When your name has no capital letters: our tool saw nothing in Japan or Korea
Measured 17 August 2026 across six markets. In Japanese and Korean, 100% of the names cited are in the local script - and our extractor required a Latin capital.
In this article
Yesterday the product started asking its questions in Japanese, Korean, Chinese, Arabic and Spanish. This morning we measured those markets for the first time. The first competitor the measurement found was our own tool.
The protocol is the series': six markets (Lyon, Mexico City, Tokyo, Seoul, Singapore, Dubai), two trades per market so that a gap cannot be explained by "that trade is special", all three assistants, and every question asked twice - the noise control the 15 August article made mandatory. Seventy-two answers.
What we found about ourselves
Before we had even read the answers, name extraction returned zero in Japanese, Korean, Chinese and Arabic - on samples where the names are obvious. The cause fits in one line of our parser: to be kept as a business name, a string had to contain a capital letter.
That is a Latin rule. Those four scripts have no case at all: there is no such thing as a capital letter in Japanese, Korean, Chinese or Arabic. No business name written in those alphabets could satisfy the condition, and all of them were dropped in silence.
A Japanese audit only saw a competitor if its name contained a Latin capital.
We fixed it the same morning: a character from a caseless script now stands in for a capital, Hebrew and Thai are included pre-emptively, and the Latin rule is untouched. The test that locks the fix was verified backwards - put the old rule back and four tests fail, one per script.
Why this is not only a code story
Here is what the measurement gives once the parser is repaired, on the question "the best bakeries", per market and per engine - the share of cited names written in the local alphabet:
Tokyo: 15/15, 16/16, 14/14.
Everything all three assistants name is written in Japanese.
Seoul: 11/11, 13/13, 10/10.
The same, without a single exception.
Singapore: 3/10, 8/16, 13/15.
A mixed market, like the city.
Dubai: 0/10, 12/16, 9/16.
And that first figure is the most telling of the day.
In Tokyo and Seoul, everything our extractor keeps today was invisible yesterday. Not "under-counted": invisible. A report would have rendered normally, with an empty competitor table and a score built on it.
Dubai: two markets in one city
Asked in Arabic about the best bakeries in Dubai, ChatGPT names ten places. Not one of the ten has an Arabic name: Ladurée, Paul, Baker & Spice, Le Pain Quotidien, Magnolia Bakery. Gemini, on the same question, cites twelve Arabic-script names out of sixteen; Perplexity nine out of sixteen, among them مخبز حبيب and مخبز الريف اللبناني.
A tool that can only read the Latin alphabet would therefore have seen ChatGPT's list in full, part of Gemini's, and concluded that bakery in Dubai means Ladurée. The reading defect and the market fact have exactly the same shape: what is not written in your alphabet does not get counted.
The rest of the measurement confirms what the series has shown for ten days. ChatGPT named no bakery at all in Korean or in French: it answered with questions - "what kind of bread do you prefer?", "which neighbourhood?". On the same question, Gemini names eleven in Lyon and eleven in Seoul.
What to do with this
Ask your AI-visibility tool whether it reads your alphabet.
The question sounds absurd until the day it isn't. Ours was built by people who write in Latin script, and it showed in the code.
A zero in a non-Latin language deserves a manual check.
Ask the question yourself in the app, in Japanese, Korean or Arabic, and compare with what the tool displays.
In a mixed market, measure in both scripts.
In Dubai the Arabic list and the Latin list barely overlap: that is two playing fields, not one.
Do not judge your presence by ChatGPT alone
- across our six markets it is the engine that names the fewest businesses, and sometimes none.
The limits, and one that stays open
Six markets, two trades, one morning, one model per engine: that is a sounding, not a census. We did not verify that each named establishment exists - we classified what is named, we did not audit the businesses.
And one hole stays open, so we may as well say it here. The filters that discard advice mistaken for businesses - "ask for a written quote", "check the licence" - are lists of French and English phrases. They do not cover the four new scripts. The counts above therefore mix real shops with fragments of advice, which changes nothing about the alphabet measurement (advice in Japanese is also written in Japanese) but forbids reading them as a number of businesses. That is the next job, and it is named rather than hidden.