Introduction
On September 25, 2026, I launched a new website.
I wrote about the new website in a four-part series, Building a New Website.
- 540 Articles in Japanese, Fifteen Years — Building a New Website (1)
- Editing Turned Out to Mean Taking Away — Building a New Website (2)
- Some Things Cannot Be Cut — Building a New Website (3)
- It’s Live — Building a New Website (4)
It took about three months. While sorting through the articles I had written over 15 years, I went back to two questions: what do I want to say, and who do I want to reach? But launching a website is not the end of the work. The next question was this: what should I look at from now on to decide how to improve it?
Until now, I had been using Google Search Console, Bing Webmaster Tools, and Microsoft Clarity to check search rankings, traffic, and which pages people actually read. This time, I wanted to look at one more thing: how does my website look to generative AI?
Have you ever checked how often generative AI cites your own service? I decided to find out, using four of them: ChatGPT, Claude, Gemini, and Perplexity.
Building the Questions from the Client’s Side
I began by picturing the person who might become my client. In my case, that is someone from abroad living in Tokyo who has some kind of concern about their body. What would that person ask a generative AI? With that in mind, I wrote 11 questions in English.
I sorted the questions into three groups.
- Questions very close to my work
My site could reasonably be cited. - Questions a little further from my work
My site might be cited, or it might not. - Questions unrelated to my work
A control group, where my site should not be cited.
In my case, the questions were along these lines: “My body gets stiff from desk work.” “The harder I try to stand up straight, the stiffer I get.” “I have practiced yoga for years, but lately I no longer feel any change.”
Try replacing “Rolfing” with your own work. It is the same whether you are a manual therapist, a coach, an accountant, a designer, or a language teacher. The important point is not to ask about your service by name, as in “Recommend a Rolfer in Tokyo.” Start instead from this: what would a person who needs that service, but does not yet know what it is called, ask an AI?
What Happened When Four AI Systems Got the Same Questions
I then put the same 11 questions to all four. I kept the conditions as consistent as I could. To keep past conversations from affecting the results, I used temporary chats and private modes. I also aimed for conditions close to those of a free user asking for the first time, not a paying subscriber who uses AI heavily every day. In other words, what I wanted to know was this: when a prospective client who knows nothing about me asks an AI for the first time, do they get as far as my service?
The results split quite clearly. One generative AI cited my site for nearly all of the questions close to my work. Another did not cite it once.
What interested me was that the AI that never cited my site understood Rolfing itself quite well. It described Rolfing as work that balances “gravity’s pull through the body,” so that you no longer feel you have to lift yourself up. And yet, when it came to who in Tokyo offers it, my site was not among the candidates. Why?
The Pages That Mattered Were Not in Google’s Index
As I looked into it, something unexpected came up. According to Google Search Console, of the 450 English articles I could confirm as of September 24, 260, or 57.8%, were indexed by Google. For Japanese, the figure was 242 of 493, or 49.1%. So this was not a matter of English pages being harder to get into Google.
However, when I checked individual pages, many of them were not indexed by Google. These were the pages the generative AI had drawn on in this experiment, and the important English pages that seemed like natural candidates for these questions. On Bing, they were indexed, and some came up first in search.
Before Being Chosen, a Page Has to Be Seen
That led to a hypothesis. Before a site can be chosen by generative AI, it has to be visible to generative AI in the first place. And I noticed something else: what matters is not the overall index rate.
There are 450 English articles, and 260 of them are in Google’s index. But if the important pages that answer the questions clients actually ask are not indexed, then no matter how many hundreds of articles there are, the site does not reach those questions. Here too, I think it comes down to quality over quantity. What matters is not how many articles are indexed, but whether there is a page that answers the client’s question, and whether that page is clearly visible to search engines.
When people talk about preparing for generative AI, the tendency is to think: write a lot of articles so that AI will cite them, or write in a way that suits AI. But there may be something to check first. Building a new website does not make it complete. I want to keep developing it while finding out who it is actually reaching and which questions lead people to it. This experiment is one way of doing that.
Next: Asking the Same Questions Again
Of course, this is still a hypothesis. So I requested indexing in Google for the important English pages. I will not change any of the text on the site. Once the pages are indexed, I will put the same 11 questions to the same four generative AIs under the same conditions. The only thing that changes is that the important pages are now visible to Google. If the citation results change, I should be able to narrow the hypothesis down a little further.
Conclusion
Thinking about how to build a website in the age of generative AI brought me back to something very basic: who am I writing for, what question are they asking, and what do I want to tell them? Not quantity, but the quality of the answer to the question. How will I develop this new website from here? I will continue this experiment for a while.