What is AEO, and what do
AI engines actually cite?

What we did, and what we found.

Two independent bodies of research sit behind this project. Part A is an original measurement study: we asked 3 AI engines the same 30 buyer questions, 3 times each, and recorded every source cited — 270 answers in total. Part B is an adversarially-verified review of the public literature: 28 sources, 124 candidate claims, narrowed to 22 that survived independent 3-vote fact-checking.

• The engines cite almost entirely different sources — ChatGPT, Gemini, and AI Overviews overlap on only 9–14% of cited domains. AI visibility is fragmented per engine.
• Cite rates diverge sharply — AI Overviews cites a source on 100% of answers, ChatGPT 81%, Gemini just 58%.
• Reddit is the #1 cited domain on every engine, but the engines otherwise split: ChatGPT leans editorial/review sites; AI Overviews leans community + video.
• The proven content levers (from the peer-reviewed GEO paper) are citing sources, adding quotations, and adding statistics. Keyword-stuffing measurably backfires.

The original citation study.

A first-party measurement of what three AI answer engines cite when asked real buying questions — designed to be reproducible and sliceable.

How the study was run

• Engines (3): ChatGPT, Google Gemini, and Google AI Overviews — the three most-used AI answer surfaces. (Perplexity and Claude were scoped out: lower usage and, for Claude, 8.2% share of AI-chat visits.)
• Prompts (30): real buyer questions spanning 6 verticals × 5 query types (see A.2). A deliberate blend of generic ("best protein powder") and brand-specific ("Purple vs Casper").
• Runs (3 per prompt/engine): AI answers vary run-to-run, so each prompt was asked 3× per engine and results aggregated — 90 samples per engine, 270 total.
• Measurement: for each answer we recorded every cited domain (the sources named in the answer, distinct from the broader retrieval pool). Location US, language EN, captured 16 Jun 2026.
• Taxonomy: each domain was bucketed (forum/UGC, encyclopedia, video, review/comparison, social, news/media, brand/other). Heuristic — raw domains are retained so the classification can be revised.

The 30 prompts

Full citation matrix — every prompt × engine

Each cell = the number of distinct domains that engine cited for that prompt (3 runs combined), with the single most-frequent one named. Gemini's blank cells and AI Overviews' spikes are both visible at a glance.

Per-engine profile

Cite rate — % of answers citing ≥1 source

Mentions vs links — being named ≠ being linked

A linked citation and a brand mention are different outcomes — an engine can recommend a brand by name without linking any source. A companion run (270 cells) captured the full answer text and extracted the brands each engine names, alongside the links:

Every engine names brands more than it links them, and Gemini's gap is the widest — it recommends a brand by name on 82% of answers but links a source on only ~53%, so ~1 in 3 Gemini answers name a brand with zero links. (Link rates here corroborate the headline study — AIO 100 / ChatGPT 81 / Gemini 58 — within run-to-run variance.) This refines "Gemini cites least": it gives the fewest links, but still recommends — visibility without attribution. The implication for AEO: on Gemini, being named (in-model knowledge / grounding) is a different and arguably more valuable target than being linked.

Cross-engine overlap — shared cited domains (Jaccard)

Top cited domains

Reddit ranks #1 on all three engines. Cross-engine totals: reddit.com (92) · youtube.com (46) · forbes.com (32) · google.com (29) · techradar.com (18).

Source type by query type

How-to questions spike on video (people want to be shown); best/alternatives pull the most news/media (editorial listicles); vs/use-case lean hardest on specific brand pages.

Source type by vertical

Concentration & the long tail

Citations are not concentrated in a few big aggregators. The top-10 domains account for only 27–35% of each engine's citations; the remaining two-thirds spread across 123–144 distinct domains per engine, with 64–80% going to specific vendor / niche "brand/other" pages.

What this study does and doesn't claim

• Snapshot in time — captured 16 Jun 2026, US/EN. AI citation behavior shifts quickly; treat absolute shares as a point-in-time reading.
• Commercial-query bias by design — these are buyer/transactional questions, so the picture differs from informational queries (e.g. Wikipedia, prominent in mixed-query studies, barely appears here).
• Heuristic taxonomy — "brand/other" lumps official sites, product pages, and long-tail blogs; raw domains are retained for re-classification.
• "Cited," not "retrieved" — we measured the sources named in the answer, not the wider pool an engine may have read.
• Self-references aren't filtered — on AI Overviews, google.com (Shopping/Flights) and youtube.com are Google's own surfaces, not independent publishers; they lift AI Overviews' domain counts and aren't actionable AEO targets (flagged in A.5).
• 30 prompts × 3 runs — robust for direction and ranking, not a census; per-cell counts are small.

Every domain and brand, ranked

The two tables aggregate the entire dataset — every linked source (citation) and every named brand (mention) across all 270 answers, three runs combined. The head is short and the tail is very long. The complete, unabridged lists ship as raw JSON alongside this report — citation-study-data.json and mention-study-data.json — so you can run your own analysis.

Most-cited domains — who AI engines link

339 distinct domains, 987 total citations — but long-tailed: 44% were cited only once, and the top-10 draw barely 28% of all citations. Reddit (92) is cited about twice as often as #2 (youtube.com, 46); AI Overviews drives the overwhelming share of the Reddit, YouTube and Google volume.

⚠ Google's own surfaces: google.com (29 — all but one from AI Overviews) and youtube.com (46 — 41 from AI Overviews) are Google-owned. The google.com hits are almost all product-shopping queries — Google's AI linking its own Shopping / Flights surfaces, i.e. citing itself. Counted as-is, but these aren't third-party AEO targets: you reach them via a Shopping/Merchant feed or a YouTube video, not by publishing citable content.

Most-named brands — who AI engines recommend

606 distinct brands, 1,613 total mentions, with 54% named only once — recommendations spread even wider than source citations. Strikingly, the most-named brands split evenly across all three engines (YNAB and Away are 6/6/6; Mailchimp 6/5/6): so while the engines cite wildly different sources (9–14% overlap, see A.3), they broadly agree on which brands to recommend.

The background literature research.

An adversarially-verified review of what's publicly known about AEO — so every external claim in the video and article is independently checked, not repeated on faith.

Multi-source sweep → 3-vote verification

The question was decomposed into 6 angles (definition/origin · the academic GEO paper · empirical citation data · citability levers · market growth · skeptical/contrarian). Sources were gathered per angle, key falsifiable claims extracted, then each ranked claim was put through independent 3-vote adversarial verification — voters instructed to refute; a claim needed a majority to survive.

What survived fact-checking

Claims that did not survive

The "why now" numbers

Reading the evidence honestly

• Time-sensitivity: every empirical citation statistic is a dated snapshot of a fast-moving field — cite with dates.
• Source tiers: the GEO-paper facts are peer-reviewed (gold); the citation-behavior data (Semrush, Ahrefs, Profound) is vendor-published (credible, commercial); the AEO-origin material is self-published first-party (use as "X claims").
• The GEO 40% must be hedged — best-case, relative, proxy-metric, benchmark-based.
• AEO ≈ GEO in practice; any "AEO vs GEO" distinction is editorial framing, not established fact.
• Open gap: schema / llms.txt impact appears overstated in early sources, but that specific claim is pending the verification pass before it ships.

Sources (28)

Tiered by quality. primary peer-reviewed / institutional · secondary reputable reporting · first-party self-published (use as attributed claim) · weak excluded from claims.

1. Jason Barnard — "The Trustpilot white-paper that started AEO" first-party
   jasonbarnard.com
2. Profound — "AEO vs GEO" blog
   tryprofound.com/blog/aeo-vs-geo
3. arXiv — "GEO: Generative Engine Optimization" (abstract) primary
   arxiv.org/abs/2311.09735
4. ACM Digital Library — KDD 2024, DOI 10.1145/3637528.3671900 primary
   dl.acm.org/doi/10.1145/3637528.3671900
5. arXiv — GEO paper PDF (v2) primary
   arxiv.org/pdf/2311.09735v2
6. Princeton institutional repository — GEO publication record primary
   collaborate.princeton.edu
7. arXiv — GEO paper HTML (v2) primary
   arxiv.org/html/2311.09735v2
8. Generative-Engines.com — GEO project page primary
   generative-engines.com/GEO
9. Semrush — "Most-cited domains in AI" (230K prompts) primary
   semrush.com/blog/most-cited-domains-ai
10. Semrush — "AI Mode comparison study" primary
    semrush.com/blog/ai-mode-comparison-study
11. Profound — "AI platform citation patterns" primary
    tryprofound.com/blog/ai-platform-citation-patterns
12. Peec.ai — "Top domains cited by AI search (30M sources)" primary
    peec.ai/blog/top-domains-cited-by-ai-search…
13. Ahrefs — "ChatGPT vs Google citations" (referenced in overlap finding) primary
    ahrefs.com/blog/chatgpt-google-citations
14. Pew Research — AI summaries reduce link clicks (Jul 2025) primary
    pewresearch.org
15. Gartner — search volume to drop 25% by 2026 primary
    gartner.com
16. Ahrefs — schema & AI citations primary
    ahrefs.com/blog/schema-ai-citations
17. Wikipedia — "Generative engine optimization" secondary
    en.wikipedia.org/wiki/Generative_engine_optimization
18. Search Engine Land — AI engines cite Reddit, YouTube, LinkedIn most secondary
    searchengineland.com
19. Search Engine Journal — AI Overview citations from top-ranking pages drop secondary
    searchenginejournal.com
20. Search Engine Land — "No, llms.txt is not the new meta keywords" secondary
    searchengineland.com
21. The Digital Bloom — 2025 AI-citation / LLM-visibility report secondary
    thedigitalbloom.com
22. Google AI Overviews reach — 2B (Pichai, Q2-2025 earnings) → 2.5B (Pichai, Google I/O, May 2026) secondary / primary
    pymnts.com (2B) · blog.google (2.5B) · cnbc.com (2.5B video)
23. Tech.eu — ChatGPT 800M+ weekly active users (Altman) secondary
    tech.eu
24. Stan Ventures — schema markup has no meaningful AI-citation impact secondary
    stanventures.com
25. arXiv 2406.18382 — follow-up generative-engine work primary
    arxiv.org/html/2406.18382v1
26. Contentful — llms.txt & search visibility secondary
    contentful.com/blog/llms-txt-search-visibility
27. SEOPress / DataVessel — keyword-stuffing has negligible/negative AI effect secondary
    (corroborating sources for the keyword-stuffing finding)
28. Adobe — AI-driven traffic surges weak — excluded
    business.adobe.com

How much work this was.

Most "AI SEO" content is a hot take. This project is two measured studies. Here's the full accounting.

Every external claim was cross-examined by independent verifiers; only the 22 that survived appear in the deliverables, and the 3 that failed are listed above in full. That's the standard for credibility — and it's why the article that comes out of this is itself built to be cited.