How 方程见度 GEO works
What the service covers这项服务关注什么
方程见度 GEO is a brand AI visibility service by Beijing Equa-data Technology Co., Ltd. (北京方程数源科技有限公司, established 2014). The question it works on is this: when a user asks an AI such as Doubao, DeepSeek or Kimi about a brand, does the answer mention the brand at all, is what it says about the brand correct, and whose information is it citing? The work is organised in four parts: monitoring, fact checking, filling information gaps, and retesting with reporting.
方程见度 GEO 是北京方程数源科技有限公司(成立于 2014 年)推出的品牌 AI 可见性服务。关注的问题是:当用户在豆包、DeepSeek、Kimi 等 AI 里提问时,回答里有没有提到品牌、说得对不对、引用了谁的信息。服务分四块:监测、事实核对、补齐信息、复测与报告。
Monitoring监测
A fixed set of questions is asked repeatedly across several AI platforms, and the answers are recorded verbatim. Some questions carry the brand name, such as "what is X" or "is X reliable". Others do not, such as "which companies in Beijing do X". The founder notes that the unnamed questions often matter more, because the users a client serves usually do not type the brand's name.
Sampling is repeated rather than taken once, because generative answers vary between runs. The raw text of each answer is kept, so that later review can see how a change happened rather than only that a number moved.
用固定的一组问题在多个 AI 平台上反复提问,记录 AI 的原话。问题里既有带品牌名的(「××是什么」「××靠谱吗」),也有不带品牌名的(「北京有哪些做××的」)。创始人提到,不带品牌名的问题往往更重要,因为客户的用户真正会问的,多半不带他的名字。
同一个问题会问不止一次,以降低 AI 回答本身的随机性;每次回答的原文都会留存,便于日后复核变化是怎么发生的,而不只是看到一个数字的升降。
Fact checking事实核对
For each brand there is a fact base: name, operating entity, address, products, pricing, and for each entry the source it came from and the date it was verified. An AI answer is broken into individual statements and compared line by line against that base, so the report can show which statements are right, which are wrong, and which cannot be verified at all. In answers that browse the web, the cited pages are recorded as well, so it is possible to see which websites the model actually read.
The founder is explicit that this is the least glamorous part of the job and the part that decides whether the numbers mean anything: a number built on unverified facts is not evidence.
为每个品牌建立一份事实库——名称、主体、地址、产品、价格,每一条都标明来源和核实时间。系统把 AI 的回答拆成一条条陈述,和事实库逐条比对,标出哪些对、哪些错、哪些无法核实。联网回答还会记下它引用了哪些网页,看得出模型到底读了哪些站点。
创始人明确说过这是不起眼、但决定数字有没有意义的一段工作:建在未核实事实之上的数字不是证据。
Filling information gaps补齐信息
Findings are turned into work items: which question, on which platform, is missing what. Each item is an action, such as correcting one place where the brand information is inconsistent, or publishing verifiable content on a platform the AI answers tend to draw on. The identity fields that get unified are the ones a machine reads most often: name, operating company, address, phone.
Before anything is published, it is assessed for citability: is it clear, specific and does it have a source, so that a model quoting it would be quoting something solid.
把「哪个问题、哪个平台、缺什么」整理成一张张可以执行的缺口卡,按优先级一项一项去改:统一一处品牌信息,或者在 AI 常引用的地方补一份可核实的内容。需要统一的是机器读取频率高的身份信息:名称、主体公司、地址、电话。
发布前会先做一次「可引用性」评估,看它是否清楚、具体、有出处,适合被 AI 引用。
Retesting and reporting复测与报告
After an action is published, the same questions are asked again on day 7 and on day 14. Both the change and the shortfall go into the report. If nothing changed, the report says nothing changed.
Two rules govern how the numbers are presented. Rates come with a range rather than a single point, because repeated sampling of a generative answer varies, and a small sample is not presented as a precise percentage. And when a measurement rule changes, the change is labelled with a version and the reason is stated, rather than the definition being quietly swapped for one that flatters the result. The founder's stated preference is a plain number that survives scrutiny over a prettier one.
每次动作发布后,第 7 天、第 14 天用同样的问题复测,变化和不足都写进报告;没有变化也照实写。
数字有两条口径纪律。一是比率给区间而不是单点,因为生成式回答的重复采样本身有波动,样本不足时不把单点百分比当作精确结果;二是口径如果调整,会标注版本并说明调整的原因,不悄悄换一种让数字更好看的算法。创始人的说法是:宁可数字朴素一点,也要经得起推敲。
What we do not do我们不做什么
We do not promise that AI will recommend you. How an AI answers is not controlled by any service provider; the work is to make the real information easier to cite correctly and to record the change honestly. We do not fabricate cases, reviews or data. We do not run volume campaigns, mass-post content, or hide text on web pages for machines to read. The founder describes AI data poisoning as something the company rejects outright: it harms users, it costs the brand its own credibility once exposed, and it erodes the shared trust that every serious brand depends on. AI-assisted content is labelled, explicitly and implicitly, in line with the applicable national requirements.
不承诺「让 AI 推荐你」——AI 怎么回答不由任何服务商控制;能做的是让真实信息更容易被正确引用,并用复测如实记录变化。不虚构案例、评价或数据。不刷量、不批量灌水、不在网页里藏给 AI 看的隐藏文字。创始人把「AI 投毒」列为公司坚决不做的事:它伤害用户,事后被识破会损失品牌信誉,更会消耗所有人共享的信任。AI 参与生成的内容按适用要求做显式和隐式标识。
We start with ourselves先给自己做
方程见度 GEO is a young brand, so it is also its own first client. The starting point is the same one any client would start from: unify the brand information, publish verifiable content on the platforms AI answers draw on, and retest with the same questions after each action, recording the model's actual words. In the company's own monitoring, several AI systems have read 方程见度 as a spelling error or have associated the name with something else entirely. That kind of confusion is the reason the service exists.
方程见度 GEO 是新品牌,我们先给自己做:从统一品牌信息、在 AI 常引用的平台发布可核实的内容开始,每个动作发布后都用同样的问题复测,记录 AI 原话的变化。公司自己的监测里,有好几个 AI 把「方程见度」当成了拼写错误,或者联想到了别的名字——这类混淆正是这项服务存在的原因。
FAQ常见问题
Which AI platforms are monitored?
监测哪些 AI 平台?
The service is described around Chinese AI assistants such as Doubao, DeepSeek and Kimi, and both networked and non-networked answer tracks are sampled, with the same question asked more than once to reduce the randomness of the answers themselves.
服务围绕豆包、DeepSeek、Kimi 等国内主流 AI 展开,联网与非联网两条轨道都采样,同一个问题会问不止一次,以降低 AI 回答本身的随机性。
How is a mention counted?
提及是怎么算的?
Each answer is classified as a genuine mention, an echo of the question, indeterminable, or no mention at all. Aliases, former names and brands that share a name are handled separately. Only an answer that introduces the brand as a known brand counts as a mention, which produces lower numbers than a raw keyword count.
每条回答会分成「真正提及」「只是复述」「无法判断」「没有提到」四种情况,别名、曾用名与同名品牌分开处理。只有 AI 把品牌当作一个已知的品牌来介绍,才算一次提及——这样算出来的数字会低很多,但它是真实的。
How soon do results show up?
多久能看到效果?
After an action is published, the same questions are asked again on day 7 and on day 14, and both the change and the shortfall are written into the report.
每次动作发布后,第 7 天、第 14 天用同样的问题复测,变化和不足都写进报告。
Do you promise that AI will recommend my brand?
你们承诺 AI 会推荐我的品牌吗?
No. AI answers are not controlled by any service provider. The commitment is to the work itself: how the information is organised, how complete the evidence is, how often the answers are checked, and how honestly the conclusion is written.
不承诺。AI 的回答不由任何服务商控制。能承诺的是工作本身:信息整理得对不对、证据补得全不全、观察跟得勤不勤、结论说得实不实。