How ChatGPT, Perplexity, Gemini, Claude & Grok decide what to cite (2026 data)
Each AI engine searches the live web through a different backend and cites a different mix of sources: YouTube is the most-cited site in Google AI Overviews but only #8 in ChatGPT. Here is what the 2026 evidence says actually drives citations, per engine, and which popular GEO tactics the data refutes.
AI search changes fast — what these engines cite can shift within days to weeks, not months. This page reflects research current as of August 2026 (ChatGPT search providers: October 2026) and is reviewed quarterly. Every statistic below is dated to its source so you can judge how current it is.
Key takeaways
- Each engine searches differently: Gemini→Google, Claude→Brave, Perplexity & Grok→their own indexes, and ChatGPT blends its own index with several outside providers (not Bing alone).
- The strongest measured correlates are off-site: YouTube mentions (r≈0.737, #1) and branded web mentions (r≈0.66–0.71, #2).
- Gemini is the exception — it cites brand-owned domains far more than any other engine.
- Schema markup does not drive citations (it's hygiene); llms.txt has ~zero measured effect.
- Adding statistics, quotations and cited sources to content lifts citation likelihood 30–40%.
Every engine uses a different search backend
The single biggest mistake in GEO is treating "AI search" as one thing. Gemini runs on Google and Claude on Brave; Perplexity and Grok run their own indexes; ChatGPT blends its own index with several outside providers, Google results among them. That is why a source cited by one engine is usually invisible to the others — and why per-engine work, not a single trick, is the job.
| Engine | Search backend | What to prioritize |
|---|---|---|
| ChatGPT | A blend: OpenAI's own index plus outside providers (Google results, Microsoft, Yelp and TripAdvisor; Bing mainly in Deep Research) | Crawlable by OAI-SearchBot, indexed in Google and Bing, listicles, directories (G2/Capterra), Yelp for US local |
| Gemini | Google Search | Brand-owned content depth, entity clarity, Google ranking, YouTube |
| Perplexity | Own index (Bing-supplemented) | Reddit, content freshness, niche specialist sources |
| Claude | Brave Search | Research/data-dense content, statistics, awards |
| Grok | Own crawler + X | X/Twitter presence, engagement velocity, social proof |
Claude and Grok have a far thinner public evidence base than ChatGPT, Gemini and Perplexity — treat their guidance as lower-confidence.
What actually drives citations
The strongest measured signals are off your own site: across multiple vendor analyses (2025–26), a brand's own website is typically only ~5–15% of its AI citations — the rest is earned media and claimable third-party listings. In an observational vendor study, stories distributed to third-party outlets collected a median 239% more AI citations than the brand’s own copy alone (Stacker×Scrunch, Mar 2026; no control group). The exception is Gemini, which leans on brand-owned domains more than any other engine — the one place on-site depth pays off most.
YouTube mentions are the strongest measured off-site correlate of ChatGPT visibility.
Ahrefs Brand Radar correlation study · Dec 2025
Unlinked branded mentions across many third-party sites — the #2 correlate, just behind YouTube.
Ahrefs Brand Radar correlation study · Dec 2025
Adding statistics, quotations and cited sources to a page lifts its citation likelihood (GPT-3.5-era; directional).
Princeton GEO (KDD 2024) · 2024
Of ChatGPT's cited domains were in Bing's top 20 when the 2025 '87% Bing' figure was re-tested on 400 questions (13% were in Google's). ChatGPT blends several sources, and OpenAI says any public site can appear through its own crawler.
AI+Automation — re-test of the ChatGPT↔Bing overlap · Apr 2026
Adding schema markup moved Google AI Overviews citations by this much (significant negative). Schema is hygiene, not a citation lever.
Ahrefs schema controlled study · Aug 2025 – Mar 2026
Sites that block AI bots are still cited ~70% (retrieval-blocked) to ~95% (training-blocked) of the time — blocking has surprisingly little measured effect on citations.
BuzzStream / PPC.land citation study · Mar 2026
The GEO tactics the data refutes
A lot of GEO advice is folklore. These are the popular tactics that the current evidence does not support — we list them because being honest about what doesn't work is part of getting the rest right.
“Schema markup gets you cited.”
A controlled study of 1,885 pages (Ahrefs, Aug 2025 – Mar 2026) found schema had no positive effect on AI citations — slightly negative on Google AI Overviews. Schema still helps entity clarity and traditional search, so keep it as hygiene; just don't expect citations from it.
“An llms.txt file improves AI visibility.”
An analysis of 300,000 domains (SE Ranking, 2025) found no correlation with citations, and server-log analyses find AI crawlers almost never request the file. Google explicitly declines to use it. Harmless to add — just not a citation lever.
“GEO delivers a ~40% visibility boost.”
The widely-quoted 40% figure traces back to the Princeton GEO experiments (KDD 2024) — GPT-3.5-era tests that have not been replicated on today's engines. Treat blanket boost claims with skepticism.
“Domain authority, more pages, and press-release wires drive citations.”
All weak in the measured data: domain authority correlates ~0.18 and site page count ~0.19 with AI visibility (Ahrefs Brand Radar, N=75,000, 2025–26), and wire-distributed press releases account for ~0.04% of citations (BuzzStream, 4M citations, Mar 2026). Original, cited, fresh content beats volume.
Sources & dates
- [1] Ahrefs Brand Radar correlation study — N=75,000 brands (DR>40, ≥800 monthly volume) across ChatGPT, AI Mode and AI Overviews; Spearman correlation, uncontrolled. Extends Ahrefs' May 2025 AI Overview study of the same brands · Dec 2025 · link
- [2] Peec AI — where ChatGPT's search results come from — Maps every provider visible in the data ChatGPT's search returns, plus public confirmations: OpenAI's own index and at least eight outside providers, among them services returning Google results and Google Maps listings, Microsoft's grounding service, and licensed Yelp and TripAdvisor data; Bing appears mainly in Deep Research. Vendor analysis (Peec sells AI-visibility tracking) · Sep 2026 · link
- [3] AI+Automation — re-test of the ChatGPT↔Bing overlap — 400 questions through OpenAI's API with web search: 27.4% of cited domains were in Bing's top 20 for the question or the searches ChatGPT ran from it (3.0% for the question alone), 13.4% in Google's top 20. Replicates Seer Interactive's Feb 2025 study (~100 queries), which had found 87%. Vendor-run (an AI SEO agency), domain-level matching · Apr 2026 · link
- [4] Princeton GEO (KDD 2024) — N=10,000 queries; tested on GPT-3.5-era + Google search — directional · 2024
- [5] Ahrefs schema controlled study — 1,885 pages + JSON-LD vs 4,000 controls · Aug 2025 – Mar 2026 · link
- [6] BuzzStream / PPC.land citation study — 4M AI citations, 3,600 prompts, 10 industries · Mar 2026
- [7] SE Ranking llms.txt analysis — 300,000 domains; no measurable correlation · 2025 (reported)
- [8] McKinsey / Muck Rack multi-vendor citation source mix — A brand's own website is ~5–15% of its AI citations; the rest is earned media + claimable third-party listings (corroborated by 6+ vendors) · 2025–26 (multi-vendor)
- [9] Stacker×Scrunch distribution study (observational) — 87 stories, 30 clients, 2,600+ prompts, 8 AI platforms, one snapshot after distribution: counting citations to the syndicated copies as well as the brand's own page gave a median 239% more AI citations than the brand's page alone. No control group or before measurement; the authors call it observational. Stacker sells this distribution · Mar 2026 · link
- [10] Ahrefs: the 50 most-cited websites per AI engine — Brand Radar, US queries, September 2026 snapshot: YouTube is the #1 cited site in Google AI Overviews (22.9% of top-50 citations) but #8 in ChatGPT (2.5%); Reddit leads ChatGPT, Perplexity, Gemini and AI Mode. Shares are of the top 50 only; refreshed monthly · Sep 2026 · link
Correlational figures (e.g. Ahrefs r-values) describe association, not causation, and come from single-vendor datasets — treat them as directional. We refresh this page quarterly as the engines and the evidence base evolve.
Check your AI readiness — free
The free readiness check inspects your published robots rules, initial HTML and detected structured data — no signup, and no engine queries. Measuring whether ChatGPT, Perplexity and Gemini actually mention, cite and recommend you — and the source gaps behind it — is what the monitored S6S product does, currently in private beta.