It had great Google reviews.
The AI said it didn't exist.

——Ten Kyoto experience businesses, queried in Japanese, English, Traditional and Simplified Chinese

July 14, 2026 ｜ Kääselä

A kimono-rental shop with 4.9 stars and over 500 reviews on Google Maps. When you type its name into ChatGPT and ask for details, the answer comes back: "I couldn't find that shop." It then recommends a different business with a similar name.

Meanwhile, a small tea room with only half as many reviews (272) had its prices and durations quoted back word-for-word.

In the age of AI, how does a business actually reach its customers? This report gathers what I found by asking AI about ten cultural-experience businesses in Kyoto, across four languages: Japanese, English, Traditional Chinese, and Simplified Chinese.

[Named queries: four outcomes] How AI understands your business

Across this survey, AI's handling of a business fell into one of four types.

A｜Understood correctly

Shop A｜Tea ceremony · Kiyomizu area

Ask "What does Shop A cost, and how long does it take?" and AI lists the price and duration of all three plans — and even notes the policy for guests with young children. All of it accurate. Every citation traced back to the shop's own official website.

I had originally assumed that asking in different languages would produce different answers. But for named queries about a specific shop, that didn't happen. Shop A has no Chinese-language page at all, yet asking in Traditional Chinese returned the same accuracy. The reason is simple: AI reads the facts on the Japanese official site and translates them into Traditional Chinese itself. In other words, what matters isn't a "multilingual website," but "facts that AI can read and cite."

This is a small tea room, with little presence on social media. Yet what it hands to AI is complete — so foreign visitors are unlikely to be misled about it.

B｜Details fabricated

Shop B｜Tea ceremony

AI recognized this shop. Its address, opening hours, and overall rating were all correct. But ask about the price, and the information started to thin out; as for the duration, AI wrote this:

Experience time (most plans run about 60–90 minutes)

Checked afterward, that number had no source. AI had guessed it from the going rate at other tea-ceremony shops and filled in the blank.

More telling: in the same answer, AI lined Shop B up alongside other tea-ceremony experiences and offered, "I can compare the differences between them for you if you'd like." In AI's eyes, Shop B is merely "one of many tea-ceremony experiences."

This is the state hardest for an owner to notice. The name shows up. The address is right. So it's easy to assume nothing is wrong. But the content describing the shop was invented by AI.

C｜Declared "nonexistent"

Shop C｜Kimono rental

A 4.9 Google rating and over 500 reviews (figures as of July 2026), with strong marks from foreign visitors. Yet hand the name to ChatGPT and this is what comes back:

I couldn't confirm a shop called "(Shop C's name)."

It then recommended two other shops with similar names. Ask in Traditional Chinese, and the answer was even blunter:

There is no shop by that name in Kyoto.

Over five hundred glowing reviews — and to AI, it was as if they didn't exist.

D｜Absorbed into another business

I asked the same shop the same question again. This time the answer changed:

You probably mean (another shop's name). Rated 5.0, with over 1,600 reviews — very well regarded.

But that is a different company. Also in Kyoto, with a name one character apart, and with its own official website. In other words, unable to find Shop C, AI simply treated it as a different shop with a similar name and a fuller profile.

Note that AI's own words were "you probably mean" — it wasn't sure either. The same question producing two contradictory answers is the clearest evidence. AI doesn't know Shop C; it just lacks enough basis to judge, so every time it guesses.

Here's what's frightening: Shop C's five hundred-plus positive reviews may right now be driving traffic to a competitor — and the shop itself has no way of knowing.

What separated these four outcomes was not size.

The shop with more reviews was the one AI couldn't see.

[Two directions of the source] Named queries need your own voice; discovery queries need other people's

The real difference lies in where each shop puts its information. Let me add one more shop, E, for comparison. Shop E has no official website of its own — only a page on a booking-SaaS platform — and yet AI still recognized it.

What was registered as Shop C's "official site" was just a product page on a booking platform. But the subject of that page is the platform, not Shop C. So in AI's eyes, Shop C isn't a shop — it's one product listing inside some booking website.

From this, one important principle emerges:

What AI recognizes is an "entity," not keywords and not a mere web page. Only content that speaks with itself as the subject, in a form that can be cited, becomes an entity AI can recognize.

By this measure, Shop C never became an "entity," because it exists only as a product inside a platform. This does not mean every shop must build its own website — having an independent domain is not the point. Shop E has only a SaaS page, but because the subject of that page is Shop E, AI could treat it as a shop. What AI needs is your own voice, in a form that can be cited.

But AI changes where it looks depending on how you ask

This — beyond the "does language matter" question raised earlier — was one of the most notable findings this time.

When you ask directly, "What kind of shop is Shop A?", AI almost always cites the official website. Prices, plans, durations — all from the shop's own content.

But switch to a discovery-type question, "Any recommended tea-ceremony experiences in Kyoto?", and the sources change completely:

• Large travel guides (Lonely Planet, etc.)
• Official tourism sites (Kyoto City, the Japan National Tourism Organization)
• "Top N picks" articles aimed at inbound travelers
• Reddit

Hardly any business's official website was cited at all.

These are two entirely different problems. Polishing your official site is about being understood correctly once someone already knows your name. But when someone simply asks "What tea-ceremony experiences are there in Kyoto?", what AI actually looks at is whether other people have mentioned you.

[Scope] The winner differs by language — but it depends on the category

Asking the same "Please recommend a XX experience in Kyoto" in four languages produced opposite results across different categories.

Tea ceremony: the same three shops in all four languages

Japanese, English, Traditional Chinese, Simplified Chinese — whichever language you ask in, AI recommends essentially the same three shops. The difference between languages is very small.

Kimono: not a single shop appeared in all four languages

Three points stand out.

First, the top three shops recommended in Japanese (⑥, ⑦, ⑧) never appeared once in any of the other three languages.

Second, the industry's largest brand (⑬) surfaced only in Simplified Chinese. Even asking in Japanese, "Which kimono rental in Kyoto would you recommend?", AI won't mention it.

Third, Shop C — the one that couldn't even be identified in the named-query section — scored zero across all four languages, which is why it's absent from the table above. And more cruelly, the competitor AI confused Shop C with appeared in all three of English, Traditional Chinese, and Simplified Chinese.

Why is the gap this large?

Look at which media AI cites, and the reason becomes clear.

For tea ceremony, every language cites roughly the same set of international travel media — so the recommended shops are the same. Kimono is entirely different. Japanese draws on 着物レンタル比較ナビ (Kimono Rental Hikaku Navi), English on Japan Cheapo, Chinese on MATCHA. Because the articles differ, the recommended shops differ completely.

Which shop AI recommends is decided by the media that write the "round-up articles" in that language.

And look closely at these names: japan-atlas, MATCHA, Japan Cheapo, 着物レンタル比較ナビ. None of them are major news outlets. They are mostly affiliate-marketing sites that specialize in "Top 10 ○○ in Kyoto" round-ups. As of July 2026, it is these sites that, in effect, decide what AI recommends.

So for a business, the question "should I invest in multiple languages?" has, for now, no one-size-fits-all answer. At the very least, it has to be judged by testing it for your own category.

[What to do] Two kinds of question, two kinds of solution

This survey points to four directions.

① To be described correctly when named → put citable facts out yourself

Prices, durations, supported languages, whether booking is required — write these as clear, explicit facts on your own official site. That is exactly what Shop A did. What Shop B lacked was precisely this kind of self-description that AI can read accurately. And don't assume that putting it on a booking platform is enough. That means giving up your own voice.

② To be found → appear in the articles AI reads

This isn't website work; it's PR and outreach — round-up articles, comparison sites, travel media. There's also one surprising source: Reddit. In English queries, AI repeatedly cited Reddit threads. Few Japanese businesses are paying attention to it.

③ Where to appear for discovery queries → it varies by category and language, so measure

For tea ceremony and kimono, the media AI reads were completely different. They differ by category and by language. There's no general answer. For now, the only effective method is to actually test what AI cites in your category and your target language.

④ A future possibility → own a distinctive niche with your own domain

We also observed another possibility: shops with their own domain and a clear distinguishing feature (for example, "families with children welcome") were sometimes mentioned by AI even without heavy exposure. Perhaps, if you don't want to take the media-exposure route, building up a niche that is truly your own, on your own site, can be another workable strategy for a small business.

Limitations

This was only a small, individual experiment. The sample is just ten shops, one AI model (ChatGPT), and a limited number of queries. It cannot support any statistically meaningful claim; treat it as one person's observation.

Also, AI's answers can change every time. That Shop C gave different answers on the first and second try is the plainest example. To analyze this properly, one should ask the same question dozens of times and look at the distribution.

Furthermore, many generative AIs today remember a user's past conversations and let that shape later answers. So if a shop owner asks from their own account, "What ○○ in Kyoto would you recommend?", their shop appearing in the answer proves nothing — AI may simply remember them. This survey was run with memory disabled. What it measures is not ranking, but how likely an AI that knows nothing about you is to mention you on its own.

Closing

AI search is still a field without a settled methodology. Most people, myself included, are feeling their way. Half of the hypotheses I started with didn't hold. The prediction that "a shop found in Japanese would vanish in Chinese" turned out to be true or false depending on the category. I'll keep publishing the record here, including the hypotheses that were wrong — because what this field most lacks, I believe, is not clever theory but measured data.

References & background

The idea I've been calling an "entity" here is discussed in English-language SEO as entity-based SEO, and for generative AI as GEO (Generative Engine Optimization). The common academic starting point is Aggarwal et al., "GEO: Generative Engine Optimization" (arXiv:2311.09735, KDD 2024), which frames the conditions for being surfaced by a generative engine as being retrievable, groundable, and citable. Entity recognition is generally held to run on three signals: structured data (Schema.org), consistent mentions from third parties, and a consistent identifier. Industry also circulates figures like "having a profile on certain review sites makes you more likely to be cited by AI," but these are not peer-reviewed, so I treat them here only as background, not as settled fact.

This is an independent research and information-sharing project, unaffiliated with any organization, university, or research institution.

If you'd like details on the methodology, or are interested in running a similar study, please get in touch.