Introduction: a power nobody noticed they had
Companies used to explain themselves through three channels: the website, the sales team, and the press. A prospective client wanting to understand a company would open its website. To go further, they would search for news and look through case studies. Still unsure, they would ask a salesperson or someone they knew.
Throughout that process, the company could count on one thing: the client was looking at it directly. The right to explain stayed with the company. Whatever the website said was what the client read.
That premise is now loosening. A growing number of people, before they open any company's website, start by asking an AI. What does this company actually do? How is it different from its competitors? Which Chinese brand design firms are worth considering? Which firms are right for a manufacturing rebrand? Is this company worth working with?
ChatGPT, Claude or Gemini then does a first round of reading, summarising and judging on the user's behalf. What the user sees is already a conclusion.
Which raises a question that has rarely been asked: is the company AI knows the same company you believe you are?
Often, it is not. AI may know a company's name but not its real positioning. It may know the products but be unable to say what sets them apart. It may have read an introduction from several years ago and missed today's strategy. It may even treat an outdated line on some third-party site as the company's defining description.
This is not entirely the model's fault. The more common cause is simpler: companies have never prepared their brand information to be understood by machines.
1. AI does not open your website and read it from the first page to the last
Understanding this is the precondition for everything that follows.
When a question needs current information, today's leading AI products increasingly rely on web retrieval and source citation to build their answers. But they read the world very differently from people, and the three companies behind them have each described how.
OpenAI states in its developer documentation that OAI-SearchBot is the crawler used to surface websites in ChatGPT's search features. Its publisher FAQ is explicit on the broader point: "Any public website can appear in ChatGPT search." Sites that want their content summarised and cited should not block OAI-SearchBot.
Anthropic explains in its help documentation that when web search is enabled, Claude invokes a search tool to ground its responses in content from the live web, and provides direct citations and source links so that users can check what they read.
Google made a series of changes during 2026 to how sources appear in its generative search. In May, Preferred Sources — the sites a user chooses in their search settings — was extended to AI Overviews and AI Mode, with those sources clearly labelled in AI responses. At the same time Google introduced a "Highly Cited" label to identify articles that are frequently referenced by other stories. In June it launched Generative AI performance reports in Search Console, giving site owners their first view of how many impressions their pages receive in AI Overviews and AI Mode, and from which countries and devices.
The implication is significant. A website's cited status in generative search is turning from a vague impression into something that can be observed.
So an AI's understanding of a brand today is not the product of reading a website. It is assembled, at the moment of answering, from the website, third-party pages, news coverage, industry directories, social content, public case studies, structured data and live search results. What emerges is a temporary picture.
That is exactly where the problem lies. If those sources disagree with one another, the model has to decide for itself. And most companies, without realising it, have handed the most important right they have — the right to explain themselves — to that assembly process.
2. The real question is not whether AI knows you. It is how AI defines you.
The first time many companies test an AI, they ask whether ChatGPT knows their company. That question is too shallow. Knowing a name and holding an accurate model of a brand are two completely different things.
The question worth testing is what kind of company ChatGPT thinks you are — and then, in turn, what it believes your core strengths are, what kind of client it thinks you are right for, how it thinks you differ from your main competitors, and why it thinks someone should choose you.
You will usually find the answers start to drift at the second question, and by the fourth are barely recognisable.
The reason lies in how models classify. If a company is described again and again across public sources as a "brand design company", a model will very likely file it under traditional design services — even if the company has long since changed, and now works across brand strategy, brand systems, GEO, AI brand workflows, knowledge bases and long-term governance. If those new capabilities exist only in a slide deck, a single talk or a handful of scattered articles, a model will struggle to form a stable understanding of them. It will keep using whichever definition has been repeated most.
So brand-building in the AI era has to ask a question nobody used to ask: how do we want machines to define us? The nine actions below are all answers to that question.
3. Action one: give the brand a stable, explicit, repeatable official definition
The most obvious problem on most company websites is that the company has never actually defined itself. The About page says one thing, the homepage says another, the recruitment site says a third, press coverage a fourth, LinkedIn a fifth. Three years on, five versions of the company's self-description are in circulation.
People can accommodate the difference; they assume the company has changed over the years. AI cannot necessarily tell which version is current.
So the first step is very plain: establish a single, stable definition, and use it consistently.
That means no longer describing yourself as "an international brand design agency" on one page, "an innovative brand consultancy" on another, and "a strategic design practice", "an AI brand company" or "a full-service brand provider" elsewhere. If these phrases have no clear relationship to one another, a model cannot tell which is the primary definition. It treats them as five weak, parallel signals.
The better approach is a fixed definitional sentence. A good official definition usually does four things: it names the category the company belongs to, its core method, who it serves, and where its difference comes from. Xinming Design's reads:
Xinming Design is a brand systems company grounded in brand strategy and powered by AI, helping organisations build brand systems that run sustainably through Brand OS.
That sentence should appear word for word wherever the company is introduced. On the homepage, it is the first landing point for both people and machines. On the About page, it is the self-description most often retrieved and cited. In the description field of the Organization schema, it is the structured definition that machines read directly. On LinkedIn and other social profiles, it sits on high-weight third-party platforms. In the company overview and press kit, it becomes the source that media coverage draws from. In major interviews and public talks, it becomes raw material for third-party reporting. And in industry directories and partner platforms, it becomes the basis on which outsiders classify the company.
The point is not to repeat keywords mechanically. It is to establish entity consistency, so that every source gives the same answer to the question: what is this company?
4. Action two: replace "we're very good" with "here is our evidence"
The kind of brand content AI finds hardest to use is adjectives: professional, leading, international, innovative, premium, excellent, pioneering. These words carry almost no information, because any company can describe itself this way. A word everyone can use cannot distinguish anyone.
What AI can actually use is facts that can be verified. That includes basic entity information such as founding year, place of registration and office locations; the scope of the business, meaning the sectors, services and geographies it covers; named clients, named projects and a description of what the work involved; public identifiers such as a listed company's stock code, or formal certifications; third-party recognition in the form of awards, directory listings and media coverage; the methodology, the key principals and their published writing; and, above all, links where each of these can be checked.
Google's 2026 guidance for site owners lists "unique, non-commodity content" as the first priority for performing well in AI search. The Highly Cited label introduced in May likewise gives explicit visibility to original and widely referenced sources.
So the way a brand website is written has to change. The old way was to write: We have extensive experience in manufacturing brand work. The new way is to write: Xinming Design has delivered brand identity upgrades for manufacturers including Pearl River Piano Group (珠江钢琴集团), Shouhang New Energy (首航新能源), Ritmüller (里特米勒) and Strauss (施特劳斯). Each project can be viewed in the portfolio section of our website.
The second version answers four questions the first leaves open: who you worked for, what you did, where the work can be seen, and whether it can be verified independently.
Put simply: the brand claim is what makes people remember you. The evidence is what makes AI believe you.
5. Action three: a website can no longer be a company brochure
This is the area most companies most need to rebuild.
The traditional website is organised as Home, About, Products, Cases, News and Contact. For people, that is enough. For AI, it falls well short. People browse a website to form an impression; machines read a website to extract knowledge. What a machine actually needs is a knowledge structure.
The websites of the next few years should therefore progressively add pages of several new kinds. A definition page states plainly what the brand is, and is what a model reaches for when asked what the company does. A methodology page explains how the company works, and answers the question of how it differs from competitors. Service pages set out the problem each service solves, so that a model can answer what the company can do for a given client. Industry pages describe the sectors served, which is how a model answers "who has worked in my sector?" A proof page gathers the facts that support the company's capability, and is where credibility questions get resolved. Case pages describe the specific problem each project solved, so that a model can match the company to a similar brief. An FAQ page captures what people genuinely ask. And knowledge pages hold the company's accumulated professional perspective, which is what models draw on for sector and method questions.
In the guidance for site owners it updated in June 2026, Google emphasises — alongside original content — how content is organised, a good page experience, and high-quality images and video. The focus has clearly moved from keywords to the structure and quality of content.
This is the judgment we have been making for some time: the website is changing from a showcase into the brand's cognitive infrastructure.
6. Action four: give the brand citable answers
Traditional SEO has long asked how many times a keyword appears. The more useful question in GEO is whether your page contains a single sentence that directly answers the question being asked.
Suppose someone asks what Brand OS is. If the website devotes a thousand words to the topic without one clear definition, the AI has to summarise it — and what it produces may not be what you meant. If instead the company simply states that Brand OS is a method for integrating brand judgment, structure, visual identity, knowledge and governance into a long-running system, the model can cite that sentence directly and accurately.
The same applies to every core concept a company owns or relies on. What is an AI brand design company? What is GEO? What is a brand knowledge base for? Each needs a direct definition somewhere on the site.
A citable answer usually has four qualities. It sits near the top of the page, in a prominent position, rather than being buried in the fifth paragraph. It is complete as a single sentence, so it can be understood out of context. It is a firm statement rather than a hedge such as "to some extent" or "in a certain sense". And it is consistent across the whole site, so that no concept ever carries two definitions.
High-quality brand content will therefore increasingly contain two languages at once: one for people to read, with a point of view, rhythm and story; and one for machines to extract, clear, direct and verifiable. They do not conflict. Genuinely good GEO content does both in the same article.
7. Action five: build external consistency, not just the website
This matters a great deal and is the part most often overlooked.
What a company says about itself is only one layer of evidence. AI also looks at how the news describes you, how clients describe you, how partners describe you, how industry sites classify you and how social media introduces you. If the website says "an AI-driven brand systems company" and dozens of third-party pages still say "a logo design studio", the model sees conflicting signals — and in a conflict, the more numerous side usually wins.
That is why GEO cannot be reduced to writing more articles. What actually needs building is source consistency, and it has several layers, each with its own lever.
Official sources — the website and the company's own social accounts — can simply be unified and edited directly. Client sources, such as client websites and client announcements of a collaboration, cannot be edited by the company, but it can supply standard wording and ask for updates. Media sources are best managed through a single, maintained press kit that journalists draw on. Industry sources, such as directories, associations and platforms, usually allow a company to update its own classification if it takes the initiative. And social sources, including staff profiles and third-party discussion, improve when the company provides a standard, citable form of words for people to use.
The more consistent the information across these layers, the more likely a model is to form a stable picture.
One more thing is almost always missed: cleaning up outdated information. After a rename, a pivot or a change of business, the old descriptions persist on third-party platforms for years. Building new information and retiring old information have to happen together.
8. Action six: make sure AI can actually reach your content
This is the most basic layer, and the one where things most often go wrong.
Many company websites run Cloudflare, a web application firewall, bot protection, JavaScript challenges, CAPTCHAs or geographic restrictions. The result is that people can reach the site normally while search crawlers and AI crawlers are shut out.
OpenAI's documentation is specific here. Its developer documentation recommends allowing OAI-SearchBot in robots.txt and also allowing requests from OpenAI's published IP ranges, and notes that it can take around 24 hours for a change to robots.txt to take effect. In separate guidance for advertisers, OpenAI states plainly that protection services such as Cloudflare and Akamai "can mistakenly block legitimate crawlers, often returning 403 Forbidden errors", and that CAPTCHAs and JavaScript challenges can block access in the same way.
Google added another setting to check in June 2026. Search Console now includes a control that lets site owners decide whether their site appears in, and helps ground responses in, Google's generative AI Search features. Sites that opt out receive no traffic or impressions from those features.
At a minimum, a company should work through the following. Confirm that robots.txt allows OAI-SearchBot, since this governs appearance in ChatGPT search. Confirm that OpenAI's published IP ranges are allowed through the firewall, because editing robots.txt alone is not enough. Check whether the CDN, firewall or bot protection is blocking crawlers, which typically shows up as crawlers receiving a 403 response. Check whether CAPTCHAs or JavaScript challenges are being applied to crawlers as well as people. Make sure core pages return an HTTP 200 status. Make sure core content does not depend on heavy JavaScript to appear, since some crawlers execute JavaScript only partially or not at all. Verify that canonical tags point to the right pages, so authority is not split. Look for noindex tags left behind by mistake, which is common after a redesign. And confirm the state of the generative AI setting in Search Console, so that it has not been switched off by accident.
This looks technical. In practice it is now brand infrastructure, because brand content a machine cannot reach does not exist, as far as AI is concerned.
9. Action seven: structured data is valuable — but schema is not a cheat code
Once companies start working on GEO, many ask whether adding schema will make ChatGPT start recommending them. That is the wrong way to think about it.
Schema helps search systems understand the semantic relationships on a page more accurately: who the organisation is, what is a product, what is an article, who the author is, what is a question and answer, and how pages relate to one another. It is a semantic aid, not a recommendation switch.
At a minimum, a company should describe itself with Organization markup, carrying its core entity information and the official definition; identify the site itself with WebSite markup; attribute its articles with Article markup that records author, publisher and publication date; show each page's place in the site hierarchy with BreadcrumbList; describe what it offers with Service or Product markup; identify its principals and authors with Person markup; and use FAQPage markup — but only where questions and answers genuinely exist on the page.
That last condition deserves emphasis. Structured data added for its own sake, that does not match what is on the page, is not merely useless; it can damage credibility.
What ultimately determines how a brand is understood is whether the content itself is accurate, substantive, consistent, evidenced and genuinely distinctive. Schema makes that content easier to read. It cannot substitute for it.
10. Action eight: don't overlook images, video and multimodal content
This will be one of the most important layers over the next two to three years.
In its one-year review of AI Mode, published in May 2026, Google reported that more than one in six searches in AI Mode in the U.S. now use voice or images, with image searches growing more than 40% month over month. Search and AI are moving quickly into multimodality.
Brands will therefore need to consider not only whether their text can be understood, but what their images are actually saying and whether they have an identifiable subject and context; whether their videos serve as atmosphere or as evidence of capability; whether the data in their charts is locked inside an image or explained in text; whether product images are labelled with name, model and application; and whether project visuals carry the project name and the problem solved alongside them.
One problem is especially common. Many companies present their service process, certifications, client lists and key figures as beautifully designed long images. These are friendly to people and close to unreadable for machines. Core information should always exist in text as well.
On most brand websites, images have mainly served as atmosphere. They will increasingly have to serve as knowledge.
11. Action nine: what really matters is continuous re-testing
GEO is not a one-off optimisation. Models change, search changes, and the company itself changes. The effect of any single round of work may drift after the next model update.
The good news is that re-testing is becoming feasible. Google's Generative AI performance report in Search Console now shows site owners which pages receive impressions in AI Overviews and AI Mode, and from which countries and devices. It should be noted that the report currently shows impressions only; it does not include clicks.
Beyond official data, companies should run their own re-testing. The recommended approach is to put the same set of questions to ChatGPT, Claude and Gemini once a month. The set should cover four kinds of question: one testing entity recognition, such as asking what kind of company yours is; one testing how the model understands your strengths; one testing category recall, such as asking which companies in your sector are worth considering; and one testing scenario matching, such as asking which firms suit a particular type of client with a particular need.
For each answer, record whether you appear at all in the category and scenario questions; how far the description deviates from your official definition; which sources the model cites, and whether it is relying on your website or on third parties; which details are out of date, whether old services, old positioning or old figures; and whether the three models agree, or which one's picture deviates most.
This step is critical. What GEO really has to manage is not ranking. It is model perception. A ranking is a position; a perception is a definition — and it determines what you are described as when you are mentioned at all.
12. Do ChatGPT, Claude and Gemini need separate optimisation?
You need to understand how they differ. But building three different sets of content for three platforms is not recommended.
Each system has its own retrieval, indexing and answering mechanisms. ChatGPT's search reads public web pages and cites its sources; the technical point to watch is that OAI-SearchBot and its IP ranges are allowed. Claude's web search draws on live web content and gives direct citations; the requirement is that pages can be crawled and read normally. Gemini and Google's AI features are deeply integrated with Google Search, where AI Overviews and AI Mode have become major entry points; the points to watch there are the generative AI setting in Search Console and the Preferred Sources mechanism.
But what all three need is strikingly similar: accurate information, a stable entity, a clear definition, sufficient evidence, high-quality original content and consistent third-party sources.
So companies should not build "ChatGPT SEO", "Claude SEO" and "Gemini SEO" as three separate systems. The more sensible approach is to get the brand's own knowledge structure right first, and then check technical accessibility and actual performance platform by platform. One set of content; checks run per platform.
13. Through the lens of Brand OS: where the nine actions sit
In Xinming Design's Brand OS v1.5, a brand system is made up of six layers. The nine actions above are not nine isolated tasks. Each sits within one of those six layers, and together they form what could be called a company's machine-readable layer.
The Brand Kernel answers the question of who we are, and it is where action one — the official definition — belongs. Brand Context answers the question of the context in which we are understood; it is home to action three, the website's knowledge structure, and action four, citable answers. Brand Asset answers the question of what we can show, and it holds action two, evidence, and action eight, multimodal content. The Agent Protocol layer answers how machines read us correctly, and it is where action six, technical access, and action seven, structured data, sit. Brand Governance answers what belongs to us and what does not, and it contains action five, external consistency and source management. The Interface & Learning Loop answers whether we are being understood correctly, and it is where action nine, continuous re-testing, lives.
Within Brand Governance, the brand constitution is the highest-order rule. It determines which descriptions belong to the brand and which do not, even when they are effective. In an AI context, it decides which descriptions should be corrected and which third-party sources need clearing first.
Together, the six layers determine one thing: when ChatGPT, Claude and Gemini face a user's question, whether they can form a picture of the brand that is close to the company's real state.
So GEO, in the end, is not about pandering to AI. It is about taking brand information that was scattered, vague and dependent on human interpretation, and reorganising it into a clearer knowledge system.
That work would be worth doing even without AI. AI has simply made the cost of not doing it visible for the first time.
14. A self-check
Nine questions correspond to the nine actions, and a company can answer them in an afternoon.
First, does a single, word-for-word official definition exist, and does it appear on the homepage, the About page, in schema and on the major social platforms? Second, do the capability statements on the website rest mainly on verifiable facts rather than adjectives? Third, does the website include knowledge-type pages — definition, method, industry, proof, case and FAQ? Fourth, does every core concept have a direct definition near the top of its page that stands on its own? Fifth, do the major third-party sources describe the company consistently with the official definition, and have outdated descriptions been cleared? Sixth, is OAI-SearchBot allowed, are protection systems letting crawlers through, and has the generative AI setting in Search Console been confirmed? Seventh, is Organization, Article and other structured data in place, and does it match the page content? Eighth, does the core information in images and video also exist as text? And ninth, is a fixed set of questions re-tested across all three platforms every month, with the results recorded?
A company that can answer yes to fewer than five of these is where most companies currently are.
Frequently Asked Questions
We're a small company with no dedicated team. Where should we start? Start with actions one and six. The official definition is a single piece of writing and the technical check is a single configuration task. Neither needs ongoing effort, but both determine whether everything else works. If crawlers cannot reach your website, the value of all other content work is sharply reduced.
How long before we see results? Fixes to technical access are usually fastest; OpenAI states that changes to robots.txt take around 24 hours to take effect. The effect of content and source consistency depends on re-crawl cycles and model update schedules, and is usually observed over months. That is why action nine recommends monthly re-testing.
Can we just work on the website and leave third-party sources alone? The effect will be limited. When one new definition on your website faces an old definition on dozens of third-party pages, the model sees a conflict, and the more numerous side usually wins. Work on the website and on external consistency needs to move together.
Will GEO replace SEO? No; it builds on it. The foundations of SEO — technical access, structured data, content quality — still hold in GEO, and matter more. What GEO adds is active management of how models define you.
If AI describes us incorrectly, can we ask for it to be corrected? There is currently no direct channel for changing a model's perception. The only effective route is to correct the sources the model relies on: unify the official definition, add verifiable evidence, update third-party sources, and then observe the change through re-testing. That is the common logic behind all nine actions.
Closing: AI will not define your brand for you — but it will explain your brand to your customers
This is the thing companies most need to recognise today.
You can choose not to participate, and AI will still explain you, using whatever it finds. You can choose not to invest in GEO, and models will still form a judgment about you. The only real question is whether that judgment is something the company built deliberately, or something the internet assembled at random.
So brand work is gaining a new and important responsibility: managing how machines perceive you. Brand teams used to be responsible for how people see the company. From now on, they are also responsible for how machines understand it.
A genuinely mature brand should achieve three things at once. People understand it. The organisation can use it. AI describes it accurately.
This may be the new basic discipline of branding in the AI era.
发表评论