The Three-Layer Framework Behind AI Brand Recommendations
This is a digest. Read the full article in Chinese.
A practical three-layer framework for observing how AI engines pick brands: retrieval recall shapes the candidate pool, answer assembly shapes what gets said, and the platform ecosystem shapes how a recommendation is displayed. This is the framework we use to audit real-device answers. It is not disclosed system architecture, and it cannot be read back into model internals.
We apply it to China's leading AI engines: Doubao (ByteDance), Qwen (Alibaba), and DeepSeek. In our real-device tests, with web search on or triggered, buying-intent questions returned answers carrying citation labels; DeepSeek offers a web-search toggle, and its state changed whether answers carried retrieval citations. Checking answers against their cited sources question by question showed a clear qualitative pattern: brands absent from citable sources had much lower odds of being recommended, though engines also draw on stored knowledge, so absence from citations does not make a mention impossible. Within our sample, brands echoed consistently across several independent sources were more often described accurately across rounds, while brands resting on sparse or conflicting sources drifted more. The ecosystem layer shapes the display: in our tests, Qwen's app surfaced Taobao product cards in some categories, while Doubao maps to the Douyin ecosystem. More in AI shopping recommendations.
In a three-engine benchmark we ran in July 2026, the same questions produced noticeably different recommendation lists across the three engines, and the citation source types displayed also differed clearly. Being visible in one engine tells you little about the other two.
Influencing AI Recommendations Is Content Supply Engineering
We do not claim a shortcut into AI recommendations. The path we use, and can verify step by step, is to monitor where you are absent, fill the citation sources the engines actually display, and re-test.
The full measurement method, from frozen question sets to scheduled re-tests, is covered in how to measure AI visibility, and the GEO fundamentals in what GEO is. What the three-engine view adds: monitor each engine separately, match content to the source types that engine actually cites, and re-run the same questions on the same app so results stay comparable.
No method can guarantee a recommendation; retrieval fluctuates and answers drift. The work improves the conditions for being findable and described accurately, and whether visibility changes is established through repeated observation, regressions included. The full engine-by-engine breakdown of Doubao, Qwen, and DeepSeek is available in the Chinese edition of this post.
FAQ
- How does AI choose which brands to recommend?
- We use a three-layer observation framework: retrieval recall (which sources the answer draws on), answer assembly (which brands get named and how they are described), and ecosystem display (plain text or product cards). In our tests, brands absent from cited sources had much lower odds of being recommended. This is an analysis framework, not disclosed model internals.
- Why do different AI engines recommend different brands?
- In a three-engine benchmark we ran in July 2026, the same questions produced different recommendation lists across engines, and the citation source types displayed also differed clearly. Visibility in one engine does not imply visibility in another, so each engine needs its own measurement.
- How do I influence AI brand recommendations?
- Treat it as content supply engineering: monitor where you are absent and which sources are cited, place verified content on those source types, then re-test the same questions. No method can guarantee a recommendation; whether visibility changes is established through repeated observation.
Want to know how AI engines answer buyer questions in your category right now? We will run a real-device test and lay out, question by question, where you are absent and which sources the AI actually cites.
Get your free Growth Audit