Most brands tracking their visibility in Chinese AI search are measuring the wrong layer. They send a prompt — "推荐几个高端护肤品牌" — record which brands appear in the answer, and compute a score. That score is real, but it describes the output of a process the brand never saw.
Since DeepSeek shipped Deep Thinking 2.0 in April 2026, introducing multi-turn reasoning that automatically decomposes complex questions into sub-questions and solves them step by step, the mechanism that determines brand presence has moved upstream. The model no longer answers the prompt. It answers three to six sub-questions it wrote itself, each with its own retrieval pass, and then assembles an answer from whatever those passes returned. A brand that loses all of the sub-queries is absent from the final answer — and the brand's monitoring dashboard records a zero with no explanation attached to it.
This matters more in China than anywhere else, because Chinese models draw their citations from tightly enclosed ecosystems. Query fan-out multiplies that enclosure rather than diluting it.
What fan-out actually does to retrieval
AI answer engines increasingly decompose complex queries into multiple retrieval tasks, and reasoning models do this most extensively — each sub-query runs its own retrieval, and each pulls different sources. The practical consequence is that a single prompt is not one visibility event. It is several, chained, with the early ones invisible.
Consider what a reasoning model plausibly does with "which mid-range hotel in Shanghai is best for a business trip." It does not retrieve once on that string. It decomposes: what counts as mid-range in Shanghai pricing; which districts serve business travelers; which properties have meeting facilities; what recent guest reviews say about each. Four retrieval passes. A hotel brand with excellent positioning in review content but no presence in district-level or pricing-tier content loses two of the four passes — and if the shortlist is formed at the district pass, the brand never reaches the review pass where it was strong.
This is a different failure mode from the ones brands currently plan for. It is not a content-quality problem or an entity-recognition problem. It is an elimination-round problem: the brand was competitive in the final, and never got there.
Why China's citation concentration makes this worse
In an open web, fan-out would tend to diversify sources — more sub-queries, more corners of the internet, more chances for a brand to surface somewhere. Chinese models work differently. A July 2026 study of citation source distribution found that Doubao draws roughly 60% of its citations from ByteDance properties (Toutiao at about 35%, Douyin at about 25%); ERNIE draws around 55% from the Baidu ecosystem (Baijiahao 35%, Baidu Baike 20%); DeepSeek leans on CSDN (20%) and Zhihu (15%); and Kimi favors Zhihu (18%).
Run fan-out inside those numbers and the effect inverts. Every additional sub-query is another draw from the same small pool. If a brand has no Toutiao footprint, it does not lose one retrieval pass in Doubao — it loses every pass that routes to ByteDance properties, which is most of them. Fan-out does not give the brand more chances. It gives the same gap more opportunities to matter.
| Model | Primary citation concentration | What fan-out amplifies |
|---|---|---|
| Doubao | ~60% ByteDance (Toutiao 35%, Douyin 25%) | Absence from Toutiao/Douyin compounds across every sub-query |
| ERNIE | ~55% Baidu (Baijiahao 35%, Baike 20%) | A thin or contested Baike entity propagates into each pass |
| DeepSeek | CSDN 20%, Zhihu 15% | Technical and Q&A depth gaps block sub-queries that demand specifics |
| Kimi | Zhihu 18%, long-document preference | Shallow content fails structured sub-questions that ask for comparison |
There is a second-order effect worth naming. Because DeepSeek's Deep Thinking 2.0 also added confidence annotations and source tracing, the model now has a mechanism for declining to assert a weakly sourced claim. A brand whose supporting evidence is thin does not simply rank lower in the sub-query — it can be dropped entirely because the model is not confident enough to state anything about it. Hedging is a visibility outcome, not a neutral stance.
The measurement gap this creates
The GEO measurement field has partly caught up to this. Some tracking platforms now cover query fan-out analysis across 20 AI models, showing how a single prompt branches into sub-queries that shape citations. But fan-out analysis remains rare in China-focused measurement, where the common practice is still surface-prompt scoring across Doubao, Kimi, DeepSeek, Qwen, Wenxin, and Hunyuan.
That matters because investment is accelerating faster than measurement sophistication. Roughly 32% of digital marketing leaders now name GEO their top priority for 2026, up from about 12% of 2025 digital budgets. Budgets are tripling against a metric that reports outcomes without causes.
The specific distortion is that surface-prompt scoring produces false negatives that look like content problems. A brand scores 20 out of 100 on Doubao for a category query. The natural reading is "our content isn't good enough." The actual cause may be that the model's sub-queries all asked about price tiers and regional availability, the brand has neither documented in any ByteDance property, and the brand's genuinely strong product-quality content was never retrieved because no sub-query asked for it. Those two diagnoses lead to opposite remediation plans — improve content quality versus fill structural coverage gaps — and the score alone cannot distinguish them.
What this implies about GEO content strategy
The strategic shift is from competing on the category question to covering the decomposition space around it.
Brands currently optimize for the query they want to win: "best premium sunscreen in China." Under fan-out, winning that query requires presence across the sub-questions a reasoning model would generate from it — SPF and PA standards, formulation types, price bands, skin-type suitability, availability channels. Content that answers the headline question beautifully but addresses none of the sub-dimensions is structurally fragile, because it can only be retrieved by one pass out of five.
This also reframes the long-running question of which content formats matter. Kimi rewards longer, better-structured documents with clear sections, downloadable assets, and deeper analysis, while Doubao prefers mobile-first, conversational, socially legible content suited to the ByteDance ecosystem. Under fan-out, that difference is not merely a formatting preference — it determines how many sub-queries a single asset can satisfy. A well-sectioned long document can be retrieved by several different sub-queries, each matching a different section. A short social post can usually satisfy one. Format choice is effectively a bet on retrieval surface area.
Takeaway for Brand Marketers
Three things change in practice.
Stop reading a single score as a single verdict. When a brand scores near zero on a category prompt, the useful next question is not "what's wrong with our content" but "which sub-questions did the model ask, and where did we drop out." If your measurement vendor cannot tell you that, you are buying a thermometer and asking it for a diagnosis. Ask whether their China coverage includes reasoning traces or fan-out analysis, not just final-answer scraping.
Audit coverage by dimension, not by keyword. Build the list of sub-questions a reasoning model would generate from your top five category queries — price tier, regional availability, technical specification, use-case fit, comparison against named competitors — and check whether your brand has substantive, citable content for each, inside the ecosystems each model actually draws from. Gaps in that matrix predict invisibility more reliably than keyword rankings do.
Match asset structure to retrieval breadth. One long, well-sectioned, data-carrying document placed in a high-citation property will typically be retrieved by more sub-queries than five short posts of equivalent total length. Where budget is constrained, depth in the right ecosystem beats volume across the wrong ones — particularly for DeepSeek and Kimi, where confidence annotation and source tracing mean unsupported claims get dropped rather than discounted.
The uncomfortable implication for 2026 planning is that the brands most exposed here are not the ones with weak content. They are the ones with strong, narrow content — excellent on one dimension, absent on four — who will watch their scores fall as Chinese models get better at reasoning, and will have no idea why.
Related: track how brands in your category currently score across all six Chinese models on the GEO Hub brand index.
Sources: DeepSeek Deep Thinking 2.0 feature documentation (April 2026); July 2026 citation source distribution study across Chinese AI engines; GEO budget allocation surveys for 2026; platform content-preference analyses for Doubao and Kimi.