This article is based on a series of recent practical tests and research notes concerning personal websites, AI search, and GEO. The initial question was relatively simple: beyond traditional Google search, can a personal website enter AI search systems such as DeepSeek and Claude, and how do these systems process the content they find there? As the experiments progressed, I found that the questions involved were beginning to extend beyond what is usually understood as SEO or GEO.
As the experiments progressed, the question became broader than conventional SEO or GEO. AI was not only retrieving individual articles. It was beginning to connect ideas distributed across multiple articles by the same author and use them to analyse new questions. That suggests a possible new way of thinking about digital persona.
The first experiment used an original concept that was easy to verify. With DeepSeek’s web search enabled, entering Sustenesis allowed the system to find relevant pages and correctly connect Sustenesis, Sustenesis Theory, the Chinese term 维成论, and Geoffrey Chen. Reversing the query and entering only Geoffrey Chen also allowed DeepSeek to distinguish the relevant Geoffrey Chen from several people with the same name and associate him with Sustenesis Theory, philosophy and artificial intelligence.
When web search was disabled, DeepSeek no longer knew what Sustenesis referred to. Claude showed a similar pattern. This indicates that the effect currently exists primarily in the retrieval layer rather than in the parametric knowledge of the base model. It would therefore be inaccurate to say that this material has already been trained into the models. What can be said is that it has entered public knowledge spaces that some AI systems are able to retrieve and use.
The experiments then moved from the original theory to ordinary material on the personal website.
For example, when asked why Geoffrey Chen thinks adult piano learners may play each hand well separately but struggle when putting both hands together, DeepSeek searched the adult piano material on geoffreychen.com and built an answer from several articles dealing with bilateral coordination, musical structure, hearing and memory.
When the question changed to how Geoffrey Chen views the fact that life goals keep changing, the search moved to a different cluster of articles dealing with habits, continuity of self, cognitive structure and personal philosophy.
When asked about the performance style of a well-known pianist, DeepSeek moved again, drawing on articles about piano education, stage presentation and bodily movement.
These tests suggest that AI is no longer forming only a simple relationship between an author and a website. It is beginning to connect an author, multiple themes, several articles and the judgements expressed across them.
This differs from traditional search. Conventional search is mainly concerned with which pages match a query. AI systems can go further by asking whether those pages belong to the same person and what position those pages collectively express.
A more interesting case appeared when I asked about Geoffrey Chen’s view of older adults participating in artistic activities and the distinction between enjoying art and performing art. A newly published article directly addressing this question had not yet entered DeepSeek’s effective retrieval range.
DeepSeek explicitly stated that it could not find a direct discussion of the issue. But it did not stop there. It searched earlier articles by the same author on later-life cognition, social presentation, face and cognitive structure, then used those materials to infer how the author might approach the question.
This suggests at least three levels of AI use of personal website content.
The first is direct retrieval, finding what the author explicitly said.
The second is cross-document synthesis, combining relevant views distributed across several articles.
The third is viewpoint inference, using the author’s previously expressed ideas to infer a likely position when no direct answer exists.
The third level goes beyond ordinary web search. AI begins to treat a dispersed body of writing as a relatively continuous structure of thought.
Traditional digital-persona systems usually work by collecting a person’s articles, documents, conversations and other records into a dedicated knowledge base, then building a specialised AI system that imitates or answers on behalf of that person.
These experiments suggest another possible architecture.
A personal website can act as long-term memory. Search systems can retrieve relevant parts of that memory. A general-purpose language model can then perform understanding, integration and reasoning.
The personal website provides long-term memory.
AI search provides dynamic retrieval.
The large language model provides understanding, synthesis and reasoning.
This structure does not require a dedicated personal model. Different general AI systems can access public material when needed and temporarily construct a working model of a person’s thinking.
That also gives a large archive of personal writing a value different from traditional SEO.
From an SEO perspective, one thousand articles mainly represent one thousand possible search entry points. From an AI perspective, every article is also a sample of the author’s thinking. If the number of samples grows while core concepts, judgement patterns and intellectual structure remain relatively stable, AI may increasingly be able to reconstruct the author’s broader position from those distributed samples.
The experiments then moved to a stricter stage.
When a query explicitly contains Geoffrey Chen, DeepSeek is already quite good at finding the relevant site and articles. But ordinary users usually do not know the author’s name in advance.
The more important question therefore becomes whether AI can discover an author when the user asks only about an issue.
This proved more difficult.
For example, a general question about why adult piano learners struggle with two-hand coordination may be answered entirely from DeepSeek’s internal model knowledge, even when web search is enabled. The same happens with philosophical questions. When asked whether an AI system without subjective experience, but capable of storing, using, testing and revising knowledge, genuinely possesses knowledge, DeepSeek can answer using established epistemology and philosophy of mind without searching the web at all.
This reveals an often overlooked precondition of GEO.
Whether a website enters an AI answer depends not only on source ranking, but also on whether the AI decides to search.
If the model believes its internal knowledge is sufficient, external sources may never enter the candidate set. AI use of external knowledge therefore involves several stages — triggering retrieval, searching for sources, selecting sources, understanding them, and integrating them into the final answer.
I then changed the experiment.
After discussing AI and knowledge, I asked whether there were any independent researchers in Australia working on related questions. The query did not contain the name Geoffrey Chen. Because the system now had to identify real contemporary researchers, DeepSeek initiated web search. Among the results, geoffreychen.com appeared.
The path therefore became a conceptual question, followed by a research area, then an independent researcher, and finally Geoffrey Chen and geoffreychen.com.
This is a form of Zero-brand Discovery. The user does not know the author’s name and does not specify the website, but the AI discovers the author from the problem itself.
I then designed additional questions around AI and knowledge production — whether AI is shifting knowledge production from humans to machines, whether AI changes who or what counts as a producer of knowledge, and whether ideas similar to Knowledge Without a Knower are being explored.
In several tests, DeepSeek included geoffreychen.com as a source even though the name Geoffrey Chen was absent from the query.
This suggests that, at least within DeepSeek’s current retrieval space, an association is beginning to form among AI, knowledge production, the knower or cognitive subject, and geoffreychen.com.
The relationship is not yet stable and has not spread across every subject on the site, but it is observable.
To avoid overgeneralising from a single platform, I also tested other AI systems.
The differences were substantial.
DeepSeek currently performs best at discovering Geoffrey Chen and Sustenesis, connecting entities, and synthesising ideas across multiple articles.
Claude often fails to retrieve Sustenesis through open web search, but when explicitly directed to geoffreychen.com it can read and understand the material accurately.
Doubao currently does not effectively recognise either Sustenesis or Geoffrey Chen, and even when given sustenesis.com directly it does not show retrieval behaviour comparable to DeepSeek.
It would therefore be misleading to say that an idea has simply entered all AI systems. Different AI platforms use different search systems, indexes and retrieval mechanisms. Entry into AI knowledge space is better understood as a platform-by-platform diffusion process.
Traditional SEO has a relatively clear goal — improving the ranking of webpages in search results. If GEO is defined only as getting AI to cite a webpage, it remains close to the old search model.
These experiments suggest a more advanced set of GEO goals for personal websites.
First, the author’s name leads to the website.
Then the author’s name plus a question leads to the relevant theme and articles.
At a more advanced stage, the question itself leads the AI to discover the author.
After that, multiple articles can be used to reconstruct the author’s position.
Finally, a new question may be answered by inferring a likely position from the author’s existing body of thought.
The last two stages begin to overlap with the idea of a digital persona.
At that point, citation count is no longer the only important metric. The accuracy of the reconstructed thinking becomes more important.
Two possible measures are useful here.
Identity Fidelity measures how closely the AI-reconstructed intellectual position matches the real author.
Viewpoint Prediction Fidelity measures how closely AI can predict an author’s position on a question the author has not yet directly answered.
These may ultimately matter more than simply counting how many times an AI system cites a webpage.
In the early internet era, people mainly used the web to obtain information. With the rise of self-media, more people began building their own publishing channels in order to gain readers, traffic and influence.
The AI era may add another objective to public writing.
The aim is not only to make content visible to people today, but also understandable to machines.
If a person consistently publishes observations, judgements and ideas over time, and those materials retain sufficient conceptual continuity, they may gradually enter knowledge spaces that AI can retrieve, connect and reason over.
In that sense, what we contribute to an era does not have to be limited to products and services. It can also be thought.
Products are replaced. Services end. Human lives have clear temporal limits. But ideas that are preserved in a stable, public and machine-readable form may continue to be searched, analysed and used in future contexts.
The continuation of a digital persona therefore does not necessarily require a dedicated AI built to imitate a person. A simpler structure is already emerging.
Websites preserve thought, search systems retrieve thought, and large models reorganise and reason over thought.
These experiments are still at an early stage. The clearest effects have so far appeared in DeepSeek, and there are obvious limitations, including retrieval delay, platform differences and the risk of incorrect viewpoint inference.
But one change can already be observed.
Public personal writing is beginning to acquire a function beyond being read by people.
It is becoming material that AI can retrieve, interpret and reconstruct.
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.