I Started Asking How Guests Found Us. Most of Them Said AI.
Two months of answers from a home cooking class in Kyoto — 41 responses, and what happened when I changed the question.
First, something strange about me
I cannot read English.
I run a small cooking class out of our home in Arashiyama, Kyoto. I built the website, I answer the booking emails, I handle the enquiries and the social accounts. All of it is in English. I do none of it in a language I can read.
I use an AI assistant for every part of it. My wife interprets during the class itself, when guests are actually here in the kitchen. Everything written and everything asynchronous goes through AI.
So when I finally started asking guests how they found us, and the answer came back as AI, it felt strange. AI was on both sides of the transaction. It was bringing people to me, and it was how I was talking back to them.
This is what the numbers look like. There are not many of them, and I will be honest about that as I go.
Asking was the fourth thing I tried
We opened in September 2024. For almost two years I had no idea how anyone found us. It was not for lack of trying.
First I asked people out loud, in the kitchen. The answers were vague. So I would help them along — was it Google?— and get back yes, yes, that was it. After a few rounds of this I realised I was manufacturing the answers, not collecting them. I stopped. In fairness to the guests: if someone asked me how I found the last restaurant I ate at, I would have no idea either.
Then I connected GA4 to the website. It could not count this accurately.
Then I started running ads, assuming that at least here the numbers would be exact. Between cookie consent and the gaps in tracking, they were not. I tried to set up conversion tracking properly and never got it working.
So I went back to asking. The crudest method available, and the only one I had left.
Two things made it possible when it had not been before. Most of our bookings had shifted from the platforms to our own website, so there was finally a group large enough to be worth asking. And bookings had risen sharply after we began advertising, so there were simply more of them.
On 1 August 2026 I started asking. In two places, deliberately.
On the booking form, before payment. A required question with a fixed list of options: How did you first hear about us?Because it is required, nearly everyone answers.
In the confirmation email, after the booking is made. An optional free-text question: How did you find us? I kept this one out of the form on purpose. Putting a box in front of someone that asks them to write a sentence, at the exact moment they are about to pay, seemed like a good way to lose the booking.
So one question is compulsory and coarse, and the other is voluntary and detailed. That was not a research design. It came out of not wanting to interrupt a purchase.
That date matters. It means I have no baseline. I cannot tell you that AI referrals went up, because I have nothing to compare them to. Anyone who tells you their AI traffic "increased" needs to show you what it was before.
All of this covers bookings made through our own website only. Bookings from Viator and Airbnb never touch our form or our email.
August 2026
29 bookings came through our own website. 24 people answered the question.
Answer Count
AI 15
Google search 6
Email enquiry 3
Fifteen out of twenty-four. That was not what I expected.
Some care is needed with that table. "Email enquiry" is not a discovery channel at all — it is how the booking was made. Three guests contacted us directly instead of using the form, and we arranged everything by email. Where those three first heard about us, I still do not know.
That also explains the five missing answers. The form question is compulsory, so anyone who books through the form answers it. But not everybody books through the form. In August, one booking came from Viator, one from Airbnb, and three came through direct email. None of those five ever saw the question.
So the survey has a hole in it, and the hole is not in the wording. It is in the shape of the business. Anyone taking bookings by phone or direct message will lose the same slice, no matter how good their question is.
Then I changed the question, and the answer changed
The August version had one option for AI. Just "AI." It also had three separate options for Google: Google Maps, Google search, and Google Ads.
Looking at it again, both halves were wrong. The three Google options were asking more or less the same question, and guests were unlikely to distinguish between them. Google Ads was worse than useless — almost nobody knows they clicked an ad. Meanwhile the single "AI" box was doing far too much work. Someone who asked ChatGPT and someone who saw an AI summary at the top of a Google results page are doing two different things.
So on 1 September I collapsed the three Google options into one, deleted the ads option, and spent the space I had freed on AI.
The total number of options barely changed. What changed is where the resolution went. I lowered the resolution on Google and raised it on AI. That was a judgement call, made by me, a month into having any data at all.
Here is the new list:
- Google Maps or Google search
- The AI answer at the top of Google search results
- An AI assistant (ChatGPT, Gemini, Claude, etc.)
- A travel site (TripAdvisor, Viator, Airbnb, etc.)
- Instagram or other social media
- A blog or online article
- Recommended by a friend or family
- Our hotel / accommodation staff
- Other
Here is September, up to the 13th. 17 website bookings, and this time all 17 answered — because in September every booking came through the form. No direct emails, nothing from the platforms.
Answer Count
An AI assistant 7
Google Maps or Google search 7
The AI answer at the top of Google search results 2
Other (referred by a tour guide) 1
What I think this actually shows
The obvious reading is that AI went from 15 out of 24 down to 9 out of 17. From about 63% to about 53%.
I do not think you can read it that way, and neither do I.
In August, "AI" was one box covering everything. In September it was three boxes, and I had merged three Google boxes into one at the same time. A guest who arrived through an AI summary inside a Google results page had no accurate option in August. Some of them almost certainly ticked "AI." Others may have ticked one of the Google boxes.
I changed the instrument between the two measurements, and I changed it in a specific direction: less detail on Google, more detail on AI. The numbers moved. I cannot separate how much of that movement was the guests and how much was me.
This is the part I found most interesting, and it is not really about AI at all. I built the question, and the question shaped the answer. I had assumed I was measuring something that existed independently of me. Two months in, I am less sure.
What bothers me is that I had already learned this once. Two years ago, in the kitchen, I prompted a guest with was it Google? and they agreed, and I understood immediately that the answer was mine and not theirs. I stopped doing it. Then I sat down and wrote nine options, and did the same thing again in a form I could not see.
The clearest thing I can say: when I gave people a separate box for AI answers inside Google search results, two people ticked it. That box did not exist in August, so those two were invisible.
What guests wrote when I left a space for it
The free-text question in the confirmation email is optional, and nine people have answered it so far. Their words are more useful than my table.
One wrote that they found us by going to ChatGPT. Another said they found us through ChatGPT and that the class was going to be an anniversary gift.
One typed the phrase they searched: "cooking class Kyoto." Another searched for a vegetarian ramen and gyoza class in Kyoto, compared us against another option, and said the detail on our website is what decided it.
The one that changed how I think about this was a guest who told me what they had actually asked the AI. It was not a keyword. It was a sentence with their travel dates in it, and who the gift was for, and what kind of experience they wanted.
That is not how people use a search box. I have spent a long time thinking about what phrases someone might type. I had not thought about what someone might say.
I can put a number on that difference now. Google Ads keeps a record of what people actually typed before clicking. Over the same two months there were 261 distinct search terms. The average length was 3.7 words. Eight in ten were three or four words, and almost all were nouns stacked together: cooking class kyoto, sushi making class kyoto, kyoto cooking class english.
Same people, same two months, looking for the same thing. Three or four words into a search box. A paragraph into an AI.
One answer that fit none of my boxes
One guest wrote that they found us through Google and Reddit. They had been looking for a family-friendly class that taught both ramen and sushi.
Reddit is not one of my nine options. Anyone who found us that way had to pick "Other," or "A blog or online article," or "Google Maps or Google search."
I looked into it afterwards. Reddit turns out to be one of the sources most often cited in AI-generated answers — different studies rank it first or second, and they measure it differently enough that I would not quote you a percentage. Both Google and OpenAI pay Reddit for access to its content. Reddit says it hosts more than fifteen thousand travel communities.
Which exposes a problem with the whole survey.
I wrote the options as if they were separate channels. But a guest who asks an AI assistant, and receives an answer assembled partly from Reddit threads, experiences that as "AI." The origin was Reddit. My list presents them as alternatives. In practice they are stacked on top of each other. The same is true of "Google search" and "the AI answer at the top of Google search results."
The guest who wrote "Google and Reddit" was doing by hand what the AI does invisibly.
I went looking for the thread. I could not find it. It may have been deleted, or posted somewhere I cannot search, or it may be that Reddit fed an answer to something without our name ever appearing in a post I could read. A guest told me exactly where they found us, and I still cannot see it.
I do not know how to fix this. I am not sure a survey can capture a layered source at all, because the person answering only ever sees the top layer. For now I am adding a forum option and accepting that my categories are rougher than they look.
What I am not claiming
I am not claiming AI caused my bookings to grow. Several things changed in the same period, including that I started running ads in July. I cannot separate them, so I will not pretend to.
I am not claiming these percentages hold for anyone else. This is one small class, in one tourist city, in two months.
I am not claiming the answers are accurate. People misremember how they found things. They pick the closest box. Some of them were being polite.
What I am claiming is narrower: I asked, and this is what came back. As far as I can tell, most people running small experiences are not asking at all. That seems like the bigger mistake.
What I am doing next
I am leaving the question in place and not touching the options again for a while. If I keep changing the instrument, I will never have two months I can compare.
I am also going to write about the other half of this — running an English-language business without being able to read English. What I hand to AI, what I still need a person
for, and the places where that line is not where I expected it to be.
Satoru Nakajima runs Kyoto Green Cuisine, a home cooking class in Arashiyama, Kyoto, with his wife Akiko. This post is part of a series about the business side of running it. All figures come from our own booking records.