[{"data":1,"prerenderedAt":390},["ShallowReactive",2],{"blog-brand-corpus-audit":3,"blog-prev-brand-corpus-audit":212,"blog-next-brand-corpus-audit":313},{"id":4,"title":5,"author":6,"body":7,"canonical":196,"category":197,"cover":196,"date":198,"description":199,"draft":200,"extension":201,"keywords":196,"meta":202,"navigation":203,"noindex":200,"ogImage":196,"path":204,"seo":205,"seoDescription":196,"seoTitle":196,"stem":206,"tags":207,"updated":196,"__hash__":211},"blog\u002Fblog\u002Fbrand-corpus-audit.md","品牌语料盘点：五个客户平均 240 个来源，官网只占 30 个","元析智能",{"type":8,"value":9,"toc":185},"minimark",[10,14,18,21,24,27,30,33,36,39,128,131,134,137,140,143,146,149,152,155,158,161,164,167,170,173,176,179,182],[11,12,13],"p",{},"AI 认识你，靠的是 240 个页面，其中只有 30 个是你写的。2026 年 3 月我们给五个 B2B 软件客户做了语料盘点，把四个引擎在品牌词和产品词问答里挂出来的来源全部列出来、去重、分类。平均 240 个来源，自家官网占 30 个，12.5%。",[15,16,17],"h2",{"id":17},"盘点要回答三个问题",[11,19,20],{},"它读了什么：哪些页面在给 AI 提供关于你的信息。\n它读错了什么：哪些页面的描述和事实不符。\n它缺了什么：哪些信息本该有、但一个来源都没有。",[11,22,23],{},"这三个问题比\"我们被引用了几次\"有用得多。次数是结果，语料是原因。",[15,25,26],{"id":26},"怎么收集来源",[11,28,29],{},"四个引擎，每个跑 20 个查询：10 个品牌词，5 个产品词，5 个\"品牌加场景\"。",[11,31,32],{},"每个回答记录两样东西：挂出来的域名，以及它描述你的原话。原话必须抄，不能只记域名。",[11,34,35],{},"五个客户、400 次问答，去重后拿到 1180 个来源域名，平均每个客户 236 个。顺手还收到 300 多条描述原话，这批原话是后面判断对错的依据。",[15,37,38],{"id":38},"平均结构长什么样",[40,41,42,58],"table",{},[43,44,45],"thead",{},[46,47,48,52,55],"tr",{},[49,50,51],"th",{},"来源类型",[49,53,54],{},"平均数量",[49,56,57],{},"占比",[59,60,61,73,84,95,106,117],"tbody",{},[46,62,63,67,70],{},[64,65,66],"td",{},"官网",[64,68,69],{},"30",[64,71,72],{},"13%",[46,74,75,78,81],{},[64,76,77],{},"平台内容（知乎、公众号、掘金、CSDN）",[64,79,80],{},"96",[64,82,83],{},"40%",[46,85,86,89,92],{},[64,87,88],{},"媒体与自媒体",[64,90,91],{},"42",[64,93,94],{},"18%",[46,96,97,100,103],{},[64,98,99],{},"目录站与工具导航",[64,101,102],{},"38",[64,104,105],{},"16%",[46,107,108,111,114],{},[64,109,110],{},"GitHub 与技术社区",[64,112,113],{},"22",[64,115,116],{},"9%",[46,118,119,122,125],{},[64,120,121],{},"其他",[64,123,124],{},"10",[64,126,127],{},"4%",[11,129,130],{},"平台内容占了四成，这个数字每次都让客户惊讶。他们一直以为官网是主战场，实际主战场在别人的地盘上。",[15,132,133],{"id":133},"读错了什么",[11,135,136],{},"五个客户里有三个存在明显错误描述。",[11,138,139],{},"一个被写成\"仅支持私有部署\"，实际有 SaaS 版。错误来源是 2023 年的一篇媒体报道，被三个目录站转载。",[11,141,142],{},"一个被写成\"开源软件\"，实际只有 SDK 开源。来源是 GitHub 上一个第三方项目的 README。",[11,144,145],{},"一个被写成\"总公司在深圳\"，实际在成都。来源是某招聘网站的旧信息。",[11,147,148],{},"三条错误的共同点：都不是官网写的，但都被 AI 当成事实，因为多个来源重复了同一个说法。",[15,150,151],{"id":151},"缺了什么",[11,153,154],{},"缺口集中在三块：中文场景内容、新版本文档、客户案例的量化结果。",[11,156,157],{},"新版本文档缺得最要命。一个客户 2025 年上线了三个新功能，文档站更新了，但中文语料里一个字都没有。结果 AI 回答\"你们支持向量检索吗\"时，说\"暂不支持\"。",[11,159,160],{},"这条错误每天在替他拒绝客户。",[15,162,163],{"id":163},"盘点之后做什么",[11,165,166],{},"三类动作，按优先级：",[11,168,169],{},"改错的。找到错误来源页，发更正请求。改不动的，就在官网和平台写一版更清楚的说法，让正确信息在数量上压过错误信息。",[11,171,172],{},"补缺的。缺口里最影响成单的先补，通常是新版本文档和部署方式。",[11,174,175],{},"建台账的。把 240 个来源记成一张表，标注类型、是否准确、最后检查日期。每半年复查一次。",[15,177,178],{"id":178},"一个判断",[11,180,181],{},"语料盘点比关键词研究有用。关键词研究告诉你别人搜什么，语料盘点告诉你 AI 以为你是谁。后者错了，前者做得再好也没用。",[11,183,184],{},"这活儿一次四个小时，五个人工。不做，你就一直在猜。",{"title":186,"searchDepth":187,"depth":187,"links":188},"",2,[189,190,191,192,193,194,195],{"id":17,"depth":187,"text":17},{"id":26,"depth":187,"text":26},{"id":38,"depth":187,"text":38},{"id":133,"depth":187,"text":133},{"id":151,"depth":187,"text":151},{"id":163,"depth":187,"text":163},{"id":178,"depth":187,"text":178},null,"GEO","2026-03-30","我们把 AI 能读到的关于品牌的页面全列出来，五个客户平均 240 个来源，自家官网只有 30 个。盘点要回答三个问题：它读了什么、读错了什么、缺了什么。",false,"md",{},true,"\u002Fblog\u002Fbrand-corpus-audit",{"title":5,"description":199},"blog\u002Fbrand-corpus-audit",[208,209,197,210],"语料","品牌实体","盘点","-XSdAAZ8qz0gjHInPI5tuRSI7rtw_Kmek-os-70Xwgs",{"id":213,"title":214,"author":6,"body":215,"canonical":196,"category":197,"cover":196,"date":303,"description":304,"draft":200,"extension":201,"keywords":196,"meta":305,"navigation":203,"noindex":200,"ogImage":196,"path":306,"seo":307,"seoDescription":196,"seoTitle":196,"stem":308,"tags":309,"updated":196,"__hash__":312},"blog\u002Fblog\u002Fai-citation-tracking.md","AI 引用怎么统计：我们手工查了 60 个问题四个月",{"type":8,"value":216,"toc":293},[217,220,224,227,230,233,236,239,242,245,248,251,254,257,260,263,266,269,272,275,278,281,284,287,290],[11,218,219],{},"AI 引用没有现成报表，只能手工查。我们从 2025 年 12 月起固定 60 个问题，每月丢给 ChatGPT、Perplexity 和国内两三个助手，记录回答里有没有引用我们的页面。四个月下来，数字不大，但规律很清楚。",[15,221,223],{"id":222},"_60-个问题怎么定","60 个问题怎么定",[11,225,226],{},"20 个是品类问题，20 个是具体技术问题，20 个是竞品对比问题。每类各留 5 个作为对照，不针对我们任何内容做优化。",[11,228,229],{},"问题要选用户真会问的，词表里意图明确的。别选“元析智能怎么样”这种品牌词，那不是测引用，是测自己。",[15,231,232],{"id":232},"查的时候要固定条件",[11,234,235],{},"同一个账号级别、同一个地区节点、同一个问法。条件变了，结果可比性就没了。",[11,237,238],{},"我们一开始没固定，一个月后换了个节点，同一批问题引用次数从 11 次掉到 4 次，查了半天才发现是环境问题。",[15,240,241],{"id":241},"记什么",[11,243,244],{},"记三件事：有没有引用，引用的是哪个页面，回答里有没有提到我们的品牌名。",[11,246,247],{},"有引用但没提品牌名，价值只有一半。用户不知道这个答案是你给的。",[15,249,250],{"id":250},"四个月的粗糙结论",[11,252,253],{},"技术问题的引用率最高，20 个问题里最多一个月有 9 个引用了我们的文档页。品类问题最低，20 个里只有 2 到 3 个。",[11,255,256],{},"竞品对比问题最不稳定，因为我们和竞品的页面都在迭代，AI 引用哪边经常换。",[15,258,259],{"id":259},"手工统计的用处",[11,261,262],{},"它不精确，但不能省。你只有亲自看回答，才知道 AI 是怎么摘你的内容的：它摘的是不是那句结论，有没有断章取义，有没有拼错。",[11,264,265],{},"2026 年 2 月我们发现一个问题，AI 引用我们的页面时把适用条件漏了，回答变成“所有规模都适用”。后来我们在那一句前面补了限制条件，下个月就正常了。",[15,267,268],{"id":268},"要不要上工具",[11,270,271],{},"有工具能监控品牌在 AI 回答里的提及，可以补数量上的盲区。但工具看不到“摘得对不对”，这件事还得人工。",[15,273,274],{"id":274},"引用率的三个层次",[11,276,277],{},"第一层是“被提到”，回答里出现我们的品牌名但没有链接。第二层是“被引用”，回答里带我们的页面作为来源。第三层是“被采纳”，回答里的结论和我们的结论一致。",[11,279,280],{},"大多数公司只盯着第二层。实际上第一层更基础，品牌名在 AI 回答里出现多了，用户会自己去搜你。",[15,282,283],{"id":283},"别为了引用率乱改内容",[11,285,286],{},"见过有人为了让 AI 更容易摘，把页面写成一堆短问句，读起来像机器翻译的 FAQ。",[11,288,289],{},"这类页面短期可能被多摘几次，长期会失去人。AI 引用是结果，不是目标。",[11,291,292],{},"我们的做法是工具管量，人工管质。每月 40 分钟，够用。",{"title":186,"searchDepth":187,"depth":187,"links":294},[295,296,297,298,299,300,301,302],{"id":222,"depth":187,"text":223},{"id":232,"depth":187,"text":232},{"id":241,"depth":187,"text":241},{"id":250,"depth":187,"text":250},{"id":259,"depth":187,"text":259},{"id":268,"depth":187,"text":268},{"id":274,"depth":187,"text":274},{"id":283,"depth":187,"text":283},"2026-03-29","AI 引用没有现成报表，只能手工查。我们固定 60 个问题，每月丢给 ChatGPT、Perplexity 和国内的几个助手，记录有没有引用我们的页面。四个月下来，规律比数字更有用。",{},"\u002Fblog\u002Fai-citation-tracking",{"title":214,"description":304},"blog\u002Fai-citation-tracking",[310,197,311],"AI引用","效果评估","iqUymEig-akAfFBXMkqCCfKID2vC9Gz-jcunljj-NL4",{"id":314,"title":315,"author":6,"body":316,"canonical":196,"category":378,"cover":196,"date":379,"description":380,"draft":200,"extension":201,"keywords":196,"meta":381,"navigation":203,"noindex":200,"ogImage":196,"path":382,"seo":383,"seoDescription":196,"seoTitle":196,"stem":384,"tags":385,"updated":196,"__hash__":389},"blog\u002Fblog\u002Fchina-site-trust-signals.md","中国站点信任信号清单：备案、工商信息和真人",{"type":8,"value":317,"toc":373},[318,321,324,331,337,343,349,352,355,358,361,364,367,370],[11,319,320],{},"国内用户判断一个软件站可不可信，看的是几个具体信号，跟设计好不好看关系不大。2025 年到现在我们给 40 多个国内软件站做过信任页评审，四项基础信息不全的占了六成。",[15,322,323],{"id":323},"四项必备",[11,325,326,330],{},[327,328,329],"strong",{},"ICP 备案号。"," 放在页脚，链接到工信部备案查询页。查法很简单，去工信部系统输入域名，能查到备案主体和备案号。页面上的号必须和查询结果一致，过期、注销、写成别家公司的，都算不合格。",[11,332,333,336],{},[327,334,335],{},"公安备案。"," 通常和 ICP 并列放在页脚，带公安图标和编号。很多站只做了 ICP 没做公安，严格来说是不完整的。",[11,338,339,342],{},[327,340,341],{},"工商信息。"," 公司全称、统一社会信用代码、注册地址。至少要在关于页能查到。用户真要核实，会去国家企业信用信息公示系统搜。",[11,344,345,348],{},[327,346,347],{},"公开联系方式。"," 一个能打通的电话，一个能到门牌号的地址。这两项看着简单，实际最常出问题：电话是空号、地址只写到城市、留的是虚拟办公地址。",[15,350,351],{"id":351},"三个加分项",[11,353,354],{},"真实的办公照片或短视频，比任何文字都管用。团队页放真人姓名和职位，别用卡通插画顶替。客户 logo 墙要拿到授权，没授权就写行业和规模，别放 logo。",[15,356,357],{"id":357},"常见错误",[11,359,360],{},"最常见的是备案主体和运营主体不一致。2025 年 8 月我们评审一个客户，备案主体是个人，运营方是公司，企业客户在采购流程里直接提出质疑，最后花了两周重新备案。",[11,362,363],{},"第二常见的是电话打不通。我们随机拨了 20 个客户的公开电话，有 3 个是空号，2 个转到了语音信箱且没回电。",[11,365,366],{},"第三是地址变更没同步。公司搬了家，页脚还是老地址，客户按地址找过去是别家公司，这种信任损失很难补。",[11,368,369],{},"还有一个隐性问题是备案号被放在图片里。有些站把备案号做成一张小图，看着整齐，但用户和工具都读不到，也点不进工信部查询页。改成文字加链接，两分钟的事。",[11,371,372],{},"这份清单两个小时内能查完，但六成站点不及格。都是些小事，只是没人认真做。我们现在的做法是把这张清单放进新站上线的检查表，逐项有人签字，免得拖到被客户问起才补。",{"title":186,"searchDepth":187,"depth":187,"links":374},[375,376,377],{"id":323,"depth":187,"text":323},{"id":351,"depth":187,"text":351},{"id":357,"depth":187,"text":357},"技术SEO","2026-03-31","我们给 40 多个国内软件站做过信任页评审，ICP 备案、公安备案、工商信息、公开电话四项不全的占六成。这篇是一份可逐项打勾的清单，每项都写了怎么查、查到什么算过关。",{},"\u002Fblog\u002Fchina-site-trust-signals",{"title":315,"description":380},"blog\u002Fchina-site-trust-signals",[386,387,388],"备案","工商信息","站点可信度","4cV3lE2YBlcvujmiNILjk_XLfAb2vWv0iBPyy6JjASw",1794088107195]