wikis / Agentic SEO / wiki / summaries / ai-citation-studies.md view as markdown report a mistake
AI Citation Studies
Key Points
Almost every study here is published by a company that sells AI-visibility, SEO or content software. Each row states publisher, date, sample and a vendor-interest note. Numbers are never merged across rows: "AI Overview citations" (Google's product, per query), "AI assistant citations" (ChatGPT, Gemini, Copilot, Perplexity) and "ChatGPT retrieval" are different populations.
Table
| Study | Publisher (date) | Sample / method | Finding | Vendor interest |
|---|---|---|---|---|
| 76% of AIO citations from top 10 | Ahrefs (21 Jul 2025) | 1.9M citations from 1M AI Overviews; top 3 citations per response | 76.10% of cited pages ranked top 10; 9.50% ranked 11-100; 14.40% outside top 100. Median rank of the three citations 3 | Sells Brand Radar |
| Update: 38% | Ahrefs (2 Mar 2026) | 863K keyword SERPs, 4M AIO URLs, Brand Radar data, improved parsing (not like-for-like with the 2025 study), after AIOs moved to Gemini 3 (Jan 2026) | 37.9% of cited URLs in the first 10 SERP blocks (blue links only: 37.1%); 31.2% in 11-100; 31.0% beyond. 18.2% of non-ranking citations were YouTube (5.6% of all AIO URLs) | Same |
| 12% overlap, AI assistants | Ahrefs (11 Aug 2025; data collected early July 2025) | 15,000 long-tail prompts to ChatGPT, Gemini, Copilot, Perplexity, compared with Google and Bing top 10 | About 12% of cited links in Google top 10, but that is a mean of five figures that includes Perplexity's 28.6% (the other assistants are around 8%, so "12% for ChatGPT, Gemini and Copilot" overstates them); Bing top-10 overlap about 10%; about 80% did not rank anywhere for the original query | Same |
| AI Mode vs AI Overviews | Ahrefs (15 Dec 2025; Sept 2025 US data) | 540,000 query pairs (citations) and 730,000 pairs (content similarity) | Same URLs only 13.7% of the time (16.3% for top 3 citations); Jaccard similarity of response text 0.16 | Same |
| ChatGPT fan-out | Ahrefs (12 Oct 2026) | 66.6M logged-out free ChatGPT responses, 15M fan-out queries, Aug-Sep 2026 | 14.4% of responses contained fan-out queries (about 1.6 each); 6.8% of fan-outs used site:; 75.9% of common site: targets were brand sites; Reddit citations down 88% in six months (separate citation data) |
Same |
| Schema test | Ahrefs (11 May 2026) | 1,885 pages that added JSON-LD (Aug 2025-Mar 2026) vs 4,000 matched controls; 30 days before/after; difference-in-differences | No meaningful lift: AI Overviews -4.6%, AI Mode +2.4%, ChatGPT +2.2% (the last two indistinguishable from noise). Separately, across 6M URLs, cited pages were about 3x more likely to have JSON-LD | Same |
| AIO one-year review | BrightEdge (undated page; data to Feb 2026) | Generative Parser over tracked keywords; nine industries | AIO presence about 31% (Feb 2025) to about 48% (Feb 2026); average height about 1,200 px; about 17% of AIO-cited sources in organic top 10, flat over a Feb-Jul 2025 window (about 49% to 53% ranking anywhere in the top 100); by industry Healthcare 24.0%, Finance 11.3%, eCommerce 13.4% | Sells AI Catalyst |
| AIO citations vs organic | BrightEdge (18 Sep 2025; data May 2024-Sep 2025) | Nine industries | Overlap with organic rankings grew from 32.3% to 54.5%; Healthcare 75.3%, Education 72.6%, E-commerce 22.9% | Same |
| AIO CTR v3 | Seer Interactive (Apr 2026) | 53 brands, 5.47M queries, 2.43B organic impressions, Jan 2025-Feb 2026; OLS projection to Mar 2026 | Organic CTR on AIO queries fell to 1.3% (Dec 2025), rose to 2.4% (Feb 2026); no-AIO queries 3.8%. Cited brands got +120% more organic clicks per impression than uncited, still 38% below no-AIO. AIO shown on 36% of informational, 8% commercial, 5% transactional queries | Agency; sells search services |
| Fan-Out Effect | AirOps (undated landing page) | 16,851 queries, 353,799 pages, 10 industries, each query sent to ChatGPT 3 times | Page at retrieval position 0 cited 58% of the time vs 14% at position 10; strong heading match 41% vs 29%; JSON-LD 38.5% vs 32.0%; no positive link between domain authority or backlinks and citation | Sells an AEO content platform |
| Do engines cite the same sources? | Writesonic (22 Jul 2026) | 161,286 prompts, ChatGPT, Gemini, Perplexity, AIO; domain-level Jaccard | 3.8% of domains cited by all four (on the 70,879 prompts where all four returned citations); pairwise 0.119 to 0.237; Perplexity-AIO highest | Sells citation tracking |
| 5-engine overlap | Wellows (29 Jul 2026, updated 6 Aug) | 22.7M citations, 1.15M questions, Jan-Jun 2026, 531,889 questions cited by all five engines | 79.6% of sources on one engine only; 0.31% on all five; same page 6.8%, same brand 30.3%. Vendor's own random-baseline test weakened part of two claims | Sells tracking |
| Domain authority | Wellows (5 Aug 2026, updated 2 Oct) | 20.5M citations, 386,000 domains, 42 markets, 8 Mar-31 Jul 2026 | 47.9% of citations from domains below DA 40; 301 DA 90+ domains took 17.0% | Same |
| Citation economy | OtterlyAI (Jan-Feb 2026 data) | 1M+ citations, ChatGPT, Perplexity, AIO | Brand share of citations: AIO 59.8%, ChatGPT 44.7%, Perplexity 28.9%; 73% of sites with technical barriers to AI crawlers. Internal conflict on the 52.5% figure | Sells tracking |
| Most-cited domains | Semrush (10 Nov 2025) | 230K prompts, 13 weekly snapshots 14 Jul-12 Oct 2025, 100M+ citations, ChatGPT search, AI Mode, Perplexity | ChatGPT cited Reddit in about 60% of responses in early Aug, about 10% by mid-Sep; Wikipedia about 55% to under 20%; AI Mode and Perplexity more stable (Perplexity's Reddit share declined slightly) | Sells AI visibility toolkit |
| ChatGPT citation factors | SE Ranking (24 Nov 2025) | 129,000 domains, 216,524 pages, 20 niches | Referring domains, traffic, content freshness and Reddit/Quora presence correlate with citation; FAQ schema averaged 3.6 citations vs 4.2 without; llms.txt negligible | Sells a ChatGPT tracker |
Conflicts to keep visible
- How much AIO citation overlaps the top 10. Ahrefs: 76% (Jul 2025) then 38% (Mar 2026); Ahrefs says Gemini 3 and fan-out "may" explain it, and it also improved its parsing between the two. BrightEdge: about 17% top-10 overlap (Feb-Jul 2025 window, one-year report) and 54.5% "from organic rankings" (to Sept 2025) in another. The 54.5% is not a top-10 figure: that page says "Only 16.7% of citations come from top 10 results" and that growth comes from pages ranking 21-100; the one-year report shows about 53% ranking in the top 100. The 54.5% page does not state its cutoff. Wiki reading.
- Structured data. AirOps reports +6.5 percentage points for JSON-LD pages; Ahrefs' before/after test found no lift; SE Ranking found no gain for FAQ schema. Observational comparison (AirOps, Ahrefs' 3x) and a matched test answer different questions.
- Authority. SE Ranking reports high referring-domain sites 3.5x more likely cited by ChatGPT; AirOps found no positive link between authority and citation; Wellows found both small-site entry and large-site concentration.
- ChatGPT retrieval. Ahrefs originally said 14.4% of responses "searched the web", then corrected to "contained fan-out queries" in October 2026; the fetched copy's headings still say ChatGPT "answers 6 in 7 prompts without searching the web", so treat that framing as unreliable.
- Manipulation resilience. Semrush's Head of Organic and AI Visibility attributes the Reddit/Wikipedia drop to ChatGPT becoming "more resilient to manipulation"; this is opinion, not a tested result.
Relevant Concepts
- generative engine optimization
- ai visibility measurement
- ai search retrieval and fan out
- structured data and ai
- ai visibility tools
Source Metadata
| Item | Detail |
|---|---|
| Publishers | Ahrefs, BrightEdge, Seer Interactive, AirOps, Writesonic, Wellows, OtterlyAI, Semrush, SE Ranking |
| Date range | Studies published Jul 2025 to Oct 2026 |
| Independence | None independent of a software or services vendor. Raw data is not published in the gathered pages |
| Replication | No study was replicated by another in the gathered sources |
| Population note | Google AI Overviews, multi-assistant citations and ChatGPT-only studies are separate rows; do not average them |
| Gaps | AirOps report date; Google's own statements on citation selection are in ai search retrieval and fan out, not here; BrightEdge methodology definitions |
