How AI Search Recommends: Query Rewrites and Sources in 5 Chinese AIs

何瑜 — Sat, 07 Nov 2026 08:06:37 +0000

This is the English translation of the Chinese original, first published on 2026-11-05. Search queries and on-figure labels are kept in Chinese, with English glosses in brackets or captions.

Revised 2026-11-07: added a comparison of four platforms' final answers (new section "Who did the AI end up recommending?"), corrected the Kimi citation count, and aligned collection times and session history with the dataset card.

Bottom line: AI search does not search with your exact words. It first works out where you are, rewrites the question into several search queries, and then picks part of what it finds to hand to the model. Each platform does these steps differently, and what you said earlier gets mixed in too, so for the same question only 2 providers appeared in all four recommendation lists.

I asked 5 Chinese AI assistants the same question, "What medical aesthetics clinics are there in Suzhou?" (苏州当地有哪些医美？), and captured the network requests behind each answer. All 5 were asked, but the data is uneven: search queries can be compared for Doubao, Qwen and Kimi; final answers for those 3 plus Tencent Yuanbao; DeepSeek is not compared for this question, and its multi-round retrieval comes from the next one. The differences visible in this one question are concentrated in the three steps below.

Figure 1: From question to answer. The black boxes mark the three steps where this capture showed differences. Labels in Chinese: 你的问题 your question → 位置识别 location detection → 搜索词改写 query rewriting → 检索 retrieval → 来源筛选 source filtering → 交给模型 sent to the model → 回答 answer.

Key points

Location detection: 3 platforms handled location-related information beyond the question itself, such as IP geolocation, browser location permissions, and a region-search tool.
Query rewriting: from the same sentence, one platform produced 1 search query, another 3, another 5, and they went in different directions.
Conversation history gets into the queries and the answer: one platform carried a company name from the previous turn into this turn's search queries, and reshaped its whole answer around what had been said earlier.
Source filtering: one platform retrieved 20 items and sent only 5 of them to the model.
Final picks: the four platforms recommended between 6 and 16 providers; only 2 appeared in all four.
What it means for GEO: what you need to study is the queries after the AI rewrites them, not just the keywords users type; when testing recommendation rates, the conversation context has to be controlled as a variable.

Note: clinic and pharmaceutical company names are replaced with XX. [Observed] means seen directly in the capture, [Inferred] means inferred from what was observed, and [Unknown] means not captured or not possible to determine.

Does AI search get your location?

Yes. At least 3 platforms do.

Platform	What was observed	Evidence
Doubao	The request carried 3 browser location-permission states, but no coordinates	[Observed]
Qwen	Before the question, it looked up the IP's location; the analytics field user_city was "苏州市" (Suzhou), and the intent was labeled "本地生活-地点周边信息查询" (local services: nearby place information)	[Observed]
Tencent Yuanbao	Called a region_search tool	[Observed]
Kimi	Not observed	[Unknown]
DeepSeek	Not observed	[Unknown]

But because the question already said "Suzhou", I can't tell from this test whether location affected the recommendations [Unknown]. For local service businesses, the point worth noting is this: even when the user doesn't name a city, the AI may work it out on its own.

What does AI search actually search for?

Not your exact words. The AI first rewrites the question into search queries, a step known as query rewriting.

Platform	Queries	Actual search queries
Qwen	1	苏州当地医美机构及医疗美容医院 [local medical aesthetics providers and cosmetic hospitals in Suzhou]
Kimi	3	苏州医美机构知名排行榜 2025 [Suzhou medical aesthetics providers, well-known, ranking 2025]; 苏州整形医院公立医疗美容科 [Suzhou plastic surgery hospital, public, medical aesthetics department]; 苏州医美产业企业 [Suzhou medical aesthetics industry companies]
Doubao	5	苏州正规医美机构公立三甲皮肤科连锁医美敏感肌 [licensed providers in Suzhou, top-tier public hospital dermatology, chain clinics, sensitive skin]; 苏州姑苏XX医美诊所介绍 [about the XX medical aesthetics clinic in Gusu, Suzhou]; 苏州吴中高新区医美机构推荐敏感肌 [Wuzhong and High-Tech Zone, Suzhou: recommended providers, sensitive skin]; 苏州敏感肌医美面诊避坑 [Suzhou, sensitive skin: pitfalls to avoid at consultations]; 苏州光子嫩肤红血丝机构测评 [Suzhou IPL photorejuvenation for facial redness: provider reviews]

Tencent Yuanbao's rewritten queries were not captured this time, and DeepSeek is not counted for this question.

Figure 2: Search queries rewritten from the same sentence: 1 from Qwen (千问), 3 from Kimi, 5 from Doubao (豆包). The figure drops the leading 苏州 (Suzhou) from each query and shortens long ones; the full queries are in the table above.

The three approaches differ a lot [Observed]:

Qwen basically searches for the original meaning.
Kimi adds rankings, public hospitals, and an industry angle.
Doubao breaks it down the most, searching separately for sensitive skin, consultation pitfalls, and IPL photorejuvenation. But skincare had come up earlier in this conversation, so words like "sensitive skin" and "facial redness" most likely came from the history rather than from Doubao's own breakdown [Inferred].

The user asks about "medical aesthetics in Suzhou", but what the AI actually searches for may be "pitfalls to avoid at sensitive-skin consultations". Content built only around the original keyword won't be found by these searches.

Does what you talked about earlier affect AI search?

Yes, and not only the search queries. This time Doubao showed the whole chain.

A company from the previous question showed up in the queries. Doubao's second query was "苏州姑苏XX医美诊所介绍" [about the XX medical aesthetics clinic in Gusu, Suzhou]. I never mentioned this clinic in this question, but it shares its name with the pharmaceutical company in my previous question, "XX药业是什么" [What is XX Pharmaceutical?] [Observed].

Turn	Content
Question 2	"XX药业是什么" [What is XX Pharmaceutical?]
Question 3	"苏州当地有哪些医美" [What medical aesthetics clinics are there in Suzhou?]
Query generated by Doubao	"苏州姑苏XX医美诊所介绍" [about the XX medical aesthetics clinic in Gusu, Suzhou]

The answer was tailored to the history too. Doubao's answer opened with "based on your situation", listed several skin concerns, and said one provider was close to my district. I said none of this in the question; it came from skincare talk earlier in the same conversation [Observed]. The clinic sharing its name with the pharmaceutical company from the previous question got its own entry in the answer, marked as a key pick [Observed].

De-identification: the skin concerns, the clinic's name, and what the answer said about that clinic are not given here.

Conversation history showed up in the search queries and in the final answer [Observed]. From this I infer that the AI draws on conversation history both when rewriting queries and when composing the answer [Inferred].

One caveat: in this Doubao conversation the same three questions had already been asked once before, so "XX Pharmaceutical" appeared in the history more than once, which may have amplified the effect. Qwen's and DeepSeek's conversations also had an earlier round, but showed nothing similar this time; that is not evidence that they don't use conversation history [Unknown].

This matters a lot for GEO testing: asking directly in a new conversation and asking after a few turns may lead the AI to search for different things.

Does AI search use everything it finds?

No. Being retrieved, being read, and being cited are three different things.

Qwen: 20 items retrieved, 5 sent to the model. The material_stat field shows 20 items retrieved, of which only 5 were finally sent to the model [Observed]. How those 5 were chosen can't be reconstructed [Unknown].

Figure 3: Qwen retrieved 20 items, and only 5 were sent to the model. Labels in Chinese: 检索到 20 条素材 20 items retrieved → 筛选（规则未知） filtering (rules unknown) → 送进模型 5 条 5 sent to the model. Data: the material_stat field in Qwen's request.

Doubao: Douyin and business-registration information make up a sizable share. The page said it referenced 29 sources, and I captured 17 of them with links [Observed]:

Source type	Count
Related to the same-named clinic (Qichacha, a district government filing notice, a yellow-pages listing, Aiqicha, a job posting, a review page)	6
Douyin videos	4
Hospital websites	2
China.com (中华网)	2
Other (Hangzhou Net 杭州网, Chao News 潮新闻, Xiaohe 小荷, 1 each)	3

Earlier, when I asked about skincare products, 10 of the 29 sources also came from Douyin. Entity information such as business registrations and official filings also made it into the citations. But these are only the sources that were cited, not all the candidates, and how much weight each one carries can't be seen [Unknown].

Kimi: one round of retrieval, citations pinned to paragraphs. Kimi searched once; its answer carried 12 citations drawn from 8 distinct web pages. Each citation can be linked to specific paragraphs in the original web pages, not just to a URL [Observed].

DeepSeek: searches again based on the first round of results. This data comes from the next question: with deep thinking (深度思考) turned on, I asked about the price of IPL photorejuvenation in Suzhou. It first searched 5 queries, got 30 results, and opened 7 web pages; it then searched 4 more specific queries, as specific as "苏州市立医院东区光子嫩肤 M22 价格单次" [Suzhou Municipal Hospital, East Campus, IPL photorejuvenation, M22, price per session], opened 3 more pages, and only then answered [Observed]. This is a different question and can't be compared directly with the other platforms, but it shows that some AIs run additional rounds of retrieval [Inferred].

Who did the AI end up recommending?

The four recommendation lists barely overlap.

Platform	Providers recommended	What the answer was like
Doubao	16 (6 public)	Tailored to the conversation history; a separate entry for the clinic sharing the pharmaceutical company's name
Qwen	12	A consumer-leaning list, plus a price table by treatment
Kimi	10, plus 7 names listed at the end	All 12 citations carry paragraph anchors; warns that online "top 10" rankings mostly come from medical aesthetics marketing sites
Tencent Yuanbao	6 (2 public)	The most restrained, with an address for each

Only 2 providers appeared in all four: a municipal public hospital and a national medical aesthetics chain [Observed]. 5 more appeared in three of the four (not counting the 7 names Kimi only listed at the end) [Observed].

The provider list itself isn't what this test compares, so beyond the counts no names are given. The answer texts contain personal information and judgments about commercial providers, and are not published.

This brings us back to the opening question: location, queries, sources and conversation history all differ, and so do the providers recommended in the end. Being recommended by one AI for a question doesn't mean another AI will recommend you for the same question [Inferred].

What does this mean for GEO?

Study the queries after the AI rewrites them, not just the user's original words. Of Doubao's 5 queries, 4 added a group of people (sensitive skin), a treatment (IPL photorejuvenation), a concern (pitfalls to avoid) or a district, and none was the original wording. Content that covers these narrower questions is more likely to be found.
Control the conversation context when measuring AI recommendation rates. Because skincare had come up earlier in the same conversation, both Doubao's queries and its recommendation list changed. Recommendation-rate figures that don't say whether they came from a new or an ongoing conversation are hard to reproduce.
Look at retrieved, read, cited, and recommended separately, and per platform. Qwen retrieved 20 items and sent only 5 to the model; only 2 providers appeared in all four recommendation lists. Looking only at one AI's final picks can neither explain why an optimization worked nor tell you where you stand in other AIs.

FAQ

How is AI search different from a traditional search engine?

A traditional search engine returns ranked web pages directly from the keywords. AI search first rewrites the question, may search in several rounds, and then picks part of the results for the model to generate an answer. So an AI recommendation is not the same as "the pages that rank highest in search".

Why do different AIs recommend different things for the same question?

Judging from this capture, there are at least three reasons: they handle location differently, they rewrite the question into different search queries, and they pick different sources from the search results. What you talked about earlier also changes the picks. This time, only 2 providers appeared in all four lists.

Does AI search take earlier conversation into account?

Yes. This time one platform carried a company name from the previous turn into its search query, and also reorganized its answer around what had been said earlier. The other platforms weren't seen doing this, but it can't be ruled out.

Where should GEO start?

First find out which search queries AI rewrites for your industry, then see which types of sources tend to get cited, such as official websites, business registration filings, short videos, and media coverage. When testing, keep new conversations and ongoing conversations apart.

Method and limitations

Time: 2026-10-30 23:22 to 23:35 (CST); Kimi was collected separately, 10-31 01:34 to 01:37
Conditions: one account, one continuous conversation, China Telecom network in Suzhou
Context: the target question was the 3rd question in the conversation; the first two asked about well-known skincare products from Suzhou and "What is XX Pharmaceutical?". In the Doubao, Qwen and DeepSeek conversations, the same three questions had already been asked once (not captured); Tencent Yuanbao and Kimi were new conversations
Visible: search queries, some search results and citations, some analytics fields, tool-call names; the final answers from Doubao, Qwen, Kimi and Tencent Yuanbao
Not visible: system prompts, which page text the model actually read, internal ranking rules; Tencent Yuanbao's search queries for this question, and DeepSeek's answer to it (DeepSeek did search 4 queries for this question; they are kept in the data but not counted in this article)

This was tested only once. The conclusions don't represent how the platforms behave in general and can only serve as hypotheses for later controlled experiments.

The de-identified data extracted from the captures (search queries, sources, tool calls, and analytics fields; CC BY 4.0) is public on Hugging Face: Maodian/opengeo-observation-001. The dataset card maps each figure in this article to its records and gives details such as collection times. The final answer texts contain personal information and judgments about commercial providers, and are not in the dataset.

Originally published on OpenGEO by He Yu (何瑜). Text licensed CC BY 4.0; credit "Source: OpenGEO project". Dataset: Maodian/opengeo-observation-001 (CC BY 4.0).

DEV Community: 何瑜