LLM Citations: How Each AI Engine Actually Cites Sources (Data Study)

Every conversation about “getting cited by AI” skips the first question. What does an LLM citation even look like, and how many does each engine attach? The answer varies by an order of magnitude between engines. It also shifts dramatically depending on what you ask about.

LLM citations are the source links an AI answer engine attaches to its response. They are the receipts — the evidence trail a model shows for what it just told you.

For any brand, they are also the scoreboard. If your domain is in the LLM citations, the engine used you to build the answer. If it isn’t, you are invisible to that engine on that prompt.

cloro parses LLM citations for a living, so we ran the numbers. This study covers ~13,000 AI answers across six engines — ChatGPT, Google AI Overview, Google AI Mode, Microsoft Copilot, Gemini, and Perplexity — spanning six topic verticals (commerce, brand-intent, travel, voice-assistant queries, restaurants, and a cross-category product panel). For every answer we parsed the full citation list. Here is the anatomy of how large language models actually cite.

Per-engine citation anatomy

What counts as an LLM citation

An LLM citation is any source the engine attaches to a specific answer. It is not the same as a link buried in a training set. It is a live, per-answer reference the model surfaces to justify what it said. In practice that takes four shapes: a clickable pill, a numbered footnote, a source card, or a span of grounded metadata.

The distinction matters for measurement. A brand mention inside the answer text is not an LLM citation. A citation points at a URL you can attribute to a domain. That domain-level attribution is what makes LLM citations countable — and what makes absence from them measurable.

Two engines can cite the same fact and disclose it differently. One shows a clickable source card. Another hides the reference in grounded metadata a human never sees. Both count here as LLM citations, because both name a retrievable, domain-attributable source.

The six citation formats, engine by engine

Before the counts, the formats — because an LLM citation is not one thing. Each engine exposes its sources differently. The format decides whether the LLM citations are machine-readable at all:

Engine	Citation format	Placement	Machine-readable?
ChatGPT	Citation pills backed by a sources array	Inline + end-of-answer	Yes — structured array
Perplexity	Inline numbered footnotes [1][2]	Inline, tied to a source list	Yes — when present
Gemini	Grounded-source metadata	Attached to grounded spans	Yes — grounding metadata
Google AI Overview	Source cards	Right-side / inline cards	Yes — card links
Google AI Mode	Inline citation markers	Throughout the answer	Yes — dense inline links
Copilot	Numbered superscripts + source list	Inline + footer	Yes — source list

Each of cloro’s engine product pages (ChatGPT, Gemini, AI Mode, AI Overview, Perplexity, Copilot) documents the exact response fields. What matters for this study is that all six expose something parseable. That is what lets us count and categorize LLM citations at scale.

The format gap matters more than it looks. A pill or a numbered footnote is easy to read off. Grounded-source metadata sits deeper in the API response. So the same LLM citations a reader skims in a second can take real engineering to extract programmatically — which is why most rank trackers never attempt it.

Citation depth varies ~20× by engine

The single biggest finding: engines disagree wildly on how many sources an answer needs. Averaged across all six verticals:

Google AI Mode is the citation maximalist. It attaches roughly 15–22 sources per answer depending on the vertical — about 3× what ChatGPT, Gemini, or Copilot return. AI Mode’s design casts a wide net: many sources, each carrying proportionally less weight. If your domain is in the retrievable set at all, AI Mode is the engine most likely to surface it. But being one of 20 LLM citations is very different from being one of 4.

ChatGPT, Gemini, and Copilot are selective. All three cluster around 4–8 sources per answer. These engines make sharper choices about what to cite, which means each citation carries more weight — and being included is harder.

Perplexity is the citation minimalist — sometimes to zero. Its ~1.3 average hides enormous topic variance (below). Perplexity often produces a confident, named answer with no LLM citations at all. That is the hardest situation for anyone trying to monitor or influence AI presence: there is nothing to appear in.

The same engine cites differently by topic

Averages lie here, because LLM citations are strongly topic-dependent. The same engine that cites 20 sources for one query cites 4 for another. Here’s the full matrix — average citations per answer, engine × vertical:

Engine	Commerce	Brand	Travel	Voice	Restaurants	Product panel
AI Mode	14.9	7.7	19.6	20.4	21.6	21.6
AI Overview	8.0	6.3	11.5	14.3	9.7	13.6
ChatGPT	4.9	3.0	6.5	6.9	7.7	6.2
Copilot	7.7	4.9	4.9	7.1	8.0	6.9
Gemini	6.4	4.0	7.5	3.6	7.7	4.5
Perplexity	2.3	0.6	0.9	3.8	0.0	0.01

Two patterns jump out. First, brand-intent queries suppress citations across every engine — when someone asks “is [brand] legit,” models cite fewer sources (ChatGPT 3.0, AI Mode 7.7) than for any other topic, pulling from a narrower, complaint-weighted source set. Second, Perplexity’s zero. For dining (0.0) and cross-category product recommendations (0.01), Perplexity answers with essentially no citations, while for voice-assistant and finance queries (3.8) it does attach sources. Whatever gates Perplexity’s citation surface, it’s topic-driven, not a global setting.

AI Overview decides whether to answer at all

There’s a second axis the citation counts hide: AI Overview frequently declines to generate an answer. Its trigger rate — the share of queries where it produced an AI Overview at all — swings from near-total to almost nothing:

For shopping and voice-assistant queries, AI Overview answers almost always (98%). For local dining, it answers 3% of the time — Google defers to the local pack instead. So “does my domain get cited in AI Overview” has a prerequisite question: does AI Overview even trigger for my query class? For some verticals, the answer is mostly no. (Every non-Google engine in the study answered 100% of prompts regardless of topic; only AI Overview gates.) The AI Overview Trigger Index breaks this down by intent, and AI Overviews Around the World shows how both trigger rate and citation depth shift across 15 markets.

Who actually gets cited: the source anatomy

Count aside, which domains do engines cite? Aggregating every citation across the corpus, two sources form a universal backbone — and then each vertical has its own authority layer.

The universal backbone: Reddit and YouTube. Reddit is a top-2 cited domain in every single vertical we measured — 22% of restaurant answers, 19% of commerce, 18% of travel, 16–19% elsewhere. YouTube is #1 in commerce (19%) and voice (19%). User forums and video, the two source types classic SEO underweights most, are the connective tissue of AI citations. We’ve covered Reddit’s citation dominance in depth separately; this study confirms it holds across every topic, not just a few.

Why these two? Both are engines of first-hand, recent, high-volume text. Reddit threads read as unvarnished user experience. YouTube captions give models spoken, demo-style detail.

AI systems lean on that kind of content when they assemble LLM citations. It answers the did-someone-actually-use-it question a marketing page never can. The lesson for GEO is blunt: earn presence where real users talk, or watch the LLM citations route around you.

Each vertical has a distinct authority layer. Below the backbone, the cited domains are topic-specific and reveal what each engine treats as authoritative:

Vertical	Signature cited sources (beyond Reddit/YouTube)
Commerce	RTINGS (16%), PCMag, CNET, Tom’s Guide, TechRadar, Wired — the review-tech press
Travel	Expedia (14%), TripAdvisor (13%), Hotels.com, Condé Nast Traveler, Booking.com
Voice / finance	NerdWallet, Healthline, Mayo Clinic, Experian, Bankrate — YMYL finance + health
Restaurants	OpenTable (13%), TimeOut, The Infatuation, Michelin, Eater city editions
Brand intent	Trustpilot, BBB, ConsumerAffairs — review + complaint boards
Local (all)	google.com / Maps — heavy in travel (16%) and dining (19%) as a grounding source

The practical read: there is no single “get cited by AI” playbook. Landing in commerce answers means being reviewed by RTINGS and covered on tech YouTube. Landing in travel answers means presence on Expedia and TripAdvisor. Landing in dining answers means the local Eater and TimeOut “best-of” lists. The engines are drawing from different authority sets per topic, and your GEO strategy has to match the vertical you compete in.

Each engine cites a different web

The vertical view above holds across engines. Cut the same corpus the other way — by engine rather than by topic — and a second pattern appears: the engines don’t just cite different amounts, they reach for different domains. Aggregating the most-cited source domains per engine over a complementary parse of the live monitoring corpus (roughly 2,500 answers per engine over a 7-day window), the top lists barely rhyme:

Engine	Most-cited source domains
ChatGPT	reddit.com, ikea.com, techradar.com, whowhatwear.com, zara.com, apple.com, nike.com
Google AI Mode	google.com, youtube.com, instagram.com, reddit.com, facebook.com, linkedin.com, ikea.com
Perplexity	youtube.com, ikea.com, zara.com, target.com, shopify.com, hm.com, nike.com
Google AI Overview	youtube.com, reddit.com, instagram.com, facebook.com, google.com, ikea.com, quora.com
Copilot	ikea.com, nike.com, walmart.com, amazon.com, apple.com, zara.com, alibaba.com
Gemini	reddit.com, ikea.com, youtube.com, apple.com, forbes.com, pcmag.com

• ChatGPT is a Reddit-and-brands engine — community opinion (reddit.com is its #1 source) stacked with retail and category-media pages.
• Perplexity is video-first — youtube.com leads by a wide margin before it reaches for retail.
• Copilot is retail-only — the outlier: its top domains are almost purely commerce, with no Reddit and no social platforms in the list at all. Where every other engine surfaces community content, Copilot goes straight to the store.
• Gemini leans Reddit plus tech media — reddit.com, then an Apple/Forbes/PCMag editorial tilt.

And Google’s AI engines cite Google’s own properties. Google AI Mode’s two most-cited domains are google.com and youtube.com (YouTube being Google-owned), and AI Overview cites youtube.com most with google.com prominent — while google.com appears in the top domains of no non-Google engine (ChatGPT, Perplexity, Copilot, Gemini). We can only report the pattern, not the intent, but the effect is real and it’s a per-engine GEO consideration: on Google’s surfaces, Google’s own properties are part of the citation mix in a way they aren’t anywhere else. The one universal thread is Reddit and YouTube near the top of most lists; past that shared backbone, each engine adds its own layer — so a domain that’s a top source in one engine can be absent from another’s entirely.

Why citations diverge from rankings

The reason all this matters for SEO teams: LLM citations are not a re-skin of the Google top 10. The most-cited AI sources — Reddit, YouTube, RTINGS, NerdWallet, OpenTable — are frequently not the pages ranking #1 for the same query in classic organic search. AI engines retrieve and synthesize from a source set weighted toward forums, video, review aggregators, and category authorities. They then attach LLM citations drawn from that set.

That divergence is the whole case for generative engine optimization as a discipline distinct from SEO. A page can rank first on Google and never enter an AI answer; a Reddit thread that ranks nowhere can be cited by four engines at once.

Optimizing for LLM citations means optimizing for a different retrieval system with different favorite sources. That is why measuring AI share of voice requires watching the citation surface directly, not inferring it from rankings. (And none of it works if AI crawlers can’t reach your pages in the first place — the supply side of citations.)

Monitoring your citations

The operational takeaway falls out of the data. LLM citations differ by engine and by topic — in depth, format, trigger behavior, and favored sources. So you can’t infer your AI presence from any single engine, and a general rank tracker won’t capture it either. You have to:

1. Run your priority prompts across all six engines — the citation-heavy ones (AI Mode) and the citation-shy ones (Perplexity) tell you different things.
2. Parse each answer’s full citation list — pills, footnotes, source cards, and grounding metadata normalized into one structure.
3. Track whether your domain appears, per engine and per prompt, over time — and which third-party sources are cited in your place.
4. Match the vertical’s authority layer — being absent from RTINGS (commerce) or Eater (dining) shows up as citation absence long before it shows up in traffic.

That normalization across engines is exactly what cloro’s AI visibility tracking does — the same pipeline that produced this study, pointed at your domain. For the tooling landscape around it, the LLM visibility tracking tools comparison covers the options.

How to track your LLM citations over time

Tracking LLM citations is a monitoring problem, not a one-time audit. Engines rewrite answers constantly. A domain cited in ChatGPT this morning can drop out by tomorrow. The only reliable signal is a time series, never a single snapshot.

The workflow itself is simple: pick the prompts your buyers actually type. Run them across every engine on a fixed schedule. Capture the full LLM citations list from each answer. Then diff the results over time — appearances, position in the list, and which rival domains showed up in your place.

The parsing is where most tools stop. Normalizing LLM citations across six formats — pills, footnotes, source cards, grounded metadata — is the hard engineering. That is the exact step cloro’s pipeline is built to handle, at the scale this study required.

One more practical note: measure position, not just presence. Being the 18th of 20 LLM citations in AI Mode is not the same win as being one of ChatGPT’s four. A good tracker records where in the list you land, per engine and per prompt. That is what turns a raw LLM citations feed into a GEO roadmap instead of a vanity metric.

Methodology

The study parses citations from ~13,000 successful AI answers captured in cloro’s monitoring corpus between 2026-07-03 and 2026-07-04, across six engines (ChatGPT, Google AI Overview, Google AI Mode, Microsoft Copilot, Gemini, Perplexity) and six topic verticals (commerce, brand-intent, travel, voice-assistant queries, restaurants, and a cross-category product panel), US, English-language. “Average citations per answer” counts parsed LLM citations per triggered response, per engine, within each vertical; the headline per-engine averages are means across the six verticals. Trigger rate is the share of prompts for which an engine produced an answer (only AI Overview varied materially; every other engine answered ~100%). Cited-domain shares are the share of triggered answers in a vertical in which a domain appeared at least once. The per-engine “different web” cut is a complementary parse of the same monitoring corpus (roughly 2,500 answers per engine over a 7-day window), ranking each engine’s most-cited source domains regardless of vertical; it reports domain rank order, not the per-answer citation counts above, so the two views don’t share a scale.

Caveats, stated plainly. This is one measurement from one corpus with a US, English prompt mix skewed toward commercial, brand, and recommendation intents; treat the point estimates as cloro-corpus signals, not universal constants. Citation count is not citation quality — a 20-source AI Mode answer isn’t necessarily better-grounded than a 4-source ChatGPT one, and being one of 20 LLM citations confers less influence than being one of 4.

The findings are robust across verticals and any reasonable prompt weighting: AI Mode’s citation-maximalism, Perplexity’s topic-gated citation surface (down to zero for dining and cross-category product queries), AI Overview’s variable trigger rate, and the Reddit/YouTube backbone beneath vertical-specific authority layers. Citation formats are current as of mid-2026 and change as engines ship UI updates.

If you want this measured continuously for your domain — which engines cite you, how deep in the LLM citations list, for which prompts, and which sources are cited in your place — that’s what cloro’s AI visibility tracking is built for.