When someone asks ChatGPT, Gemini, or Perplexity what product to buy or how to solve a problem, the engine builds its answer from a set of web sources — and increasingly, those citations decide which brands exist in the buyer's world at all. CommunityMentions tracks AI-search visibility for client brands as part of its core service. Every tracked brand gets a battery of buyer-intent queries run against the major answer engines on a recurring schedule, and we record every source each engine cites. This report aggregates that tracking corpus: 41,592 queries executed between January 9 and July 14, 2026, across OpenAI GPT models with web search (ChatGPT), Google Gemini with search grounding, and Perplexity Sonar. The question this data answers: when AI engines go looking for evidence, where do they actually go?

Queries were generated per tracked brand to mirror what real buyers ask: organic discovery questions (41.1% of the corpus), competitive and comparison queries (32.0%), and pain-point questions (24.2%), with the remainder covering solution discovery and alternatives research. Engine mix: 16,305 queries against OpenAI GPT with web search, 14,628 against Gemini with grounding, and 10,659 against Perplexity Sonar. For every response we stored the full answer text and the list of cited source URLs. Counting rules: a domain is counted at most once per answer (an answer citing three Reddit threads counts as one Reddit-citing answer), and percentages are computed against the 32,745 answers (78.7%) that cited at least one source. One artifact worth knowing about: Gemini's grounding API returns many citations as vertexaisearch.cloud.google.com redirect URLs that mask the destination domain. Those redirects (present in 8,368 answers) are excluded from domain-level rankings, which means Gemini's per-domain numbers — including its Reddit rate — are undercounts of its true behavior. Overall figures are therefore conservative. Because queries derive from real client categories (B2B SaaS, health and wellness, consumer products, ticketing, fitness, and services), the domain mix reflects commercial buyer intent rather than general knowledge questions.

Across all engines, Reddit was cited in 6,202 of the 32,745 sourced answers — 18.9%, the highest of any real domain in the corpus. For commercial, buyer-intent questions, the single most likely place an AI answer pulls evidence from is a Reddit thread. The top ten cited domains across all engines (share of sourced answers): 1. reddit.com — 18.9% 2. youtube.com — 14.4% 3. en.wikipedia.org — 5.8% 4. healthline.com — 3.9% 5. forbes.com — 3.4% 6. pmc.ncbi.nlm.nih.gov — 3.3% 7. facebook.com — 2.8% 8. g2.com — 2.8% 9. goodrx.com — 2.4% 10. techradar.com — 2.0% Note what's missing: brand-owned blogs. Individual vendor domains almost never crack the top of the citation stack. Engines overwhelmingly prefer third-party surfaces — communities, video, reference sites, review platforms, and publishers — over anything a brand publishes about itself.

Averaged together, sourced answers cite 8.14 sources each — but the engines behave nothing alike: - ChatGPT (GPT + web search): cited sources in 54.4% of answers; 5.2 sources per sourced answer; cited Reddit in 29.6% of sourced answers — nearly 1 in 3. - Perplexity Sonar: cited sources in effectively every answer (99.9%); 7.8 sources per answer; Reddit in 23.5%. - Gemini (with grounding): cited sources in 90.4% of answers; 10.4 sources per answer — the widest net; Reddit in 8.1% of measurable citations (an undercount — see methodology on redirect masking). The practical read: Perplexity always shows its work, Gemini casts the widest net per answer, and ChatGPT is the most selective — but when ChatGPT does cite, it reaches for community content at the highest rate of the three.

Counting all community and user-generated platforms together — Reddit, YouTube, Quora, Stack Exchange, LinkedIn, Facebook, Medium, Substack, review communities like G2, Trustpilot, and Capterra, and similar UGC surfaces — 32.7% of sourced answers cited at least one. For buyer-intent queries, roughly one AI answer in three is built partly from what real people said in a community, not from official marketing content. This is the mechanism behind the rise of community-focused search optimization: the conversations happening in threads are the raw material of AI answers about products.

Wikipedia ranks third among real domains overall (5.8% of sourced answers) — but that average hides a stark split. ChatGPT cited Wikipedia in 18.6% of its sourced answers, while Perplexity (1.1%) and Gemini (1.0%) barely touch it for commercial queries. For brands, the implication cuts two ways: encyclopedic presence matters disproportionately for ChatGPT visibility, while for Gemini and Perplexity the citation battle is fought almost entirely on communities, publishers, and review platforms.

Three tactical conclusions follow directly from the data: 1. Third-party presence beats owned content for AI visibility. The citation stack is dominated by places brands don't control — communities, video, reviews, publishers. Publishing on your own blog and hoping engines cite it runs against the grain of how engines actually source answers. 2. Reddit is the highest-leverage single surface. As the most-cited domain overall — and nearly a 1-in-3 citation rate on ChatGPT — authentic presence in relevant Reddit threads is the most direct route into AI answers for commercial queries. (Our Reddit Marketing Benchmarks report covers what that looks like operationally.) 3. Engine strategy is not one strategy. ChatGPT rewards encyclopedic and community presence; Perplexity and Gemini reward breadth across publishers and review platforms. Brands tracking their AI visibility should measure per engine, not in aggregate. This is the thesis CommunityMentions is built on: getting brands genuinely mentioned in the communities AI engines actually cite. The tracking system that produced this dataset is the same one we use to prove those mentions show up in AI answers.

This corpus reflects commercial buyer-intent queries for tracked client categories, not general search behavior; domain shares would differ for informational or general-knowledge questions. Gemini's redirect masking means its per-domain rates are floors, not exact values. Engine model versions advanced over the six-month window, and the engine mix was not identical across categories. Citation counts measure what engines showed users as sources, which can differ from every page an engine consulted internally.