. Six models, six numbers, one dashboard.

That assumption is now wrong for a growing slice of Chinese AI search traffic — and the reason is 360's Nano AI (纳米AI), a product that does not have a model of its own. Nano AI integrates over 80 large language models from domestic and international vendors, decomposes an incoming query, and dispatches sub-tasks to whichever model it judges best suited, then synthesizes a single answer. AI product tracker AICPB has repeatedly placed Nano AI at or near the top of China's AI search category and second globally behind Microsoft's Bing, ahead of Perplexity. In July 2026, 360 extended the same architecture to the enterprise with Nano Work, a multi-agent platform for cloud-executed task workflows.

For brand marketers, the important detail is not 360's market position. It is the architecture. On a router product, your brand is never evaluated by "a model." It is evaluated by a selection function that chooses models, and then by a synthesis step that reconciles what they said. Those are two ranking events your visibility score has never measured.

The measurement gap in one sentence

A per-model brand score tells you what a model would say if asked directly. A router decides whether that model gets asked at all.

Consider a Chinese consumer query like "推荐适合敏感肌的高端面霜" (recommend a high-end face cream for sensitive skin). On Doubao, this hits one model with one retrieval stack, and your brand either surfaces or it doesn't. On a router, the query may be split — ingredient safety to one model, price comparison to another, user-review synthesis to a third pulling Xiaohongshu and SMZDM content — and the final answer is a merge.

If your brand scores 85 on Model A and 20 on Model B, your dashboard reports an average somewhere in the sixties. But if the router's task decomposition sends the recommendation sub-task to Model B and only the price lookup to Model A, your real exposure on that surface is closer to 20. The average was never the number. Routing weight was.

This is a different problem from the one we covered in the query-routing trap, where the issue was coverage across separate apps. Here, a single app is doing the routing internally, invisibly, and per-query.

Three ranking events, not one

Layer	What decides it	Does your current score measure it?
Model selection	Router's task classifier — matches query type to model capability	No
In-model retrieval	Each model's search stack and source priors	Yes (this is what you test)
Cross-model synthesis	Merge and reconciliation logic when models disagree	No

The synthesis layer deserves particular attention, because Chinese models genuinely disagree about brands. We have documented divergence repeatedly: Wenxin systematically discounts international brands relative to peers, and Qwen, Doubao and Kimi judge brands on materially different evidence bases. A router that queries three of them and merges is running an unweighted vote across three incompatible priors.

Two failure modes follow directly. First, majority erasure: a brand strong in one model and absent in two others gets dropped in synthesis, because the merge treats absence as evidence rather than as missing data. Second, conflict flattening: when models disagree on a factual attribute — price, availability, formulation, provenance — the router either picks one arbitrarily or hedges into vagueness. Vague answers do not convert. We have already seen how price hallucination costs sales even at a perfect brand score; synthesis makes that class of error more likely, not less, because it introduces a reconciliation step between sources that were never designed to agree.

The honest caveat: how big is this actually?

Here the data is genuinely messy, and brand teams should know it before reallocating budget.

Rankings that measure web visits and app downloads have put Nano AI at the top of China's AI search category. QuestMobile's panel-based MAU measurement, by contrast, has placed Nano AI in a far more modest 1–5 million band, with negative compound growth over one measured period, and app-store-derived figures around 8.9 million MAU, ninth domestically. Those are not small discrepancies. They are different orders of magnitude.

The gap is explainable — 360 distributes through browser and desktop search properties where visit counts inflate relative to app-panel MAU — but the direction of the correction matters. Against Doubao's roughly 382 million MAU as of mid-2026, a router product in the single-digit-millions band is not a primary channel. It is a leading indicator.

Treat it that way. The architecture is the story, not the current user count. Routing is not a 360 idiosyncrasy: enterprise buyers are already standardizing on multi-model aggregation as a scheduling layer, and Chinese workplace agents route around Doubao for the same reason. When routing logic sits between the user and the model, per-model scores stop being the unit of visibility — whether that router is a consumer app with 9 million users or an enterprise platform with 900 seats.

What actually moves brand outcomes on a router

Three properties matter more on router surfaces than on single-model surfaces.

Cross-model consistency beats single-model peak. On Doubao alone, a score of 95 is worth more than a score of 60. Across a router, a brand that scores 60 on all six models will typically survive synthesis better than a brand that scores 95 on one and 15 on the rest, because the merge step rewards agreement. This inverts the standard optimization instinct, which is to double down on the highest-traffic model. The six-platform score gap — brands ranking 68 on Kimi and 11 on Wenxin — is precisely the profile that routers punish.

Entity resolution becomes load-bearing. A router matching sub-answers across models has to determine that Model A's mention and Model B's mention refer to the same brand. Inconsistent naming across Chinese-language sources — the alias gap between official names and the nicknames consumers actually use — causes the merge to treat one brand as two, or drop the weaker variant entirely.

Factual agreement across sources is now a ranking input, not just an accuracy issue. If your product specs, pricing tiers, and availability differ across Baidu Baike, your Chinese site, and JD or Tmall listings, models will disagree, and the router will hedge. The three-baike problem has always been a credibility risk. On router surfaces it becomes a visibility risk.

Takeaway for Brand Marketers

Three concrete moves, in priority order:

Stop reporting a single averaged brand score. Report the distribution — minimum, median, and spread across models. On router surfaces, your floor predicts outcomes better than your mean. A brand with a 15-point floor is exposed regardless of where its ceiling sits.
Reallocate from peak optimization to floor repair. If you have one model scoring under 25, that model is now a liability on every router surface, not just a missed opportunity on its own app. Wenxin is the usual culprit for international brands. Fixing a 15 to a 45 will do more for router-surface visibility than moving an 85 to a 95.
Add router surfaces to your test set at low weight, now. Nano AI and comparable aggregators do not justify a budget line at current MAU. They justify a monitoring line. Run your top 20 category queries through a router surface monthly, and compare the answer to what your per-model dashboard predicted. The size of that delta is your exposure to this architecture — and it is the number that will tell you when routing stops being a leading indicator and starts being a channel.

The broader shift is straightforward: as Chinese AI products mature, fewer of them are models and more of them are orchestration over models. Brand visibility measurement built on the assumption of direct model access has a shelf life. It is worth knowing how long yours is.

Related: compare your brand's per-model score distribution across all six Chinese AI models on the hubGEO brand index.

Sources: AICPB AI product rankings (China AI search category); QuestMobile 2026 half-year China AI app report; 360 Nano Work launch announcement (July 2026); The Paper and Sina Finance coverage of Nano AI's search positioning.