日本語EN
Parallel World
Scan your site free

B. Japanese buyers

How business buyers in Japan use AI apps to find and choose a service, what they check on the official site, and which sources they trust.

Built from our knowledge base at every deploy. 31 principles, 200 claims. Newest note: 2026-11-12.

日本語Handbook index

Each principle is followed by the claims it rests on. Evidence levels are explained on the handbook index.

Show claims by evidence level

B1. AI to make the list, the official site to check

Japanese B2B buyers use AI at the start (framing the problem, building a shortlist, a comparison table), then check each vendor's official site before contacting them.

Evidence for the principleExternal source

External sourceRecorded 2026-10-14, re-check by 2026-12-14

Japanese surveys from 2026 report that 85 to 91% of B2B buyers use generative AI when choosing a service, 53.8% ask AI for comparison tables, and 56.8% want prices before meeting sales; but only 1.6% find AI alone enough, and they verify on official sites.

Basis
Five Japanese surveys from 2026 (n=300 to 9,278), all by vendors of related services, as summarised by a GEO tool vendor's blog.
Does not prove
Every survey was run by a company selling related services; questions and samples differ, so numbers cannot be added up. No survey covers buyers of LLMO measurement tools or site owners specifically. We have not read the original surveys.
Sources
What a site owner can do

Put concrete prices, comparison points and case studies on your official site, readable without filling in a form, because buyers shortlist with AI and then check your own pages before contacting sales.

Also underB3D9

External sourceRecorded 2026-10-16, re-check by 2027-01-16

Many visits from AI answers arrive without a referrer (vendor figures: 70.6% and 35.7%); ChatGPT on the web adds a utm_source tag naming ChatGPT to cited links, while the ChatGPT and Claude mobile apps drop the referrer (April 2026 snapshot).

Basis
A web analytics vendor's test (April 2026) and a newsletter tracing the 70.6% figure to one vendor's blog post; read 2026-10-16.
Does not prove
The two figures differ twofold, both come from analytics vendors and mostly European and US sites, not Japan. Nobody has measured how much 'direct' traffic is really from AI. Apps change referrer behaviour with new releases.
Sources
What a site owner can do

Count visits that carry an AI app's referrer or ChatGPT's utm_source tag as a lower bound on AI-referred traffic, not the total. A zero does not prove nobody came from AI.

Also underC3

External sourceRecorded 2026-10-16, re-check by 2027-01-16

In a July 2026 Japanese survey, 44.6% of marketers used AI tools or AI Overviews to gather information (22.6% in December 2025), their most-used channel, ahead of industry media (36.6%); those 'not interested' in LLMO fell from 49.6% to 3.4%.

Basis
A tracking survey by PRIZMA of 1,032 marketers, 2026-07-02 to 2026-07-06, as reported by CommercePick; original not read.
Does not prove
Marketers in general, not buyers of LLMO measurement tools. The survey may have been commissioned by an LLMO service seller (sponsor unknown). Multiple-choice answers. The report may have left out channels; no figure for social media was given.
Sources
What a site owner can do

Ask ChatGPT and other AI apps the questions your buyers would ask about your category, note whether your company appears, and make sure your key pages are open to AI search bots in robots.txt.

Our lab observationRecorded 2026-10-14, re-check by 2026-11-14

From 2026-10-01 to 2026-10-14, pw.migurei.ai received no human visit from an AI app: zero visits referred by ChatGPT, Perplexity, Claude, Gemini, Copilot or Bing, and none tagged with utm_source.

Basis
Our own server logs on pw.migurei.ai, 2026-10-01 to 2026-10-14; a second count to 2026-10-16 gave the same result.
Does not prove
Some AI apps may hide the referer (ChatGPT usually adds utm_source=chatgpt.com, but not every app does). Our logs only see requests that reach our server. The site had no analytics then.
Sources
What a site owner can do

Before assuming AI apps send you visitors, check your logs or analytics for referers from AI apps and for utm_source=chatgpt.com.

Our lab observationRecorded 2026-10-16, re-check by 2026-11-19

Most 'human' page views our counter recorded on pw.migurei.ai in October 2026 came from our own tools, team and scripts, with 0 visits from AI apps. In the week 2026-10-29 to 2026-11-05, purchase pages showed no rise in outside visitors.

Basis
Our server logs on pw.migurei.ai, 2026-10-02 to 2026-11-05, filtered by user agent, referrer and our team's IPs.
Does not prove
Team traffic is partly inferred from the logs, so team devices on other networks count as outsiders. A browser-like visit may still be a bot. Viewing a pricing or contact page is not purchase intent.
Sources
our own data, no outside link
What a site owner can do

Before trusting human visit counts, exclude your team's IPs, your own monitoring and test tools' user agents, and link-preview bots. Then look at what is left by page and by country.

Our lab observationRecorded 2026-11-01, re-check by 2026-11-15

We first counted 8 visits from Google search to pw.migurei.ai from Japan (2026-10-26 to 2026-10-30), mostly to the company page; a recheck found 6 of 8 were likely automated and Search Console showed 0 Japanese impressions, so this does not hold.

Basis
Our own server logs on pw.migurei.ai from 2026-10-24, and Search Console data for 2026-10-02 to 2026-10-29, read 2026-10-31.
Does not prove
Search Console hides rare queries (all 12 worldwide impressions were hidden); 8 visits is small; Google referrers carry no query. Automated traffic is inferred from signs such as HTML-only loads and a headless browser, not verified. Visitors may have sought a namesake or the parent company.
Sources
our own data, no outside link
What a site owner can do

Before writing content for visitors who seem to come from Google, check in your logs whether they load images and JavaScript, and compare with Search Console impressions for the same pages and country.

Also relevant here

B2. Asking by category name

Buyers search by category names (LLMO対策, AIO, GEO, AI検索最適化) and read SaaS comparison pages that use those same names.

Evidence for the principleHypothesis(partly: external source)

External sourceRecorded 2026-10-10, re-check by 2027-01-10

Japanese SaaS comparison sites file services that optimise for AI answers under the category 'LLMO対策', alongside the terms AIO, GEO and AI検索最適化, and these terms have no agreed definition yet.

Basis
An ITトレンド comparison of 15 LLMO services (updated 2026-10-09), a GEO tool vendor's blog, and Google's SEO Starter Guide, read 2026-10-10.
Does not prove
Both Japanese pages come from sellers of such services, not search volume data. No data shows which words Japanese users type into AI apps. Google's advice is about Google Search, not AI apps.
Sources
What a site owner can do

If your service fits the category, use the words buyers see there (LLMO対策, AIO, GEO) truthfully in your title, description, h1 and visible text, and call your product what the category calls it (for example a tool or platform).

External sourceRecorded 2026-11-01, re-check by 2027-01-31

Google Trends Japan (12 months to 2026-10-18): 'LLMO' rose from 2.2 to 4.0, 'GEO対策' reached 1.0, 'AI Overview' fell from its peak of 100 (week of 2026-06-28) to 1 from 2026-08-09, and 'SEO対策' dropped from 9.2 to 5.8.

Basis
A Markenote article on Google Trends (2026-10-31), and our own re-check on Google Trends on 2026-11-04 (LLMO 3.8, GEO対策 0.9).
Does not prove
Trends is a relative index, not search counts; values of 1 to 10 sit on a scale set by a term's very high peak. People searching on Google may not use the same words when asking AI. Site owners may still search with technical words like GPTBot or robots.txt.
Sources
What a site owner can do

In Japanese titles, headings and FAQs, use 'LLMO' and Google's Japanese name 'AI による概要' rather than 'GEO' or 'AI Overview'. Avoid 'AIO' as your main term; it has other meanings.

Our lab observationRecorded 2026-11-04, re-check by 2026-11-18

For two Japanese statistics questions, free ChatGPT wrote its web search query in plain English in 5 of 6 runs (morning of 2026-11-03). Yet the source it cited most was a Japanese page (6 of 6).

Basis
The search queries ChatGPT displayed in 6 test runs on 2026-11-03 (free account, temporary chat, Japanese IP), with blind-graded citations.
Does not prove
That the English query was the only one: ChatGPT may also run hidden Japanese queries. That an English page would be chosen: none of our English pages was cited. One morning, 6 runs, one account. Japanese shopping questions got Japanese queries.
Sources
our own data, no outside link
What a site owner can do

For technical or statistics topics, add an English title or summary using the words ChatGPT searches with (Japan, websites, block, GPTBot, percentage, study) next to your Japanese page. Then check AI answers again.

Also relevant here

B3. Prices and evidence before meeting sales

Buyers want prices, real cases and numbers on the page; when the page lacks them, AI fills the gap itself.

Evidence for the principleExternal source(partly: hypothesis)

Also relevant here

B4. Trust in named sources and real authors

Japanese buyers trust PR TIMES, note, trade media and review sites more than a company's self-description, and AI apps in Japan cite those same sources (PR TIMES second in ChatGPT, note first in Gemini).

Evidence for the principleHypothesis(partly: external source)

Also relevant here

B5. Fear that AI says something wrong about you

Site owners check what AI says about their company and lose customers when AI gets it wrong; that worry is a reason to pay for measurement.

Evidence for the principleExternal source

Our lab measurementRecorded 2026-10-28, re-check by 2026-11-28

Asked about electricity in Miyagi Prefecture without naming a company, Gemini named the region's leading power company in 20 of 20 answers (2026-10-28 and 2026-11-03) but never linked to its website, citing comparison and affiliate sites instead.

Basis
Our graded test: two questions, 10 Gemini answers on each of 2026-10-28 and 2026-11-03, criteria fixed before the first question.
Does not prove
Small sample: one prefecture, one industry, two questions. ChatGPT linked the company's site in 2 of 2 answers on 2026-10-28 but 0 of 2 on 2026-11-03. One question asked about cheap prices, which favours comparison sites. It says nothing about sales.
Sources
What a site owner can do

Ask Gemini the questions your customers ask about your region and industry, without naming your company, and note which sites it cites. If only comparison sites appear, publish pages that answer those questions directly, with your own data.

Our lab observationRecorded 2026-11-04, re-check by 2026-12-04

Asked the price of an item in a Japanese Shopify shop (2026-10-31 to 2026-11-02), DeepSeek read the line "free shipping on orders of 22,000 yen or more" as the price in 7 of 24 tries and was right 0 times; ChatGPT was right 23/24, Gemini 24/24.

Basis
Our blind-graded test: 600 questions to four AI apps from a Japanese connection, 2026-10-31 to 2026-11-02, criteria fixed before the first question.
Does not prove
Why DeepSeek picked that line (its position, the missing JSON-LD price data, or how it summarises). It is one shop; the test covered five Shopify fashion shops, free accounts, two days. Apps took prices from the shop page mainly when the question named the shop or gave its URL.
Sources
our own data, no outside link
What a site owner can do

Put the price in structured data (JSON-LD Product with offers) and keep shipping thresholds and other yen amounts clearly apart from the item price. Then ask DeepSeek the price of one of your items and check the answer.

Our lab observationRecorded 2026-11-04, re-check by 2026-12-04

For 20 products whose maker changed a price or figure after a PR TIMES press release, at least one AI app gave an old or wrong figure as current for 6 of 20 on 2026-11-02, mostly Gemini (5 of 19), rarely ChatGPT (1 of 19).

Basis
Test questions we asked ChatGPT and Gemini on 2026-11-02 from a Japanese IP, with criteria fixed in advance and blind grading.
Does not prove
That the old figures came from PR TIMES: only 3 of 6 wrong answers cited PR TIMES. How often AI is wrong in general. The sample is 20 small-company releases, one try each, two apps; 4 of 20 reference values were already outdated.
Sources
our own data, no outside link
What a site owner can do

After you change a price or figure that you announced in a press release, publish the new figure clearly on your own site, then ask Gemini about it and check which number it gives.

Also relevant here

B6. Free trial first

Free checkers, free plans and the reports of Google and Bing are the first step; a service with no quick way to try it is skipped.

Evidence for the principleHypothesis

External sourceRecorded 2026-10-13, re-check by 2026-12-13

More than ten vendors already sell tracking of how AI apps mention and cite a brand, plus fix suggestions; at least eight have Japanese interfaces (public prices from free to about ¥130,000 a month), and Google and Cloudflare offer part of this free.

Basis
Vendors' pricing and product pages, Japanese agency blogs and press releases, and Google and Cloudflare documentation, read 2026-10-13; plus our own Search Console readings.
Does not prove
Mostly vendors' own pages or press releases; features were not tested by hand and some prices came via third parties. A Japanese interface says nothing about measurement quality on Japanese AI answers. Google's free report shows impressions per page, not which questions.
Sources
What a site owner can do

Before paying, open Search Console's Generative AI performance report (AI Overviews and AI Mode impressions per page) and, if you use Cloudflare, AI Crawl Control. Then compare paid trackers' prices and coverage.

External sourceRecorded 2026-10-14, re-check by 2027-01-14

Bing Webmaster Tools has an AI Performance report (public preview since 2026-02) showing which of your pages are cited in Microsoft Copilot, Bing's AI summaries and some partners; since 2026-10-13 a Topics Portfolio lists questions where other sites are cited and yours is not.

Basis
Two Bing Webmaster blog posts (2026-02-11 and 2026-10-13), read 2026-10-14, plus our own readings of the report, 2026-10-21 to 2026-10-30.
Does not prove
For our sites we have only seen zeros, with data starting 2026-10-15, so about two weeks, not months. Partners are not named. Grounding queries are sampled. Topics appear only with enough data. Neither post mentions an API.
Sources
What a site owner can do

Verify your site in Bing Webmaster Tools and open AI Performance. Check which date the data starts before reading a zero. Verify each subdomain separately: a verified domain did not give access to its subdomain.

External sourceRecorded 2026-11-02, re-check by 2026-12-02

A leading Japanese SEO blog (海外SEO情報ブログ, 2026-10-29) introduced a free English tool by Vanessa Fox that checks whether GPTBot, ClaudeBot and other AI crawlers are blocked, via robots.txt, requests with AI crawler user agents, and headers such as X-Robots-Tag.

Basis
The 海外SEO情報ブログ post of 2026-10-29 and its weekly digest of 2026-11-02, read 2026-11-02.
Does not prove
It does not show that the blog's readers ask AI apps this question, or that AI apps cite this blog for it. The post names no Japanese tool and gives no measurements on Japanese sites.
Sources
What a site owner can do

Check more than robots.txt: request your key pages with AI crawler user agents (or use such a tool) to see whether your CDN or server blocks them, and look at the X-Robots-Tag header in the responses.

Our lab observationRecorded 2026-11-05, re-check by 2026-11-26

In test questions phrased like a buyer's (tool comparisons, free checks, e-commerce pricing), ChatGPT and Gemini almost never named us: 2 of 18 answers on 2026-11-04, both close to our survey wording, and 1 of 18 from 2026-11-06 to 2026-11-08.

Basis
Test questions we asked ChatGPT and Gemini with free accounts from a Japanese IP, 2026-11-03 to 2026-11-08.
Does not prove
Real buyers may not use these words; we wrote the questions. Only 1 to 3 answers per question, one session per round, free accounts. The Claude part was not run yet.
Sources
our own data, no outside link
What a site owner can do

Ask ChatGPT and Gemini the questions your buyers ask and record which pages they cite. Buyer questions are already answered by list articles and vendors' own tool pages, so study what those winning pages contain.

B7. Learning from the practitioner community

People who build Japanese sites learn from real measurement posts on Zenn, Qiita and note (crawler logs, GEO trials), and those posts are also where apps find tools.

Evidence for the principleHypothesis

External sourceRecorded 2026-10-14, re-check by 2026-12-14

On one new site (reported 2026-10-11), OAI-SearchBot only read robots.txt for 6 days, fetched the first article two days after it went live, and the first ChatGPT visitor came 6 days in; 74% of ChatGPT visitors landed on article pages.

Basis
A site operator's log analysis published on Zenn and Qiita (2026-10-11), read 2026-10-14.
Does not prove
One author, a few sites, not an experiment. We do not know how OAI-SearchBot found the article URL; the post does not say whether a sitemap, Search Console or IndexNow was used. The author runs those sites and has an interest.
Sources
What a site owner can do

Publish article pages that answer real questions, not only service pages, then check your logs for OAI-SearchBot fetching those pages (not just robots.txt).

External sourceRecorded 2026-10-28, re-check by 2027-01-28

A Japanese practitioner guide advises checking by hand, once a month in about 30 minutes, whether your company appears in AI answers, using buyer-style questions (how to choose, comparisons) instead of keywords, because there is no standard way to measure it yet.

Basis
One Zenn article published 2026-10-26, read from its summary on 2026-10-28.
Does not prove
One article by one author, who may be a vendor. There is no figure on how many people follow this method, and it does not say which AI apps to check.
Sources
What a site owner can do

Write down 5 to 10 questions a prospect would ask AI (for example 'which X tool do you recommend for small companies?' and comparisons), ask them in the same AI apps each month, and record the date and whether your company is named.

Our lab observationRecorded 2026-10-29, re-check by 2026-11-12

Three articles we published on note from a new account with no followers had 6 reads and 2 likes in total by 2026-10-28, and sent one visit to pw.migurei.ai, probably from our own team.

Basis
note's read counter for our own account (one snapshot, 2026-10-28) and our server logs on pw.migurei.ai.
Does not prove
note's counter may lag; this is one snapshot of articles only 2 to 3 days old. Reads on note do not show whether AI apps read the articles, since we have no note.com logs.
What a site owner can do

Publish first on a page you control and can get indexed, and treat a post on note from a new account as a secondary copy. Check its read count after two weeks before investing more.

Our lab observationRecorded 2026-11-03, re-check by 2026-12-03

In the latest Common Crawl release (CC-MAIN-2026-43, crawled 2026-10-05 to 2026-10-18), note.com has only 36 captures, all of them robots.txt returning 403, and zero content pages, so text posted on note does not reach models through Common Crawl.

Basis
The public Common Crawl index, read 2026-11-02, plus one fetch of note.com's robots.txt on 2026-11-03.
Does not prove
Which models use which Common Crawl release. One crawl only. note's robots.txt rules do allow training bots (GPTBot, ClaudeBot, Google-Extended and others), so their own crawlers may still read it; what note's network edge returns to them is unknown.
Sources
What a site owner can do

Do not rely on note posts reaching AI training data through Common Crawl; publish important content on your own site too, and check whether your host appears in the Common Crawl index.

Also underD1

Also relevant here