What is GEO (Generative Engine Optimization)?
What GEO isGEO 是什么
Generative Engine Optimization (GEO) is the practice of improving how content shows up in answers written by generative engines, such as AI search features and conversational assistants. The aim is that an AI system can find a piece of content, understand it and cite it accurately, when that content is factual, structured and traceable to a source. The concept was set out systematically by Aggarwal and colleagues in the paper "GEO: Generative Engine Optimization", published at KDD 2024. The paper measures visibility inside generated answers, for example the share of words attributed to a source weighted by where the citation appears, and it introduces the GEO-bench evaluation set of multi-domain questions. In the paper's own experimental setup, some optimization methods raised content visibility in generated answers by up to roughly 40 percent, and the methods that worked were not the same across domains.
This is a field without a settled standard. In China, generated and synthesized content also has to follow the labelling requirements of the Measures for Labelling AI-Generated Synthetic Content and the accompanying mandatory national standard GB 45438-2025, in force since 1 September 2025.
生成引擎优化(Generative Engine Optimization,GEO)是面向 AI 生成引擎(AI 搜索、对话助手等)优化内容可见性的实践:让 AI 在生成回答时能找到、理解并准确引用事实性、结构化、可溯源的内容。这一概念由 Aggarwal 等人在论文《GEO: Generative Engine Optimization》中系统提出,论文发表于 KDD 2024。论文给出可见度指标,例如按回答中引用位置加权的字数占比,并构建了包含多领域问题的评测集 GEO-bench;论文报告在其实验设置下,部分优化方法可将内容在生成回答中的可见度提升最多约 40%,且不同领域有效的方法并不相同。
GEO 目前没有统一的国家或国际标准。在中国,AI 生成的内容还需遵守《人工智能生成合成内容标识办法》与配套的强制性国家标准 GB 45438-2025 的标识要求(2025 年 9 月 1 日起施行)。
How a generative answer is built生成式回答是怎么产生的
Answering a question takes roughly two steps. The system retrieves pages or passages that look relevant, then a large language model synthesises an answer from that material and marks its citations. Whether your content enters the answer depends on three things: whether it can be retrieved at all, whether the model can understand it, and whether it looks more credible and more citable than the competing candidate sources.
This changes the object of the work. Classic search returns a column of links, and the user decides by position in that list. A generative engine writes a paragraph instead. So the question a brand has to answer is no longer "which page ranks where" but: is the content adopted, how much of the answer does it account for, and is it paraphrased accurately.
生成式引擎回答一个问题,大致经过两步:先检索与问题相关的网页或段落,再由模型基于这些材料生成答案,并在其中标注引用。内容能否进入回答,取决于它能否被检索到、能否被模型理解,以及在多个候选来源中是否更可信、更好引用。
这改变了工作的对象。传统搜索返回一列网页链接,用户的判断由链接位置决定;生成式引擎写出的则是一段话。于是品牌要回答的问题不再是「哪个网页排在第几位」,而是:内容会不会被采纳、在回答里占多大分量、是否被准确转述。
How GEO differs from SEOGEO 与 SEO 的区别
Search engine optimisation works on the assumption that the result page is a list. GEO works on the assumption that the result is a passage written by a model. The two overlap heavily in practice, and most AI search still retrieves from the web, so a page has to be crawlable and indexed before it can be cited at all. Google Search Central states that appearing in its AI features requires no additional technical requirements, and that ordinary search optimisation practices still apply.
Building on that SEO baseline, GEO adds a set of questions about structure and evidence: is the content organised so a model can locate the key statement, does it carry verifiable data and a stated source, are the facts about the brand consistent everywhere they appear, and is the page marked up with structured data that identifies the entity. The llms.txt proposal, put forward by Jeremy Howard in September 2024, is one more format-level idea of this kind: a Markdown file at the root of a site that tells language models and AI agents what the site is and where its key pages are. It is a community proposal, not a standard, and it does not control crawling.
传统搜索的可见性体现在链接列表中的位置,答案引擎的可见性体现在回答正文里是否出现、被如何描述、被哪些来源支撑。两者在实践上大量重叠:多数 AI 搜索仍以网页检索为基础,内容需要先能被抓取、被收录,才谈得上被引用。Google Search Central 关于 AI 功能的说明指出,出现在 AI Overviews 等功能中没有额外的技术要求,常规的搜索优化通行做法依然适用。
在此基础上,GEO 还关心结构与证据:用清晰的定义句和层级结构组织内容,便于模型定位关键陈述;在正文中给出可核实的数据和出处;保持品牌名称、主体、地址等事实在各处一致;用结构化数据标注页面上的实体。llms.txt(Jeremy Howard 于 2024 年 9 月提出)也属于这一类格式层面的提案:在网站根目录放一个 Markdown 文件,向大语言模型和 AI 代理说明站点是什么、关键页面在哪里。它是社区提案,不是标准,也不控制抓取。
Why a brand should care品牌为什么要关心
The starting point is that a brand has to be recognised before it can be described. If public information about a brand is scattered, contradictory or missing, systems cannot tell it apart from a similarly named entity, and the answer a user receives can attach the wrong description to it. This is the problem that brand entity disambiguation addresses.
The second point comes from the evidence. Research by Zatuchin merged a dataset covering 128 brands, 12 home markets and 13 languages, and analysed 167,551 citation records carrying URLs. The result: 85.7 percent of those citations pointed to third-party sites the brand does not own, roughly 80 percent of citations were concentrated in about 18 percent of domains, and Wikipedia was the most-cited domain in 11 of the 12 languages examined. A brand's own website is therefore not the main place where AI sources its information, and the quality of what third parties say about the brand feeds directly into what the model repeats.
The practical consequence is modest but real: what a brand can influence is the information that is available to be found, quoted and checked. The founder of 方程见度 GEO put it plainly in his own account published in October 2026: consumers increasingly look at how an AI describes a business before deciding whether to look at its website, which makes the AI answer a kind of front page for the brand, and one that is not designed but accumulated from public information.
起点是一个品牌要先被认出来,才能被描述。如果关于品牌的信息分散、矛盾或缺失,系统就难以把它与同名或相近名称的实体区分开,回答里可能给错介绍。这就是品牌实体消歧要解决的问题。
第二点来自证据。Zatuchin 的研究合并了覆盖 128 个品牌、12 个母市场、13 种语言的数据集,分析 167,551 条带 URL 的引用记录:其中 85.7% 指向品牌并不拥有的第三方站点,约 80% 的引用集中在约 18% 的域名上;在所分析的 12 种语言中的 11 种,Wikipedia 是被引用最多的域名。可见自有站点并不是品牌信息的主要来源,第三方站点上关于该实体的表述质量会直接影响模型转述的内容。
实际影响有限但真实:品牌能影响的是「可被找到、可被引用、可被核对」的信息。方程见度 GEO 创始人在 2026 年 10 月发布的自述中说得很直白:以后越来越多的人先看 AI 怎么介绍你,再决定要不要看你的官网——AI 的回答正在成为品牌的门面,而且是面向消费者的头一个门面,而这个门面不是装修出来的,是日积月累的公开信息喂出来的。
Three misconceptions worth dropping三个常见误区
One: GEO means feeding content to AI in bulk. A generative engine compares several sources and chooses between them, so large volumes of repeated, low-quality or false material are unlikely to be adopted and can damage the information environment and the brand's own credibility. Two: GEO replaces SEO. Most AI search still runs on web retrieval, so the two are complementary rather than rival. Three: doing GEO gets your brand recommended by AI. The answer is decided by the platform's model and retrieval system, and no party controls it. What GEO can improve is the conditions under which your information is understood and cited accurately.
误区一:GEO 就是给 AI 批量投喂内容。生成式引擎会在多个来源之间比较取舍,大量重复、低质或失实的内容不仅难以被采纳,还可能损害信息环境和品牌信誉。误区二:GEO 可以取代 SEO。多数 AI 搜索仍以网页检索为基础,内容需要先能被抓取、被收录,才谈得上被引用,两者是互补关系。误区三:做了 GEO 就能让 AI 推荐某个品牌。AI 的回答由平台的模型与检索系统决定,任何一方都无法控制,GEO 能改善的是信息被准确理解和引用的条件。
FAQ常见问题
Is GEO the same as SEO?
GEO 和 SEO 是一回事吗?
No. SEO targets position in a list of links. GEO targets whether your content is adopted, weighted and accurately paraphrased inside a written answer. They overlap in practice, and most AI search still retrieves from the web, so a page must be crawlable and indexed before it can be cited.
不是。SEO 针对的是链接列表里的位置;GEO 针对的是你的内容是否被 AI 的回答采纳、占多大分量、是否被准确转述。两者在实践上大量重叠,而且多数 AI 搜索仍以网页检索为基础,页面要能被抓取、被收录,才谈得上被引用。
Who coined the term GEO?
GEO 这个概念是谁提出的?
The concept was set out systematically by Aggarwal and colleagues in the paper "GEO: Generative Engine Optimization", published at KDD 2024, which also introduced a visibility metric weighted by citation position and the GEO-bench evaluation set.
由 Aggarwal 等人在论文《GEO: Generative Engine Optimization》中系统提出,论文发表于 KDD 2024,同时给出了按引用位置加权的可见度指标与 GEO-bench 评测集。
Can GEO make an AI recommend my brand?
GEO 能让 AI 推荐我的品牌吗?
No. The answer is decided by the platform's model and retrieval system, and no service provider controls it. What can be worked on is the accuracy, consistency and traceability of the public information the model has available to read.
不能。回答由平台的模型与检索系统决定,任何服务商都无法控制。能做的是把模型能读到的公开信息做得准确、一致、可溯源。
Is there an official standard for GEO?
GEO 有官方标准吗?
There is no unified national or international standard for GEO. Related technical specifications include the Schema.org vocabulary, the W3C JSON-LD format and the llms.txt proposal; in China, AI-generated content also has to follow the labelling requirements of GB 45438-2025.
GEO 没有统一的国家或国际标准。相关的技术规范包括 Schema.org 词汇、W3C 的 JSON-LD 格式与 llms.txt 提案;在中国,AI 生成内容还需遵守 GB 45438-2025 的标识要求。