{
  "datasetId": "ai-search-citation-source-stability-metrics-v1",
  "schemaVersion": "1.0.0",
  "createdAt": "2026-07-22",
  "language": "zh-CN",
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "licenseScope": "DragonAI's metric definitions, formulas, worked calculations, field descriptions, and Chinese explanations. Provider names and trademarks remain subject to their respective owners; third-party pages and complete model answers are excluded.",
  "disclosureZh": "DragonAI 运营 AI 产品经理课程，并使用这些指标监测本域名在 AI 搜索回答中的可见性。指标只描述固定样本，不保证抓取、收录、提及、引用、排名或业务结果。",
  "purposeZh": "用分子、分母和集合公式描述同一固定问题在重复 AI 搜索运行中的引用来源稳定性，避免把引用数量、来源重复或平台交集误写成来源质量和平台排名。",
  "metricDefinitions": [
    {
      "metricId": "citation_surface_run_coverage",
      "nameZh": "引用展示运行覆盖率",
      "formula": "runs_with_at_least_one_concrete_citation_url / completed_runs",
      "numeratorField": "runsWithAtLeastOneConcreteCitationUrl",
      "denominatorField": "completedRuns",
      "interpretationZh": "描述完成运行中有多少次展示了至少一个可归一化的具体引用 URL。",
      "doesNotProveZh": "来源正确、页面被答案吸收，或平台整体更好。"
    },
    {
      "metricId": "repeated_url_share",
      "nameZh": "重复 URL 占比",
      "formula": "distinct_urls_observed_in_at_least_k_runs / all_distinct_urls_in_provider_sample",
      "parameters": {
        "k": 2,
        "plannedRuns": 3
      },
      "numeratorField": "distinctUrlsObservedInAtLeastTwoRuns",
      "denominatorField": "distinctUrls",
      "interpretationZh": "描述同一平台、同一问题、同一短窗口内，至少跨两次运行复现的具体 URL 比例。",
      "doesNotProveZh": "长期稳定性、来源可靠性或跨问题偏好。"
    },
    {
      "metricId": "pairwise_hostname_jaccard",
      "nameZh": "两平台主机名 Jaccard 重叠",
      "formula": "size(hostnames_a intersection hostnames_b) / size(hostnames_a union hostnames_b)",
      "range": [
        0,
        1
      ],
      "interpretationZh": "0 表示本样本没有共同精确主机名，1 表示两组精确主机名完全相同。",
      "doesNotProveZh": "两个平台使用相同索引、排序机制或页面级来源。"
    },
    {
      "metricId": "top_hostname_concentration",
      "nameZh": "最高主机名集中度",
      "formula": "largest_hostname_citation_occurrences / all_normalized_url_occurrences",
      "numeratorField": "topHostnameCitationOccurrences",
      "denominatorField": "normalizedUrlOccurrences",
      "interpretationZh": "描述本样本中引用出现是否集中在出现次数最多的精确主机名。",
      "doesNotProveZh": "来源质量、答案准确性或平台长期权重。"
    }
  ],
  "workedSample": {
    "sourceDataset": "https://course.dragonai.tech/datasets/ai-search-rag-finetuning-citation-sample-2026-07.json",
    "queryCount": 1,
    "runsPerProvider": 3,
    "observationWindow": {
      "startedAt": "2026-07-20T23:33:18Z",
      "endedAt": "2026-07-20T23:35:22Z"
    },
    "providerMetrics": [
      {
        "provider": "qwen_dashscope_web_search",
        "model": "qwen-plus",
        "completedRuns": 3,
        "runsWithAtLeastOneConcreteCitationUrl": 3,
        "citationSurfaceRunCoverage": 1,
        "distinctUrls": 10,
        "distinctUrlsObservedInAtLeastTwoRuns": 10,
        "repeatedUrlShare": 1,
        "distinctUrlsObservedInAllThreeRuns": 10,
        "allRunRepeatedUrlShare": 1,
        "normalizedUrlOccurrences": 30,
        "topHostname": "cloud.baidu.com",
        "topHostnameCitationOccurrences": 15,
        "topHostnameConcentration": 0.5
      },
      {
        "provider": "kimi_builtin_web_search",
        "model": "moonshot-v1-128k",
        "completedRuns": 3,
        "runsWithAtLeastOneConcreteCitationUrl": 3,
        "citationSurfaceRunCoverage": 1,
        "distinctUrls": 7,
        "distinctUrlsObservedInAtLeastTwoRuns": 2,
        "repeatedUrlShare": 0.285714,
        "distinctUrlsObservedInAllThreeRuns": 1,
        "allRunRepeatedUrlShare": 0.142857,
        "normalizedUrlOccurrences": 10,
        "topHostname": "arxiv.org",
        "topHostnameCitationOccurrences": 4,
        "topHostnameConcentration": 0.4
      },
      {
        "provider": "gemini_google_search_grounding",
        "model": "gemini-2.5-flash",
        "completedRuns": 3,
        "runsWithAtLeastOneConcreteCitationUrl": 3,
        "citationSurfaceRunCoverage": 1,
        "distinctUrls": 29,
        "distinctUrlsObservedInAtLeastTwoRuns": 6,
        "repeatedUrlShare": 0.206897,
        "distinctUrlsObservedInAllThreeRuns": 1,
        "allRunRepeatedUrlShare": 0.034483,
        "normalizedUrlOccurrences": 36,
        "topHostname": "www.databricks.com",
        "topHostnameCitationOccurrences": 4,
        "topHostnameConcentration": 0.111111,
        "topHostnameTie": [
          "www.databricks.com",
          "medium.com"
        ]
      }
    ],
    "pairwiseHostnameJaccard": [
      {
        "providerA": "qwen_dashscope_web_search",
        "providerB": "kimi_builtin_web_search",
        "hostnamesA": 6,
        "hostnamesB": 5,
        "sharedHostnames": [
          "zhuanlan.zhihu.com"
        ],
        "unionHostnameCount": 10,
        "jaccard": 0.1
      },
      {
        "providerA": "qwen_dashscope_web_search",
        "providerB": "gemini_google_search_grounding",
        "hostnamesA": 6,
        "hostnamesB": 22,
        "sharedHostnames": [],
        "unionHostnameCount": 28,
        "jaccard": 0
      },
      {
        "providerA": "kimi_builtin_web_search",
        "providerB": "gemini_google_search_grounding",
        "hostnamesA": 5,
        "hostnamesB": 22,
        "sharedHostnames": [
          "arxiv.org"
        ],
        "unionHostnameCount": 26,
        "jaccard": 0.038462
      }
    ],
    "hostnameSharedByAllThreeProviders": []
  },
  "comparisonRequirementsZh": [
    "只比较查询、平台、模型、联网设置、地区与账号上下文、运行计划、解析器、URL 归一化和分母政策均可对齐的窗口。",
    "每个比例同时保存分子和分母；失败、拒答、超时、无引用与无法解析必须按预先登记的分母政策处理。",
    "URL 指标和主机名指标分开报告；不同页面不能因为属于同一主机名就合并为同一具体来源。",
    "样本很小时优先报告逐运行集合与精确计数，不输出看似精确的综合平台得分。",
    "模型、产品索引或展示机制变化时新建观察窗口，不静默接入旧趋势。"
  ],
  "limitationsZh": [
    "示例只覆盖一个中文问题、一个短时间窗口和每个平台三次运行，不能估计平台总体分布。",
    "三次重复是暴露波动的操作起点，不构成统计充分性证明。",
    "主机名重叠不等于页面 URL 重叠；引用出现也不证明内容吸收、答案正确或来源可靠。",
    "不同平台展示引用的能力和接口不同，引用数量不能直接作为平台质量排名。",
    "示例中的平台名称、模型和搜索行为可能变化；复测必须记录新版本并检查可比性。"
  ],
  "title": "AI 搜索引用来源稳定性 4 项指标与复算样本 v1",
  "description": "公开引用展示运行覆盖率、重复 URL 占比、主机名 Jaccard 重叠和最高主机名集中度的公式、分子分母、比较要求与 9 次运行示例。",
  "url": "https://course.dragonai.tech/datasets/ai-search-citation-source-stability-metrics-v1.json",
  "creator": "烛龙智元内容研究组"
}