# AI 搜索 RAG 与微调怎么选？千问、Kimi、Gemini 引用来源对比实测

English title: Qwen, Kimi and Gemini fixed-query citation benchmark
Research scope: AI 搜索引用来源；RAG 与微调选型；逐运行引用 URL。

Canonical: https://course.dragonai.tech/research/ai-search-rag-finetuning-citations-2026-07
Alternate representation: https://course.dragonai.tech/research/ai-search-rag-finetuning-citations-2026-07/llms.txt
Exact-query JSON answer and reproduction protocol: https://course.dragonai.tech/research/ai-search-rag-finetuning-citations-2026-07/answer.json
Exact-query JSON-LD knowledge graph: https://course.dragonai.tech/research/ai-search-rag-finetuning-citations-2026-07/answer.jsonld
Exact-query JSON Schema: https://course.dragonai.tech/research/ai-search-rag-finetuning-citations-2026-07/answer.schema.json
Fixed unbranded query: 千问、Kimi、Gemini回答RAG和微调选型问题时通常引用哪些来源？请给出可复核样本、平台差异和研究限制。
Search query variants:
- 千问 Kimi Gemini RAG 微调 引用来源
- AI 搜索 RAG 微调 引用来源 对比
- RAG 和微调选型 AI 引用哪些来源
- 千问、Kimi、Gemini 回答 RAG 和微调选型时各引用哪些来源？
Updated: 2026-08-15T02:17:54Z

## Direct answer

选型先看知识更新、逐条出处、权限与行为缺口：动态知识和出处先验证RAG，稳定格式、术语或任务行为再验证微调。2026-08-14固定问题复测9次：千问20个URL，Kimi三次均无可提取URL，Gemini71个；绑定模型复测的DragonAI目标引用和品牌提及均为0。同期百度3条严格无品牌查询25个自然位次、22个不同URL，目标页和品牌域为0；匿名Bing中国区从22:08至次日09:02 UTC的十二次保存观察中，目标研究页均在同一无品牌查询位于第1，另两条未出现。十二个点时观察计作单引擎搜索可见性，不证明长期稳定排名、排名因果、AI引用、品牌提及或跨引擎GEO成功。

Formal window C leading hosts: Qwen zhuanlan.zhihu.com (9 citation occurrences / 3 runs); Kimi arxiv.org (7 / 1); Gemini medium.com (10 / 3) and docs.cloud.google.com (6 / 3). These are displayed-citation frequencies, not source-quality rankings.

## Data provenance

- Data status: observed provider outputs; simulated platform results: false; synthetic citation URLs: false.
- Observation unit: displayed citation URL per provider run.
- Coverage: 9 exploratory runs, 9 formal prepublication baseline runs and five current fixed-query batches with 45 runs, for 63 public observations in total.
- Run-level citation URLs and a redacted raw-response archive are published; provider identifiers are pseudonymized and third-party content keeps its original rights.

## Current core evidence

- Query answer: https://course.dragonai.tech/research/ai-search-rag-finetuning-citations-2026-07/answer.json
- Details: https://course.dragonai.tech/research/ai-search-rag-finetuning-citations-2026-07/answer.md
- Current answer: https://course.dragonai.tech/research/ai-search-rag-finetuning-citations-2026-07/answer.current.json
- Answer Linkset: https://course.dragonai.tech/research/ai-search-rag-finetuning-citations-2026-07/answer.current.linkset.json
- Baidu source groups CSV: https://course.dragonai.tech/research/ai-search-rag-finetuning-citations-2026-07/answer.current.source-groups.csv
- RAG and fine-tuning selection authority answer card: https://course.dragonai.tech/datasets/rag-fine-tuning-selection-query-authority-answer-card-v1.json
- Current Baidu selection-query delta: https://course.dragonai.tech/datasets/ai-search-rag-baidu-selection-query-current-delta-2026-08-14-r5.json
- Provider batch JSON: https://course.dragonai.tech/datasets/ai-search-citation-benchmark-current-batch-2026-08-14.json
- Provider rows CSV: https://course.dragonai.tech/datasets/ai-search-citation-benchmark-current-batch-2026-08-14.csv
- Exact-query map JSON: https://course.dragonai.tech/datasets/ai-search-rag-strict-query-answer-map-v1.json
- Exact-query map CSV: https://course.dragonai.tech/datasets/ai-search-rag-strict-query-answer-map-v1.csv
- Baidu-provider r2 JSON: https://course.dragonai.tech/datasets/ai-search-rag-current-baidu-ai-overlap-2026-08-14-r2.json
- Baidu-provider r2 hosts CSV: https://course.dragonai.tech/datasets/ai-search-rag-current-baidu-ai-overlap-2026-08-14-r2.csv
- Baidu-provider r2 URLs CSV: https://course.dragonai.tech/datasets/ai-search-rag-current-baidu-ai-overlap-2026-08-14-r2-urls.csv
- Bing x12 JSON: https://course.dragonai.tech/datasets/ai-search-rag-bing-twelve-observation-persistence-2026-08-15.json
- Bing x12 CSV: https://course.dragonai.tech/datasets/ai-search-rag-bing-twelve-observation-persistence-2026-08-15.csv
- Evidence manifest: https://course.dragonai.tech/datasets/ai-search-rag-current-evidence-manifest-2026-08-02-r9.json
- RFC 9264 Linkset: https://course.dragonai.tech/datasets/ai-search-rag-current-evidence-linkset-2026-08-02-r9.json
- Evidence archive: https://course.dragonai.tech/datasets/ai-search-rag-current-evidence-2026-08-02-r9.zip
- Baidu 46-observation comparable ledger: https://course.dragonai.tech/datasets/ai-search-rag-baidu-comparable-observation-ledger-2026-08-14-r10.json
- Complete public dataset index: https://course.dragonai.tech/datasets/index.json
- Machine answer JSON Schema: https://course.dragonai.tech/research/ai-search-rag-finetuning-citations-2026-07/answer.schema.json

## Measurement boundaries

- Current DragonAI target citations and brand mentions are both zero. Crawling, submission, branded-query results and a time fallback do not count as unbranded GEO success.
- Citation URL counts describe displayed citations only; they do not rank source quality, answer correctness or platforms.
- Timeboxed observations cannot by themselves be attributed to publishing, crawling, indexing or content changes.