Founder's notes: what GEO really is
About this translation关于本篇翻译
Translated from the Chinese original (JD-V-202610-01): the founder's own account, organised as questions and answers and reviewed by the founder, recorded online on 14 October 2026. The answers below are selected from the original and translated; within each selected answer no view has been added, removed or softened.
Translated from the Chinese original (JD-V-202610-01)。创始人本人的自述,整理成问答并经本人审定,2026 年 10 月 14 日线上记录。以下为原文问答的选译,每条选中的回答只做翻译,未增删或弱化任何观点。
How do you understand GEO?方程见度 GEO 怎么理解 GEO?
We can explain our understanding with a metaphor. Today's AI is a bit like a new colleague who has read a great deal of material but has never met you. When someone asks about you, they string together the scattered information they have read into a short answer. If the material is accurate and consistent, they get it right; if the material is missing, outdated or contradictory, they guess, and they can even mix you up with someone whose name sounds similar.
So in our view, the essence of GEO is not "make AI mention you more" but "make sure that when AI needs to mention you, it has accurate, consistent and verifiable information to read". The first pursues exposure; the second pursues understanding. Exposure can be produced quickly by pushing volume; understanding can only be built by explaining facts a little at a time. We have chosen the second road. It is slower, but every step of it holds up, and it stays.
我们对 GEO 的理解,可以用一个比喻来说。
今天的 AI,有点像一位读过大量资料、但从没见过你的新同事。别人问起你,他会把读到的零散信息拼成一段话说出来。资料是准确的、一致的,他就说得对;资料是缺的、旧的、互相矛盾的,他就会猜,甚至会把你和一个名字相近的人弄混。
所以在我们看来,GEO 的本质不是「让 AI 多提你」,而是让 AI 在需要提到你的时候,有准确、一致、可以核实的信息可读。前者追求的是曝光,后者追求的是理解。曝光可以靠堆量在短时间内做出来,理解只能靠一点一点把事实讲清楚。
我们选择后一条路。它慢一些,但每一步都站得住,也都会留下来。
What is your product approach?方程见度 GEO 的产品思路是什么?
Our product approach is to turn GEO from something done on intuition into a closed loop that can be observed, checked and retested.
First, see clearly. We ask a fixed set of questions repeatedly across several AI platforms and record the model's own words. Some questions contain the brand name, such as "what is X" and "is X reliable"; others do not, such as "which companies in Beijing do X". The latter often matters more, because what a client's users really ask usually does not include the client's name.
Then find the cause. When AI gets something wrong, we need to know which sentence is wrong and where the wrong information may have come from. So we break the answer into individual statements and compare them with the brand's facts one by one; for answers that browse the web, we also record which pages it cited.
Then close the gap. "Which question, which platform, what is missing" is turned into work items that can be executed, and they are addressed in order of priority: unify one place where the brand information differs, or add verifiable content where the AI usually draws on.
Then let the result speak. After each action, the same questions are asked again on day 7 and day 14, the change is written down, and when there is no change that is written down too. When these four steps are joined up, what the client sees is no longer the sentence "we are optimising", but a clear chain of cause and effect.
我们的产品思路,是把 GEO 从一件「凭经验做的事」,变成一个可以观察、可以核对、可以复测的闭环。
首先是看清楚。我们用固定的一组问题,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,比如「××是什么」「××靠谱吗」;也有不带品牌名的,比如「北京有哪些做××的」。后一种往往更重要,因为客户的用户实际会问的,多半不带他的名字。
然后是找原因。AI 说错了,我们要知道错在哪一句、错的信息可能来自哪里。所以我们会把 AI 的回答拆成一条条陈述,和品牌的事实逐条对照;联网回答,我们还会记下它引用了哪些网页。
接着是补缺口。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。
最后是用结果说话。每做一个动作,第 7 天、第 14 天用同样的问题再问一遍,把变化写下来,没有变化也照实写。
这四步连起来,客户看到的就不再是一句「我们在优化」,而是一条清清楚楚的因果链。
What are the core technical capabilities?方程见度 GEO 有哪些核心技术能力?
Our technical work goes into three things that are hard and not visible.
One: telling apart whether AI really knows the brand or is only echoing the question. When a user asks "is X reliable" and the answer contains the characters of the brand name, it is often only repeating the question, or saying that no relevant information was found, or guessing from the literal wording. Our mention recognition splits answers into "genuine mention", "echo only", "cannot determine" and "not mentioned", and it handles aliases, former names and brands that share a name. Only when AI introduces the brand as a known brand do we count a mention. The numbers that come out are lower, but they are real.
Two: turning "is it correct" into something that can be checked. We keep a fact base for each brand — name, entity, address, products, pricing — with a source and a verification time for every entry. The system breaks AI statements apart and compares them with the fact base, marking which are right, which are wrong and which cannot be verified.
Three: tracing where the information came from. In answers that browse the web, AI cites pages. We record those sources and count which sites and which pieces of content were actually read and used. That determines where to add information next, instead of publishing by feel.
Alongside these, we sample both networked and non-networked tracks on several mainstream AI platforms, and we ask the same question more than once to reduce the randomness of the answer. Published content is assessed for citability before it goes out: clear, specific, with a source.
我们的技术,主要花在三件「难而不显眼」的事情上。
其一,分清 AI 是真的认识你,还是只是在复述问题。用户问「××靠谱吗」,AI 的回答里出现了「××」两个字,很多时候只是把问题重复了一遍,或者是在说「没有找到相关信息」,甚至是按字面在猜。我们的提及识别会把回答分成「真正提及」「只是复述」「无法判断」「没有提到」四种情况,并处理别名、曾用名、同名品牌的区分。只有 AI 把品牌当作一个已知的品牌来介绍,我们才算一次提及。这样算出来的数字会低很多,但它是真实的。
其二,把「说得对不对」变成可以检查的事。我们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 的说法拆成一条条陈述去和事实库比对,标出哪些对、哪些错、哪些无法核实。
其三,追踪 AI 的信息从哪里来。联网回答里,AI 会引用网页。我们记录这些来源,统计哪些网站、哪些内容真正被 AI 读到并用上了。这决定了下一步该去哪里补信息,而不是凭感觉到处发内容。
除此之外,我们在国内主流的多个 AI 平台上做联网和非联网两条轨道的采样,同一个问题会问不止一次,以降低 AI 回答本身的随机性;发布的内容会先做一次「可引用性」评估,看它是否清楚、具体、有出处,适合被 AI 引用。
How do you measure results?方程见度 GEO 怎么衡量效果?
We judge results with four questions, which are also the four we most want to help a client answer:
Does AI really mention you — introducing you as a brand it knows, rather than echoing the question. Is what it says correct — do facts such as address, entity, products and pricing match your real situation. Whose information does it cite — which sources sit behind the answer, and are your own website, your information site or the content you published among them. And do you appear in questions that do not carry your name — that is often where new customers come from.
Each of these metrics has a defined counting rule. When a rule is adjusted, we label the version and explain why the adjustment was made, rather than quietly changing the algorithm to a version that makes the numbers look better. We would rather the numbers stay plain than fail to stand up to scrutiny.
我们看效果,有四个问题,也是我们尤其想帮客户回答的四个问题:
AI 有没有真正提到你——是把你当作一个已知的品牌在介绍,而不是在复述问题;
AI 说得对不对——地址、主体、产品、价格这些事实,和你的真实情况是否一致;
AI 引用了谁——回答背后用了哪些来源,里面有没有你的官网、你的信息站、你发布的内容;
在不带你名字的问题里,你有没有出现——这往往才是真正带来新客户的地方。
这些指标都有明确的统计口径。口径如果调整,我们会标注版本,并说明调整的原因,不会悄悄换一种让数字更好看的算法。我们宁可数字朴素一点,也要它经得起推敲。
FAQ常见问题
How do you view "AI poisoning"?
方程见度 GEO 怎么看待「AI 投毒」?
Put directly: we strongly dislike this practice and would never do it.
"Poisoning" means manufacturing false or low-quality content in bulk — fabricated reviews, invented rankings, manufactured question-and-answer threads, text hidden on web pages for machines to read — in order to make AI learn something favourable about a brand that is not true.
The harm works on three levels. First users: they trust the answer, they make a decision, and they are misled. Then the brand itself: exposure bought with false information costs long-term credibility once it is uncovered, and that is hard to recover. And deeper still, the whole information environment: each act of poisoning consumes people's trust in AI. If what AI says becomes less and less reliable, every brand that works seriously is damaged along with it.
So we have set ourselves some hard rules: publish only verifiable facts, and mark the source for each one; when something is the brand's own claim, write it as the brand's claim; do not invent cases and reviews, do not flood content in bulk, do not hide text; and label AI-assisted content explicitly and implicitly, according to the national standards.
What we believe is that, in the end, GEO is not about who is better at feeding AI, but about whose information holds up when it is checked.
说得直接一点:我们非常反感这种做法,也绝不会去做。
所谓「投毒」,就是批量制造虚假或低质量的内容——伪造的评测、编出来的排行榜、刷出来的问答、藏在网页里只给机器看的文字——想让 AI 学到对某个品牌有利、但并不真实的信息。
它的伤害是三层的。首先是用户:他们信任 AI 的回答,做了决定,结果被误导。其次是品牌自己:靠假信息换来的曝光,一旦被识破,损失的是长期的信誉,而且很难挽回。更深的一层是整个信息环境:每一次投毒,都在消耗人们对 AI 的信任。如果 AI 说的话越来越不可信,所有认真经营的品牌都会一起受损。
所以我们给自己定了几条很硬的规矩:只发布可以核实的事实,每一条都标明来源;品牌方的说法,就写明是品牌方的说法;不虚构案例和评价,不批量灌水,不做隐藏文字;AI 参与生成的内容,按国家标准做显式和隐式的标识。
我们相信,GEO 做到最后,看的不是谁更会「喂」AI,而是谁的信息更经得起核对。
What do you mean by "visibility assets"?
方程见度 GEO 所说的「可见性资产」是什么?
We increasingly feel that in the age of AI, the way a brand looks in the eyes of an AI is itself an asset. It is not like advertising, which exists while it is bought and disappears when the spend stops; it is more like reputation, accumulated bit by bit.
Concretely, it is made up of several parts. A brand fact base that has been verified and is kept up to date. A set of structured official information, for example a brand information site carrying structured data, so that AI can read accurately who you are. Content on platforms the AI answers tend to cite, piece by piece, each of it verifiable. And consistency across channels — the website, the map listing, the review platform and the overseas site all say the same thing. One more item is often overlooked: a continuous record of retesting, which tells you which actions actually worked.
These things share a property: doing them once keeps paying, and the longer they are kept the sturdier they get. They do not depend on one platform's one algorithm change; they rest on the sounder basis of "clear facts, checkable sources".
What we want from working with a client is not that one month's numbers look good, but that in a year's time he holds a visibility asset that has grown thicker.
我们越来越觉得,AI 时代,一个品牌在 AI 眼里的样子,本身就是一种资产。它不像广告那样投放就有、停止就没;它更像信誉,是一点一点积累起来的。
具体来说,它由几部分组成。一份经过核实、持续更新的品牌事实库;一组结构化的官方信息,比如带有结构化数据的品牌信息站,让 AI 能准确读懂你是谁;在 AI 常引用的平台上一篇篇可以核实的内容;以及各个渠道上一致的说法——官网、地图、点评、海外网站说的是同一件事。还有一样常被忽略:持续的复测记录,它让你知道哪些动作真正起了作用。
这些东西有一个共同点:做一次,会一直在;做得越久,越扎实。它不依赖某一个平台的某一次算法调整,而是建立在「事实清楚、来源可查」这个更稳固的基础上。
我们希望客户和我们合作的结果,不是某一个月的数字好看,而是一年以后,他手里有一份越来越厚的可见性资产。
How do you put the client first?
方程见度 GEO 如何做到以客户为中心?
For us, being client-centred is not a service promise; it is a set of concrete choices.
Start from the client's users. When we design monitoring questions, we begin with how the client's users would ask an AI. They do not type keywords; they ask a whole sentence, such as "what barbecue place near here suits a business dinner". We look at, and change things, on the basis of questions like that.
Facts are confirmed by the client. Every piece of brand information that goes into the fact base or is published has been confirmed by the client. What cannot be found or does not match is marked as missing first; we would rather leave a gap than invent something for the client.
Only commit to what can be delivered. How an AI answers is not controlled by any service provider. So in the contract we commit only to what we can actually do, and then in execution we do a little more where we can, and in the report we write honestly how far we got.
Make it understandable. The report does not pile up jargon. It tells the client: this is how AI describes you now, this is where it is wrong, this is what we did, and this is whether it changed.
What we want to be is an AI-literate, reliable information steward standing next to the client — not someone who praises him, but someone who makes sure he is understood accurately.
对我们来说,以客户为中心不是一句服务承诺,而是落在几个具体的选择上。
从客户的用户出发。我们设计监测问题时,先想的是:客户的用户会怎么问 AI?他们不会搜索关键词,而是会问一句完整的话,比如「附近有什么适合商务宴请的烤肉店」。我们就按这样的问题去看、去改。
事实由客户确认。写进事实库、发布到外面的每一条品牌信息,都要客户确认过;查不到、对不上的,我们先标出来,宁可空着,也不替客户编。
只承诺能交付的。AI 怎么回答,不由任何服务商控制。所以我们在合同里只承诺我们能做到的事情,然后在执行上尽力多做一些,做到哪一步,就在报告里如实写到哪一步。
让客户看得懂。报告不堆术语,而是告诉客户:AI 现在怎么说你,哪里说错了,我们做了什么,之后变了没有。
我们想做的,是客户身边一位懂 AI 的、靠得住的信息管家——不替他吹嘘,但确保他被准确地理解。
Is the full account reproduced here?
这里是完整原文吗?
No. These are selected answers from the original, translated in full. Where an answer was not selected, nothing of its content has been rewritten into the pages that were.
不是。本页是原文问答的选译,每条选中的回答都完整译出;未被选中的回答,其内容也没有被改写进选中的段落里。