AI Visibility Does Not Cross Borders Evenly: What 30 Days of Japanese LLM Tracking Shows About Global Brands

Ask ChatGPT in Japanese which CRM a small business should use, and Salesforce comes back first — more often than any Japanese product does. Ask the same model, in the same language, which cloud accounting service to use, and no foreign brand makes the top three at all.

Both answers are generated by the same models, in the same language, on the same day. The difference is not about which company is bigger. It is about which categories let a brand's reputation travel across a language boundary — and that turns out to be the most under-discussed variable in Generative Engine Optimization (GEO), also called AIO.

The Data

The figures below come from daily tracking of ChatGPT, Gemini and Claude in Japanese, aggregated over the 30-day window from 10 July to 9 August 2026. Shares are calculated across the tracked brand set in each category — the listed figures sum to 100% — so read them as share of mentions among tracked brands, not as an absolute probability of appearing.

Cloud CRM: foreign brands take 93.5%

Brand	Origin	Mention share
Salesforce	US	37.0%
HubSpot	US	31.0%
Zoho CRM	India	25.5%
kintone	Japan	6.4%

kintone is a Cybozu product with a large installed base among Japanese SMBs. In Japanese-language AI answers it collects 6.4% while three non-Japanese vendors split the rest. Whatever this measures, it is not domestic market position. Full category detail is published at this CRM ranking page, including the note that Claude collection failed on three days in the window and those days were excluded from the averages.

Cloud accounting: foreign brands take 0% of the top three

Brand	Origin	Mention share
freee	Japan	41.4%
Yayoi	Japan	31.3%
MoneyForward	Japan	27.3%

QuickBooks and Xero are far larger businesses globally than any of these three. Neither appears in the published top ranks.

Hotel booking: a two-thirds / one-third split

Brand	Origin	Mention share
Booking.com	Netherlands	26.4%
Expedia	US	23.2%
agoda	Singapore	17.5%
Rakuten Travel	Japan	13.8%
Jalan	Japan	12.0%
Ikyu	Japan	7.1%

Foreign OTAs take 67.1% against 32.9% for domestic platforms — despite Jalan and Rakuten Travel holding contracts with the long tail of Japanese ryokan and business hotels that the global OTAs historically struggled to sign.

Why the Same Language Produces Opposite Outcomes

The mechanism is corpus, not commerce. An LLM's answer set reflects what has been written about a category and indexed — comparison articles, review platforms, documentation, forum threads, trade press — weighted by how consistently a brand appears in the structures those sources use.

That produces a testable split:

• Corpus-portable categories. The product is functionally the same everywhere, and the global brands have a decade of English-language comparison content, developer docs, integration guides and G2-style listings behind them. CRM and hotel booking both fit. Much of that authority survives translation into a Japanese answer, because the underlying entity is the same entity.
• Locale-locked categories. The product is defined by domestic rules — tax filing formats, invoicing law, banking connections, licensing regimes. Cloud accounting is the clearest case: an accounting tool that cannot produce a Japanese consumption-tax filing is not a candidate answer, no matter how much has been written about it in English. The corpus that matters is domestic, and only domestic vendors are in it.

Most categories fall between the two poles, and where a category sits is largely predictable in advance. If a competent buyer in Country B could use the Country A product unmodified, expect the foreign brands to carry. If they would need a locally compliant version, expect them not to.

Four Consequences for Anyone Running GEO in More Than One Market

1. English-language results tell you nothing about your other markets

This is the practical trap. A GEO dashboard configured out of the box samples English prompts against US-routed models. If you sell into DACH, Japan, Brazil or Korea, that dashboard is measuring one market and reporting it as your AI visibility. The prompt language, and ideally the geographic routing, has to match the market you actually care about — and each language is a separate measurement stream with its own baseline and its own competitor set.

2. Foreign AI visibility is a cheap market-entry signal

If your brand already appears in Japanese-language or German-language answers to buying questions in your category, you have demand-side awareness in that market that you did not pay for, sitting ahead of any local sales motion. That is a real input into expansion sequencing, and it costs a few dozen queries to check. Salesforce's 37% is exactly this dynamic showing up as a number.

3. Domestic incumbents are more exposed than their market share suggests

If you run a local business in a corpus-portable category, foreign competitors can occupy your AI answers without opening an office. kintone at 6.4% is a warning shot in a market where its actual commercial position is far stronger. The defence is the same as the offence: get into the comparison content, review platforms and structured listings that models synthesise from, in your own language.

4. Answer slots are scarce, so second place in a locale-locked category still pays

An AI answer typically names three to five brands. In cloud accounting, three domestic vendors split the entire top tier at 41.4 / 31.3 / 27.3 — genuinely contested, and worth fighting for. In a category where one brand holds 60%+, the calculation is different. Check your category's shape before budgeting; the free public rankings covering 56 categories and 291 tracked brands are a reasonable way to see which distribution you are dealing with.

A Minimum Viable Measurement Protocol

Before buying anything, you can validate the problem in an afternoon:

1. Write five buying-intent prompts in the target language — not translations of your English prompts, but the phrasing a local buyer would actually use.
2. Run each against ChatGPT, Gemini and Claude, three times per model, in fresh sessions with memory and personalisation off.
3. Record every brand named in each response. That is 45 observations per prompt set.
4. Calculate the percentage of responses naming each brand. Repeat weekly.

Two caveats. Single observations are meaningless — LLM outputs are non-deterministic, and only the aggregate distribution is stable. And a logged-in consumer account is not a clean instrument: personalisation, memory and region routing all bias the result, which is the main reason continuous tracking gets automated rather than done by hand. The underlying framing, including why the competitor set in AI answers is defined by the customer's job rather than by industry, is covered in this AIO fundamentals explainer.

Tooling, and What to Check Before You Buy

The AI visibility category attracted over USD 300M in funding between mid-2025 and spring 2026, and the products have diverged more than their marketing suggests.

Tool	Entry price	Best suited to
Profound	From USD 499/mo	Enterprise programmes; broadest surface coverage, 10+ engines including Perplexity, Copilot and AI Overviews
Peec AI	From USD 95/mo	Marketing teams wanting source attribution and competitor benchmarking
Otterly.ai	From USD 29/mo	Cheapest entry point; the notable feature for this article's purposes is coverage across 50+ countries and languages
Mieruca GEO	From JPY 49,800/mo + setup	Japanese market, full-funnel AI attribution; two models (ChatGPT, Gemini), annual contract
AIOPulse	JPY 19,800/mo per site, 5 keywords	Japanese market, time-series mention share across three models daily; month-to-month, no free trial

For a multi-market programme the questions that matter are narrow and easy to get wrong: which prompt languages are supported natively rather than by translating your English prompts; which models are sampled, since a tool covering only ChatGPT and Gemini is blind to a third of the answer surface; how many samples per prompt per day, which determines whether a week-on-week move is signal or noise; and whether historical data is retained long enough to correlate movements with model releases. The two Japanese options above differ on several of these — the vendor's own side-by-side comparison with Mieruca GEO is a useful starting point, read with the obvious caveat that it is written by one of the two parties.

FAQ

Does this mean we should localise content into every target language before measuring?

No — measure first. In corpus-portable categories you may already have visibility you did not build deliberately, and localisation budget is better aimed at the categories and markets where the number comes back near zero.

Is a 30-day average long enough to act on?

For establishing a baseline, yes. For attributing a change to something you did, you want the daily series rather than the average, overlaid with both your release timeline and known model updates.

Our category is clearly locale-locked. Is GEO still worth funding?

Often more so. Locale-locked categories have smaller, domestic competitor sets and thinner comparison content, which usually means the cost of moving your share is lower than in a category where you are competing against a decade of accumulated English-language authority.

Do these Japanese figures generalise to other non-English markets?

The mechanism should; the specific numbers will not. The corpus-portable versus locale-locked distinction is about product and regulation, not about Japan. But every market has its own incumbents and its own volume of local-language content, so the only way to know your German or Korean numbers is to measure them in German or Korean.

Takeaway

GEO is usually discussed as one number per brand. The data above suggests it is at minimum one number per brand per language, and that the gap between them is largest exactly where it matters most — in categories where a global competitor's English-language authority quietly carries into a market they have barely entered. If you sell in more than one language, the first thing worth checking is not your content strategy. It is whether you have ever measured the other market at all.

Disclosure: AIOPulse is developed by Marketing Cloud Lab, which also operates this site. Competitor figures for Profound, Peec AI and Otterly.ai are from publicly published pricing as of August 2026 and should be re-checked before purchase.