Monthly AI-search monitoring for brands in Japan. ChatGPT・Claude・Gemini answers in Japanese, scored on citation, Keigo & Tone Fit, structure. Free first scan.
OpenAI: ads don't influence answers, feeds don't buy ranking. The only way in is to be cited.
Who gets cited in Japanese? 1.2M-citation study: 95%+ third-party. Wikipedia, Reddit, G2: a few dozen. PR TIMES: ~12k.
Different language, different internet.
https://t.co/pv2Zkqvecl
@lilyraynyc Market-level rollouts land unevenly. In Japan the sources behind the answer skew to comparison and ranking media rather than vendors' own sites, so the brand's page loses the click twice — once to the overview, once to the aggregator cited inside it.
@sengineland Worth noting how lopsided this gets by market. In Japanese B2B, the real product detail usually sits behind a 資料請求 form, so the open-web version of a vendor's own story stays thin — and the Japanese answer gets built from comparison media instead.
@lilyraynyc Worth taking that inventory per language. Help centers and docs are often English-only, or a thin translated copy on the JP side, so the Japanese answer gets built from local media and review sites rather than your own documentation. Same brand, two different source sets.
@sengineland Agreed, and the split matters. In our scans the Japanese-language prompt alone (no mention of Japan) is where the source set changes most — that's the language effect. Adding "in Japan" shifts which brands get named, which is the market effect. Two different things to fix.
Two numbers from this month:
45.5% of AI Overview sources are new every update (Ahrefs).
50.3% of Japanese AI shoppers switched brands over an AI answer.
A scan is a snapshot. Japanly AEO now monitors monthly: 5 JP questions, 3 engines, every month.
https://t.co/IfRNGURRKM
@Moz The surface set is language-specific. In Japanese answers those slots go to note, PR TIMES, Qiita, はてな — Reddit and Wikipedia rarely appear. A brand can own the English surfaces and still be absent in JP. Worth auditing per language, not per region.
@lilyraynyc@ahrefs Worth splitting this by query language. Sites with no Google footprint can hold citations in one language and vanish in another — the source set for the same intent often differs across languages, so a drop can be a swap on one side only.
@top5seo Worth running the same prompt in another language. The prior shifts — different brands assumed, different site: guesses, different claims the model wants to verify. Inspecting the plan per language exposes gaps a translated query never shows.
@aleyda@SeenSureHQ Access is the floor, not the ceiling. Even where crawler access is identical, the cited host set for the same intent diverges by query language — the Japanese answer often cites different pages from the English one. Language would be a useful split in the next round.
@Jingg_n_Tonic@CapitalOne The underused-lever read is right. Caveat: some levers are language-bound. “Authoritative” in Japanese lives in register, not word choice, and “technical terms” splits across kanji/katakana/English forms of the same term. The set needs re-deriving per language, not translating.