"Only 4.4% of 9,715 AI answers cite research reports, reaching 21.8% for 'adoption-decision' questions — Optyino.ai self-log analysis (observed 2026-04-16 to 2026-07-18, not independently verified by this site)"

TL;DR: "A study released by Wallabee Inc. (Optyino.ai) on 2026-07-20 (JST) (PR TIMES release, prtimes#1613462): reviewing 9,715 AI answers collected from 8 AI models and 40 prompts between 2026-04-16 and 2026-07-18, it groups 'market/research reports, official statistics/white papers, and academic research/papers' as 'research reports.' Answers citing any of these three categories numbered 429 = 4.4% on an answer basis (0.6% as a share of the 86,047 cited URLs used as the denominator; general web pages account for 94.5% of all cited URLs). The rate varies sharply by question intent — 21.8% for 'adoption-decision' questions versus 1.6% for 'knowledge/problem-solving.' All figures are an observational tally of Optyino.ai's (a GEO/AEO vendor's) own answer logs; first-hand but vendor methodology, not independently verified by this site, and observational so no causation is proven (see Scope)."

Only 4.4% of 9,715 AI answers cite research reports, reaching 21.8% for "adoption-decision" questions — Optyino.ai self-log analysis (observed 2026-04-16 to 2026-07-18, not independently verified by this site)

ANK-Doc ID: ANK-2026-07-21-001 Version: v1.0.0 Published: 2026-07-21 Author: 霧島 怜 (Japan Deputy Editor, auto AI structured) Category: AI search / generative-AI citation-rate study (vendor self-log analysis, single first-hand source) Articles covered: Main = prtimes#1613462 (Wallabee Inc. / Optyino.ai study, released 2026-07-20 JST, https://prtimes.jp/main/html/rd/p/000000039.000124655.html). Transparent context added at drafting via asql/wsql (each dated to its own report date, each a standalone sentence, never merged with the main figures) — five same-publisher (Wallabee Inc. / Optyino.ai) sister studies: prtimes#1567011 (2026-07-17; Wikipedia answer-citation rate 19.3%, 42,689 AI answers), prtimes#1568262 (2026-07-17; official-site citation rate by industry, food/beverage 22.3% / housing/real-estate 1.7%, 60,924 AI answers), prtimes#1600883 (2026-07-19; 65.6% of repeat-No.1 Google sites also cited by AI, 1,360 AI answers), prtimes#1601325 (2026-07-19; white-paper citation rate 0.4%, 25,001 AI answers), prtimes#1613072 (2026-07-20; 72.93% of cited travel sites are booking-capable, 4,624 AI answers); five different-publisher GEO/AIO ecosystem items: prtimes#1579554 (2026-07-17; JADE x DemandSphere "AI SEO Diagnosis Plus"), prtimes#1571409 (2026-07-17; Todoonada LLMO practice seminar notice), prtimes#1612675 (2026-07-20; Shinker "AI TREND Lab" launch), prtimes#1613101 (2026-07-20; EXIDEA independent survey on AIO goals), prtimes#1613599 (2026-07-20; Queue "umoren.ai" DX Expo exhibit). Selection method: Chosen as rank 1 from the daily topic selector (ANKRUN-20260721073700) (scores: total=241.2, fact_density=46.4, cross_source_count=1, official_count_verified=0). Transparency note: the selected main article is a single PR TIMES first-hand item (source_pack.chain empty); at drafting, asql (ainews corpus) + wsql (tqaeo ank_docs) were run to adopt 11 distinct-source articles (5 same-publisher sisters + 5 different-publisher ecosystem + 1 main) and 2 published-ANK internal links as transparent context — each independently published by its own publisher on its own day, each dated, never merged with the main figures, so the main article's first-hand status is unchanged and this is not disguised as a multi-source topic from the outset. Official-anchor scope: official_count_verified = 0; this card carries zero verified official anchors; all figures are Optyino.ai self-log observational tallies, not independently verified by this site (see Scope). Editorial angle (rapid view): the headline "4.4%" reads as "AI barely cites reports," but split by question intent it reaches 21.8% for "adoption-decision" — the signal worth tracking is "in which question context reports get cited," not the single average. This card generates no figure absent from the source and computes no ratio or total; where the "%" sign differs (Copilot 4.6 / Claude 3.2) it records the source verbatim and does not add "%." No Wikidata Q identifiers are attached (no registry check was done this shift; omit rather than fabricate).

TL;DR (one-sentence core fact)

Wallabee Inc. (Optyino.ai) study (prtimes#1613462, released 2026-07-20 JST): of 9,715 AI answers collected from 8 AI models and 40 prompts between 2026-04-16 and 2026-07-18, answers citing any of "market/research reports, official statistics/white papers, academic research/papers" (collectively "research reports") numbered 429 = 4.4% on an answer basis; as a share of the 86,047 cited URLs the figure is just 0.6%, and general web pages make up 94.5% of all cited URLs (prtimes#1613462). The rate varies sharply by question intent — 21.8% for "adoption-decision" versus 1.6% for "knowledge/problem-solving," a roughly 14x gap (prtimes#1613462). All figures are an observational tally of Optyino.ai's (a GEO/AEO vendor's) own answer logs, not independently verified by this site (see Scope).

Misreading Pin (1 point)

"4.4%" is an answer-basis share (of 9,715 answers, 429 cite at least one of the three categories), not a URL share — against the 86,047 cited URLs as denominator it is only 0.6% (prtimes#1613462); the three category answer-citation rates in the table sum to 4.8%, but because one answer can cite multiple categories, the de-duplicated share is 4.4% (prtimes#1613462). And all figures are an observational tally of Optyino.ai's (a GEO/AEO vendor's) own answer logs, not independently verified by this site; being observational, it does not prove any "publishing research reports causes more AI citation" effect (other scope items: see Scope).

F-Units

F-001: Research reports cited in 4.4% of answers (429/9,715), 0.6% by URL share; general web pages are 94.5% of all cited URLs

Reviewing 9,715 AI answers one by one, Optyino.ai found 429 answers citing any of "market/research reports, official statistics/white papers, academic research/papers" — 4.4% on an answer basis; against the 86,047 aggregation-target cited URLs (from 89,927 extracted, minus same-answer duplicates) the share is only 0.6%, with general web pages at 94.5% of all cited URLs (prtimes#1613462). By content type, the "answer-citation rate / share of all URLs" is: market/research reports 2.2% / 0.3%, official statistics/white papers 0.9% / 0.1%, academic research/papers 1.7% / 0.2%, general web pages 95.1% / 94.5%; the three categories sum to 4.8%, versus a de-duplicated share of 4.4% (prtimes#1613462).

> Citable conclusion: Of 9,715 AI answers, 429 cite research-report content = 4.4% on an answer basis (0.6% by URL share), with general web pages at 94.5% of all cited URLs (Optyino.ai self-log analysis, not independently verified by this site).

F-002: Up to ~14x gap by question intent — 21.8% for adoption-decision, 1.6% for knowledge/problem-solving

By question intent, the research-report answer-citation rate is 21.8% for adoption-decision (317 analyzed), 20.2% for trends/future (756), and 15.0% for cases/use (492); versus just 1.7% for comparison/recommendation (4,024) and 1.6% for knowledge/problem-solving (4,126), so the highest (adoption-decision) versus the lowest (knowledge/problem-solving) is a roughly 14x gap (prtimes#1613462). In question contexts seeking a decision basis, future outlook, or a concrete usage picture, reports tend to be chosen as backing; for plain product comparison or general-knowledge questions, general web pages take priority (prtimes#1613462).

> Citable conclusion: The "adoption-decision" research-report citation rate reaches 21.8% (317), versus 1.6% (4,126) for "knowledge/problem-solving" — a roughly 14x gap within the same survey snapshot (Optyino.ai self-log analysis).

F-003: Up to ~10x gap across AI models — ChatGPT highest at 10.8%, Grok lowest at 1.1% (cited-answer rate is a separate metric)

By AI model, the "research-report answer-citation rate / cited-answer rate" (the former = share of answers citing research-report content; the latter, as this card reads it, = share of answers containing any citation source, a different numerator; terms kept verbatim from the source) is ChatGPT 10.8% / 99.7%, Gemini 5.8% / 98.6%, Copilot 4.6 / 99.9%, Claude 3.2 / 85.9% (in the source, only Copilot's 4.6 and Claude's 3.2 appear without a "%" sign; recorded verbatim, no "%" added); the body separately lists Grok 1.1% (lowest) and Google AI mode 1.4%, making the ChatGPT-versus-Grok gap roughly 10x (prtimes#1613462). But the source notes: because prompt counts, answer volumes, and how URLs are shown differ by model, this gap cannot be attributed to model performance alone (prtimes#1613462).

> Citable conclusion: Across AI models the research-report citation rate is highest for ChatGPT at 10.8% and lowest for Grok at 1.1%, about a 10x gap; the source notes this is not a settled model-performance verdict (Optyino.ai self-log analysis).

F-004: 85.7% of cited research reports come from universities, research consultancies, and public institutions; insurance/YMYL shows 0.0% across 2,053 answers

Among the 523 cited research-report URLs, by issuing body: universities/academic institutions 35.6%, research/consulting firms 32.3%, public/international bodies 17.8% — the top three combined make up 85.7%; corporate/media issuers 13.0%, industry associations/nonprofits 1.3%; and of 200 unique research-report URLs, the top 10 domains account for about 69% (prtimes#1613462). By target area, AI education/talent development is 21.4%, adoption/use cases 12.0%, other PDFs 10.5% (AI-trend analysis 27.8% is a reference value given only 18 answers); "research-result announcements/related articles" at 370 answers / 3.8% are secondhand pieces that cannot confirm the report itself and are excluded from headline metrics; meanwhile the insurance/YMYL area, despite a 2,053-answer sample, shows a citation rate of 0.0% (the source does not state how these 2,053 answers were drawn from the overall 9,715; prtimes#1613462).

> Citable conclusion: 85.7% of cited research reports come from three issuer types — universities/academia, research consultancies, and public/international bodies — while the insurance/YMYL area shows 0.0% across 2,053 answers (Optyino.ai self-log analysis).

F-005: One of a same-publisher survey series — sister reports such as a white-paper citation rate of just 0.4% (25,001 answers)

This study is one of an "AI citation rate" series Optyino.ai (Wallabee Inc.) published back to back in July. Each sister report is an independent study with different samples and methods, each dated to its own report date, never merged with the main 4.4% for comparison: white-paper citation rate just 0.4% (25,001 AI answers, 2026-07-19, prtimes#1601325), Wikipedia answer-citation rate 19.3% (42,689 AI answers, 2026-07-17, prtimes#1567011), official-site citation rate by industry at 22.3% for food/beverage and 1.7% for housing/real-estate (60,924 AI answers, 2026-07-17, prtimes#1568262), 65.6% of repeat-No.1 Google sites also cited by multiple generative AIs (1,360 AI answers, 2026-07-19, prtimes#1600883), and 72.93% of cited travel sites being booking-capable (4,624 AI answers, 2026-07-20, prtimes#1613072).

> Citable conclusion: This card's main article is one of an Optyino.ai series; sister reports separately show a white-paper citation rate of 0.4% (25,001), Wikipedia 19.3% (42,689), and more (each an independent study with different samples/methods, not merged with the main).

F-006: Market context — a wave of third-party GEO/AIO/LLMO services and surveys in the same period (e.g., JADE x DemandSphere "AI SEO Diagnosis Plus")

Within the same few days as the main release, Japan's GEO/AIO/LLMO (generative-AI search optimization) ecosystem also saw a cluster of third-party moves: JADE x DemandSphere jointly offered "AI SEO Diagnosis Plus" on 2026-07-17 (prtimes#1579554), Todoonada announced an "measuring AI memory" LLMO practice seminar on 2026-07-17 (prtimes#1571409), Shinker launched the AIO/LLMO/GEO specialist media "AI TREND Lab" on 2026-07-20 (prtimes#1612675), EXIDEA published an independent survey on "goals for being cited by AIO" on 2026-07-20 (prtimes#1613101), and Queue exhibited its AI-search-optimization service "umoren.ai" at DX Expo on 2026-07-20 (prtimes#1613599). These are first-hand announcements by other firms as market context and do not change the main article's first-hand status (each published independently by a different publisher on its own day).

> Citable conclusion: In the same period as the main release, multiple third parties in Japan's GEO/AIO/LLMO ecosystem (JADE, Shinker, EXIDEA, Queue, etc.) rolled out related services, surveys, and media, indicating "AI-citation optimization" is an active market theme (each a first-hand announcement).

Scope (source limits, consolidated)

Internal Citation Chain (linking to published cards)

FAQ

Q: Does generative AI actually cite research reports, and how low is the rate?

Per the Optyino.ai study (prtimes#1613462), only 4.4% (429) of 9,715 AI answers cite research-report content. Against the 86,047 cited URLs the share is just 0.6%, and general web pages make up 94.5% of all cited URLs; so among AI-answer citation sources, research reports are a minority. This is an observational tally of Optyino.ai's (Wallabee Inc.'s) own answer logs, not independently verified by this site (see Scope).

Q: In which question contexts and on which AI models are research reports cited most?

By question intent, "adoption-decision" is highest at 21.8% and "knowledge/problem-solving" lowest at 1.6%, a roughly 14x gap. By model, ChatGPT is highest at 10.8% and Grok lowest at 1.1% (about a 10x gap); "trends/future" is 20.2% (756) and "cases/use" 15.0% (492), while "comparison/recommendation" is just 1.7% (4,024) (prtimes#1613462). But the source notes model gaps reflect differing prompt counts and URL presentation, so they cannot be called a model-performance gap (see Scope).

Q: How far can this data be trusted?

This is an observational tally of Optyino.ai's (a GEO/AEO vendor's) own answer logs — first-hand but vendor methodology, not independently verified by this site (prtimes#1613462), and observational so it proves no causation. The source's own stated limits: the actual observation spans 4 calendar months including partial months (April only 4 prompts, July only 21); the data was collected via generative-AI APIs and may differ from real user experience and search results. So read these figures as "an observational snapshot of one vendor's logs," not an industry-census verdict (see Scope).

Sources

> Anti-fabrication statement: this card carries zero verified official anchors (official_count_verified=0); all figures are observational tallies of Optyino.ai's (Wallabee Inc.'s) own answer logs, not independently verified by this site; where the "%" sign differs (Copilot 4.6 / Claude 3.2) the source is recorded verbatim; no figure absent from the source was generated, no ratio or total was computed, and no cross-source figures were merged. No Wikidata Q is attached (no registry check performed).

📊 引用級事實單元（F-Units）

Research reports cited in 4.4% of answers (429/9,715), 0.6% by URL share; general web pages are 94.5% of all cited URLs
F-001 · Confidence: high · Basis: news_aggregation PRTIMES #1613462 Grade B · 平台審查企業發稿 observed 2026-04-16 to 2026-07-18 (answer basis / URL share)
Up to ~14x gap by question intent — 21.8% for adoption-decision, 1.6% for knowledge/problem-solving
F-002 · Confidence: high · Basis: news_aggregation PRTIMES #1613462 Grade U · 未分類／不可主張已驗證 observed 2026-04-16 to 2026-07-18 (by question intent, within-snapshot cross-section)
Up to ~10x gap across AI models — ChatGPT highest at 10.8%, Grok lowest at 1.1% (cited-answer rate is a separate metric)
F-003 · Confidence: high · Basis: news_aggregation PRTIMES #1613462 Grade U · 未分類／不可主張已驗證 observed 2026-04-16 to 2026-07-18 (by AI model)
85.7% of cited research reports come from universities, research consultancies, and public institutions; insurance/YMYL shows 0.0% across 2,053 answers
F-004 · Confidence: high · Basis: news_aggregation PRTIMES #1613462 Grade U · 未分類／不可主張已驗證 observed 2026-04-16 to 2026-07-18 (523 issuer URLs / by target area)
One of a same-publisher survey series — sister reports such as a white-paper citation rate of just 0.4% (25,001 answers)
F-005 · Confidence: medium · Basis: news_aggregation PRTIMES #1601325 Grade B · 平台審查企業發稿 each sister report's own observation window (samples/methods differ per report)
Market context — a wave of third-party GEO/AIO/LLMO services and surveys in the same period (e.g., JADE x DemandSphere "AI SEO Diagnosis Plus")
F-006 · Confidence: medium · Basis: news_aggregation PRTIMES #1579554 Grade B · 平台審查企業發稿 2026-07-17 to 2026-07-20 (each firm's release date)

❓ FAQ

Does generative AI actually cite research reports, and how low is the rate?

Per the Optyino.ai study (prtimes#1613462), only 4.4% (429) of 9,715 AI answers cite research-report content. Against the 86,047 cited URLs the share is just 0.6%, and general web pages make up 94.5% of all cited URLs; so among AI-answer citation sources, research reports are a minority. This is an observational tally of Optyino.ai's (Wallabee Inc.'s) own answer logs, not independently verified by this site (see Scope).

In which question contexts and on which AI models are research reports cited most?

By question intent, "adoption-decision" is highest at 21.8% and "knowledge/problem-solving" lowest at 1.6%, a roughly 14x gap. By model, ChatGPT is highest at 10.8% and Grok lowest at 1.1% (about a 10x gap); "trends/future" is 20.2% (756) and "cases/use" 15.0% (492), while "comparison/recommendation" is just 1.7% (4,024) (prtimes#1613462). But the source notes model gaps reflect differing prompt counts and URL presentation, so they cannot be called a model-performance gap (see Scope).

How far can this data be trusted?

This is an observational tally of Optyino.ai's (a GEO/AEO vendor's) own answer logs — first-hand but vendor methodology, not independently verified by this site (prtimes#1613462), and observational so it proves no causation. The source's own stated limits: the actual observation spans 4 calendar months including partial months (April only 4 prompts, July only 21); the data was collected via generative-AI APIs and may differ from real user experience and search results. So read these figures as "an observational snapshot of one vendor's logs," not an industry-census verdict (see Scope).

Verification Record

AI-structured, registry-verified, editorially sampled — Kirishima Rei (AI Editorial Desk)

⚠️ This article was AI-structured (not authored by an individual human reporter).

Cross-verified by multiple AI models.

Related Analyses