We took a set of questions asking for restaurant recommendations in Ginza, wrote them in Japanese and in English across three scenarios, asked four AI systems repeatedly, and counted the sources displayed alongside each answer. The results first. In Perplexity's Japanese answers, about two thirds of the returned sources were Ikyu Restaurant, appearing in 12 of 15 answers. In Claude, the source domains cited in Japanese and the source domains cited in English did not overlap at all, not one domain. Across all four AI systems, the Japanese and English source domain sets overlapped by only about a quarter (Jaccard coefficient 0.247). Research Note 001 counted how the recommendations themselves differ by language. That difference extends to the sources displayed beside the recommendations.
Design|How the measurement was designed
We used the same question set as Research Note 001, unchanged. Three scenarios, a general dinner, an anniversary dinner, and a dinner for someone who eats neither meat nor fish, written in Japanese and in English, each asked to each AI system five times. Note 001 counted the restaurants that appeared in the recommendations. This time we counted not the restaurants but the sources displayed with each answer: which sites were shown as links or citations.
| Item | Detail |
|---|---|
| Date and location | November 1 and November 6, 2026 / Ginza, Tokyo |
| AI systems | OpenAI (ChatGPT), Google Gemini, Perplexity, Anthropic (Claude), all via API |
| Scenarios | Three: a general dinner, an anniversary dinner, a dinner for someone who eats neither meat nor fish |
| Languages | Japanese and English |
| Repetitions | Five per scenario, language and AI system; 120 answers in total |
| What we counted | The domains of the sources displayed with each answer. We made no judgment about which recommendations were better |
Before counting|The word "source" meant something different for each AI
Measuring this showed that the shape of a "source" is different for each of the four AI systems to begin with.
| AI system | The shape confirmed in measurement | What this note calls it |
|---|---|---|
| Perplexity | A list of search results returned with the answer. The count is fixed at 15, and entries are not tied to specific sentences in the answer | Returned sources |
| OpenAI (ChatGPT) | URL links the model itself wrote into the answer text | Links in the answer text |
| Anthropic (Claude) | For each passage of the answer, the page and the quoted text it rests on | Citations attached to the answer |
| Gemini | Grounding information tied to ranges of the answer text | Sources presented as grounding |
For Perplexity, all 30 answers returned exactly 15 sources. Of those 450 entries, only 13 also appeared as URLs inside the answer text. That is why this note never pools the four systems' sources into a single total. Added together for reference, they come to 974 entries, but 46% of that figure is Perplexity's fixed 15-per-answer return, a different unit from a citation tied to a sentence. Every comparison below is made per AI system.
| AI system × language | Answers | Answers with sources displayed | Source entries (the unit differs per AI system) |
|---|---|---|---|
| OpenAI Japanese | 15 | 15 | 78 |
| OpenAI English | 15 (answers obtained for 14) | 14 | 73 |
| Perplexity Japanese | 15 | 15 | 225 |
| Perplexity English | 15 | 15 | 225 |
| Gemini Japanese | 15 | 6 | 81 |
| Gemini English | 15 | 5 | 81 |
| Anthropic Japanese | 15 | 14 | 107 |
| Anthropic English | 15 | 15 | 104 |
| Total | 120 | 99 | (Not totalled: the units differ) |
Caution on the measurement dates. OpenAI and Perplexity were measured on November 1, 2026, and Gemini and Anthropic on November 6, 2026. Comparisons between Japanese and English within the same AI system are between measurements taken on the same day, so they are not affected by this. Comparisons between the OpenAI and Perplexity pair and the Gemini and Anthropic pair are, at the same time, comparisons across a five-day gap. Read the cross-system comparisons below with that in mind.
Result 01|Perplexity's Japanese answers come back centred on Ikyu Restaurant
Of the 225 returned sources in Perplexity's Japanese answers, 147 were Ikyu Restaurant (restaurant.ikyu.com), appearing in 12 of 15 answers, with several Ikyu pages lined up inside a single answer. In the English answers the centre changed: the Michelin Guide led with 64 entries, followed by Tripadvisor.
| Language | Domain | Returned source entries | Answers appeared in (of 15) |
|---|---|---|---|
| Japanese | Ikyu Restaurant | 147 | 12 |
| Japanese | Michelin Guide | 10 | 2 |
| Japanese | Vegewel | 8 | 5 |
| English | Michelin Guide | 64 | 10 |
| English | tripadvisor.com | 13 | 10 |
| English | japandining.jp | 12 | 7 |
Across all of Japanese (491 entries from the four systems combined), Ikyu Restaurant was also the largest single source at 197 entries, followed by the Michelin Guide at 24 and Tabelog at 22. Tabelog's presence in this Ginza recommendation measurement stayed modest overall, at 22 entries in Japanese and 8 in English. This result marks Ikyu Restaurant as a priority subject for further study of Japanese-language visibility. It does not show that changing a listing would change the recommendations.
Result 02|Claude's sources did not overlap between Japanese and English at all
The citations attached to Claude's Japanese answers numbered 107, centred on Ikyu Restaurant with 34 entries (in 10 of 15 answers) and Tabelog with 16 (in 5 answers). Tabelog's 16 is the strongest Tabelog showing in any cell of the four systems by two languages panel. The 104 English citations centred on English-language guides and blogs about Japanese food, led by byFood with 19 entries (in 10 of 15 answers).
The 16 domains cited in Japanese and the 17 domains cited in English had zero overlap (Jaccard coefficient 0.000). This was not an artefact of pooling the three scenarios: within each scenario separately, the general dinner, the anniversary dinner and the vegetarian dinner, the overlap was also zero. In this measurement, the source domains Claude displayed did not overlap between Japanese and English.
Reclassifying the English citations showed that 91.3% were guides or media, and restaurants' own official websites were 0%. Most of what had been left as "unclassified" was, on inspection, guides and media.
Result 03|Across all four AI systems, the cast of sources differs greatly by language
Pooling the four systems, Japanese answers carried 491 source entries and English answers 483. The largest Japanese source was Ikyu Restaurant at 197 entries; the largest English source was the Michelin Guide at 84. The overlap between the Japanese and English domain sets was a Jaccard coefficient of 0.247, and by scenario it ranged from 0.184 to 0.250. About three quarters of the domains displayed in one language never appeared in the other.
Research Note 001 counted how the restaurants recommended differ by language. This measurement shows that the language difference extends to the sources displayed beside those recommendations. Reaching users of both languages means measuring each language separately. There is no other way to see both.
- Claude0.000
- All four systems combined0.247
- OpenAI0.391
Jaccard coefficient. A value of 0 means no shared domains. A value of 1 means identical sets.
Result 04|OpenAI often shows official websites, and Gemini's sources change with the scenario
The links in OpenAI's answer text numbered 78 in Japanese and 73 in English, with the Michelin Guide the most frequent in both at 14 entries. The distinguishing feature is how many were restaurants' own official websites. On reclassification, official websites made up 48.7% of the Japanese entries and 39.7% of the English entries (sites such as Ukai, Kyubey and Tempura Kondo). The overlap between its Japanese and English domains was 0.391, the largest of the four systems: OpenAI's reading was the most stable across languages.
Gemini was the contrast. Sources were displayed in only 11 of its 30 answers. Split by scenario, sources appeared in all 10 of the vegetarian answers, in 1 of 10 general-dinner answers, and in 0 of 10 anniversary answers. Within the answers that did display sources, Japanese centred on Ikyu Restaurant with 15 entries (appearing in 5 of the 6 answers with sources) and English on HappyCow with 9 entries (appearing in all 5 answers with sources). Gemini's breakdown here describes only those answers where sources were displayed. Many of Gemini's sources also carry no title information, so their classification has a floor: treat it as a lower bound.
On classification as a whole: of the 674 entries first classified as "unclassified", re-sorting assigned 509 to guides and media and 110 to restaurants' official websites, leaving only 55 that could not be classified.
- General dinner1 of 10
- Anniversary dinner0 of 10
- Vegetarian dinner10 of 10
Answers with displayed sources, out of 10 answers per scenario.
Caution|A source being displayed and the answer being correct are different things
This measurement counted which sources were displayed. It did not judge whether the content of the answers was correct. A displayed source does not necessarily support everything in the answer, and an answer with no displayed source may still be drawing on knowledge the model learned. The displayed sources are not everything the AI used. Measuring whether a restaurant appears in recommendations, and measuring whether the content agrees with the official information, are separate jobs, and we keep them separate.
For Owners
- In the Japanese answers, the centre of the sources AI showed was, for Perplexity and Claude in Japanese, Ikyu Restaurant rather than Tabelog. For Japanese-language visibility, Ikyu is a priority subject for further study. This measurement cannot say that fixing a listing would change the recommendations.
- A "source" is a different thing for each AI system. Appearing in a fixed list of 15 returned with every answer, as with Perplexity, and being cited passage by passage, as with Claude, are different meanings of "appearing as a source".
- Measure language by language. With the casts of sources overlapping by only about a quarter, an English measurement is no substitute for the Japanese answers. The reverse is also true.
- A restaurant's own official website made up a large share of the sources, especially in OpenAI's answers. Keeping basic facts such as opening hours and prices accurate on the official side at least keeps the material AI reads in order.
Is your restaurant being measured in each language?
NIXUMA's Quick Scan checks the public sources AI reads about a restaurant, and whether those sources agree with each other, measured separately for each language. The step that asks the AI systems themselves, repeatedly and in the phrasing guests actually use, is the Full Audit.
Limits of this study
This measurement is a record of one area, Ginza, fixed question texts, and two measurement dates. It was taken via API, not through the apps, and behaviour in the apps may differ. Gemini's source breakdown rests on the 11 answers in which sources were displayed, and is a reference value. The results are exploratory. They are not a ranking of restaurants.
Next, in Research Note 003, we measure what happens when sources disagree with each other: which version does AI repeat?