TL;DR — KEY OBSERVATIONS
- Cross-LLM divergence is structural: the same company is cited at materially different rates across Claude, Gemini Pro, and GPT. No single ranking predicts the others.
- Japan-US AIO adoption gap is 55 percentage points (US 73% vs Japan 18% structural markup adoption). Foundational gap, not a tail-end refinement gap.
- Three LLM behavior signatures observed: Claude — narrow, institutional, high-precision. Gemini Pro — dispersed, mid-tier-inclusive. GPT — between, institutional-weighted.
- Asymmetric early-mover opportunity for Japanese mid-market: foundational AIO markup adoption captures citation share at lower cost than will be available once peer adoption catches up.
- Three foundational steps close ~half the gap: Schema.org Organization + sameAs Wikipedia; Article + FAQPage schema; Wikipedia entity maintenance.
Background — Why This Study
AI Overview (AIO) citation behavior is becoming a primary visibility surface for B2B and consumer search. Google AI Overviews, ChatGPT's citation responses, Perplexity's source attribution, and Claude's conversational sourcing have collectively shifted a meaningful share of search-result attention from traditional SERP positions to AI-generated answers with linked sources.
The question this study addresses: are companies adopting the structural markup that supports AIO citation eligibility? And does adoption translate into actual citation share?
The comparative frame — US AI/LLM industry vs Japan M&A intermediary industry — was chosen for two reasons. First, both are knowledge-intensive industries where AIO citation visibility plausibly affects business development pipeline. Second, the two industries differ on dimensions other than AIO adoption (regulatory, capital, growth stage), allowing the AIO adoption gap to be examined against that broader context rather than as an isolated metric.
Methodology and Company Selection
Company Selection
- US AI/LLM cohort (15): OpenAI, Anthropic, Google DeepMind, Microsoft AI (via Microsoft), Meta AI, Amazon AWS AI, Salesforce Einstein, Adobe Firefly, Nvidia, Palantir, Snowflake, Databricks, Hugging Face, Cohere, Mistral
- Japan M&A intermediary cohort (15): Japan M&A Center (Nihon M&A Center), M&A Capital Partners, Strike, Onward Bridge, Integral, AGS Consulting, Nihon Houmu Solution, Tokyo M&A Park, and other Japan-based intermediation firms — all major listed and significant private firms in the Japanese mid-market M&A intermediary space
AIO Markup Adoption Measurement
For each company, we examined the corporate website root and at least 3 representative content pages for the presence of:
- Schema.org Organization markup with sameAs cross-references
- Article schema on flagship content with author and date
- FAQPage schema on relevant pages
- Wikipedia entity presence (corporate entity in en.wikipedia.org or ja.wikipedia.org)
- Entity disambiguation via Wikidata / authoritative database cross-references
A company was classified as "adopting" if at least items 1, 2, and 4 were present at the corporate site. The 73% / 18% headline figures refer to this composite adoption.
Cross-LLM Citation Behavior Measurement
For each company, we issued matched-query prompts to Claude (3.5 Sonnet), Gemini Pro (Deep Think where applicable), and ChatGPT (4o), asking for industry overview and competitor positioning where the company would naturally surface as a citation source. Citation frequency and source-type composition were recorded per LLM.
Part I — US AI/LLM Industry Observations
The US AI/LLM cohort showed substantially higher AIO markup adoption:
- Schema.org Organization with sameAs: 14 / 15 (93%)
- Article schema on flagship content: 13 / 15 (87%)
- Wikipedia entity presence (en): 15 / 15 (100%)
- Composite adoption (1+2+4): 11 / 15 (73%)
The pattern is consistent with what one would expect from a sector where founders and corporate communications teams are themselves AI-literate and have internalized structured-data practices. The single non-adopter in the cohort is in late-stage adoption (Schema markup present but inconsistent across content depth).
Cross-LLM citation behavior within the US cohort shows clear leader stratification: companies with both high AIO markup adoption and significant English-language press coverage (OpenAI, Nvidia, Anthropic) capture citation rates 3-5x higher than companies with AIO markup adoption but lower press coverage. AIO markup is necessary but not sufficient — it is the eligibility layer, not the ranking signal.
Part III — Cross-LLM Citation Patterns Observed
The three LLMs surface citations differently:
| LLM | Citation Pattern | Source-Type Bias |
|---|---|---|
| Claude (3.5 Sonnet) | Narrow, high-precision | Institutional / regulatory / academic-leaning |
| Gemini Pro | Dispersed, broader coverage | Mid-tier inclusive; pulls in secondary publications |
| GPT (ChatGPT 4o) | Between Claude and Gemini | Institutional-weighted but accepts secondary sources |
For a company seeking AI citation visibility, the practical implication is that the same content optimization will not produce uniform results across the three LLMs. Schema.org Organization markup helps with all three but matters most for Gemini's broader citation distribution. Wikipedia entity presence matters most for Claude's narrow, institutional preference. English-language major-press coverage matters across all three but disproportionately for ChatGPT's institutional-weighted pattern.
Three Foundational AIO Adoption Steps
For Japanese mid-market companies seeking to close the US-Japan adoption gap, three foundational steps close approximately half the gap:
Step 1 — Schema.org Organization Markup with sameAs
Implement Organization schema on the corporate site root with sameAs cross-references to Wikipedia / Wikidata entries, LinkedIn corporate page, corporate registry source, and authoritative industry databases. This is the foundational entity-disambiguation signal that AI systems use to identify the corporate entity.
Step 2 — Article and FAQPage Schema on Flagship Content
Add Article schema (with explicit author identification and publication date) to flagship content pages — research reports, industry whitepapers, regulatory commentary. Add FAQPage schema to FAQ-structured content, enabling extraction-style citation. Both schema types are AI-extraction-friendly and convert content from passive presence to citation-eligible structured assets.
Step 3 — Wikipedia Entity Presence Maintenance
Maintain accurate Wikipedia entity for the corporate entity — both Japanese-language (ja.wikipedia.org) and English-language (en.wikipedia.org) where notability standards are met. Wikipedia entity presence is the highest-weighted single signal for AI citation eligibility, particularly for Claude's institutional-preferring citation pattern. Maintenance involves accurate fact-table, current corporate data, and avoiding Wikipedia's notability and self-promotion violations.
Limitations and Method Disclosures
- 30-company sample is statistically small relative to either country's full universe; selection is purposive (industry-leadership focused) rather than representative
- Cross-language LLM citation behavior may be confounded by training-corpus language composition rather than purely by company adoption choices
- The AIO landscape is changing rapidly — observations reflect May 2026 snapshot and may shift with model updates
- Japanese M&A intermediary market structure differs from US AI/LLM industry structure in ways beyond AIO adoption (regulatory environment, capital intensity, growth stage) that may co-vary with citation behavior
- One Agent-execution error in initial AIO aggregation was identified and retracted before publication; the published figures reflect the corrected run
Independent replication is invited. The methodology, company list, and cross-LLM query prompts are available on request.
Frequently Asked Questions
Q1. What does the study find?
Two structural patterns: cross-LLM divergence (the same company cited at materially different rates across Claude, Gemini Pro, GPT — no single ranking predicts the others) and Japan-US adoption gap (US 73% vs Japan 18% structural markup adoption). Combination creates asymmetric early-mover opportunity for Japanese companies investing in foundational AIO markup.
Q2. How do Claude, Gemini Pro, and GPT differ?
Claude: narrow, institutional, high-precision. Gemini Pro: dispersed, mid-tier-inclusive. GPT: between, institutional-weighted. Same content optimization does not produce uniform results across the three. Schema.org helps Gemini's broader distribution most; Wikipedia entity matters most for Claude's institutional preference; English-language press matters across all three.
Q3. Why is the US-Japan 55 percentage-point gap significant?
The gap is foundational, not tail-end. US mid-market has largely completed foundational adoption (73%); Japan is materially behind (18%). For Japanese companies, this creates a 3-6 month window to capture citation share at lower cost than will be available once peer adoption catches up. The window is not permanent — Japanese adoption is forecast to accelerate.
Q4. Which US AI/LLM companies were studied?
15 US-listed and major-private AI/LLM industry companies: OpenAI, Anthropic, Google DeepMind, Microsoft AI (via Microsoft listed), Meta AI, Amazon AWS AI, Salesforce Einstein, Adobe Firefly, Nvidia, Palantir, Snowflake, Databricks, Hugging Face, Cohere, Mistral.
Q5. Which Japan M&A intermediary companies were studied?
15 Japan-listed and major-private M&A intermediary firms including Japan M&A Center, M&A Capital Partners, Strike, Onward Bridge, Integral, AGS Consulting, and other Japan-based intermediation firms. Selection mirrors the US cohort size for apples-to-apples cross-country comparison.
Q6. What are the three foundational adoption steps?
(1) Schema.org Organization markup with sameAs links to Wikipedia, LinkedIn, corporate registry; (2) Article and FAQPage schema on flagship content with explicit author and date; (3) Wikipedia entity presence maintenance. Three steps close ~half the gap in 3-6 month implementation window.
Q7. What are the study's limitations?
30-company sample small relative to full universe; purposive (not representative) selection; cross-language LLM behavior confounded by training-corpus composition; AIO landscape changing rapidly (May 2026 snapshot); structural differences between US AI/LLM and Japan M&A intermediary may co-vary with citation behavior. One Agent error in initial AIO aggregation retracted before publication.
Q8. How does this relate to JFSC's broader portfolio?
Component of "Induced Intermediation" research series. Companion studies: YMYL Crawler Cost-Tier Hypothesis (339 evaluations); SERP Industry Structure 2026 (11,116 SERP × 1,720 domains, 24 industries); Cross-Border M&A 2026; Search Funds in Japan 2026. Portfolio focuses on how information-asymmetry intermediaries interact with search-engine and AI quality signals to shape outcomes for end-users.