Measurement & metrics
Hyokiyure: how Japanese spelling variants break AI measurement — and how to count brand mentions correctly
Hyokiyure (表記ゆれ) means multiple ways of writing the same name — キヤノン, キャノン, Canon. Reliable Japanese mention matching needs an alias set, language-aware normalization, kuromoji token-aware matching and a bounded literal fallback. Morphological analysis is one layer, not the sole prerequisite.
Why Japanese brand names split into variants
Japanese offers several ways to write the same name: katakana vs Latin script (スーパーランク / Suparanku), casing (Canon / CANON), official vs colloquial spelling (Canon’s official katakana uses a large ヤ — キヤノン — yet many people write キャノン as pronounced), plus abbreviations and legacy names. All refer to the same company; as strings, they are different.
AI assistants draw on varied web text and may use different spellings depending on context. Imagine asking the same question nine times: the answers might use the Latin spelling three times, the official katakana four times, and the colloquial form twice. SparkToro’s research supports the broader instability point, but its measured result was about changes in brand lists and their order, not byte-identical full answers.
What naive string matching does
Search mentions by the single string “Suparanku”, and every katakana mention is scored as “not mentioned”. Visibility looks lower than reality; in the worst case the false conclusion “zero presence on this topic” sends budget into unnecessary work.
The opposite error exists too: a short abbreviation as the search term hits unrelated common words and inflates visibility. In either direction, broken measurement breaks every decision built on it.
Designing measurement that counts
- Maintain an alias set — record the official form, katakana and Latin forms, abbreviations and relevant legacy names. Suparanku stores these display aliases per brand.
- Normalize both sides — apply the same language-aware normalization to aliases and answer text. The Japanese path uses NFKC normalization, hiragana-to-katakana folding, case folding, separator handling and guarded legal-form removal.
- Use token-aware matching — kuromoji segments Japanese text, and the matcher walks appropriate noun-token windows before comparing normalized keys. This reduces unsafe substring matching.
- Keep a literal fallback — bounded raw-alias regular expressions catch multiword Latin names, punctuation and other surfaces the token walk misses. The fallback is part of the actual matcher, not an exception handled by hand.
- Treat ambiguous aliases separately — an alias that is also an ordinary word needs context confirmation; unambiguous aliases keep deterministic recall.
Summary
Hyokiyure is especially acute in Japanese, although Korean, Russian and other languages have their own variant patterns. Reliable counting is a pipeline: aliases, normalization, token-aware matching, literal fallback and ambiguity handling. Even then, mention matching only answers whether a brand was detected correctly; trustworthy visibility estimates also require repeated sampling, because one run is misleadingly precise.
Related terms: Hyokiyure, Visibility score, Sentiment analysis