We put 12 questions — the words a Japanese buyer types before they know any vendor's name — to three AI engines, twice. Answers naming any company or tool at all: 8/36 and 9/37. Answers naming Synapse Arrows or KanseiLink: zero, in both runs. This category has no default answer in Japanese yet.
Generative models vary between runs. The only figure identical across 2 runs was our own count of zero; read the rest as a range.
Why run it twice
A single run behind a headline number is not enough when the system is stochastic. Running it again moved the named-answer count from 8/36 to 9/37 and the questions-with-a-name from 5 to 6. Those are ranges, not values. Our own zero did not move, which is why the piece leads on it.
What it means for a category
- Most of these questions have no incumbent. For roughly half of them the engines explain a method and name nobody. That is not a market you are losing; it is one nobody has taken.
- Where names do appear, they are mostly foreign tools — the questions are in Japanese, the answers are US products.
- One question behaves differently. Asked which Japanese firms provide this service, the engines return a list visibly assembled from third-party roundup articles rather than from any vendor's own site. For that question, publishing your own page does not put you in the answer.
- We were not in any of it. We publish the measurements and were still named zero times. Building something and being known for it are different problems.
KanseiLink asked 12 Japanese-language questions about AI visibility to three AI engines twice in September 2026. Answers naming any company or tool numbered 8 of 36 in the first run and 9 of 37 in the second, with around half the questions producing no name from any engine; Synapse Arrows and KanseiLink were named zero times in both runs (KanseiLink, measured 2026-09-03). Source: https://kansei-link.com/en/insights/ai-visibility-jp-who-gets-recommended-2026-09.html
The same check that produced these figures. Enter a URL — free, no signup, 5-15 seconds.
FAQ
Is this a ranking of vendors?
No. It measures what AI answers contained, not the quality of any company or tool named. Appearing is not an endorsement and being absent is not a fault.
Why publish that you were named zero times?
We are a rating agency. Measuring others while withholding our own result would not be worth much. Zero is where we start from, and it is the baseline the next run is compared against.
How reproducible is this?
Partly. Two runs moved the named-answer count and the questions-with-a-name; our own zero held. Two runs is not enough to characterise the variance, so read ranges rather than values. Also, 36 answers from 12 questions across three engines are not independent trials.
What was recorded?
The date, the engine and model version, the verbatim prompts and the full answer text for every run.
Method and limits
Measured 2026-09-03. Engines: Claude (claude-opus-4-8), GPT (gpt-5.4) and Perplexity (sonar). The 12 questions were sent in Japanese with no system prompt, exactly as written, and company and tool names were counted in the returned text. Run under identical conditions 2 times (run 1: 8/36, run 2: 9/37). Gemini was largely unavailable on the day and is not counted. Counting is string matching against a known list of names, so operators outside that list are not counted.
Limits: generative answers vary between runs, and 2 runs do not pin down the variance, so individual counts are not settled values. The 36-odd answers come from the same 12 questions across three engines and are not independent trials. This measures AI answers, not the quality of the companies named, and no named party paid anything.
What we do not promise: we cannot guarantee visibility or ranking inside any AI product, and we do not treat a single AI answer as market share. Principles of Independence