How Generative Engines Work: How LLMs Choose What to Cite

Every generative engine runs on two very different knowledge sources, and confusing them is where most businesses go wrong.

Training knowledge is what the model absorbed from its training corpus — hundreds of billions of words scraped from the public web up to a fixed cutoff date. Training does not store pages like a database; it adjusts billions of numeric weights so the model can predict likely text. If your brand appears consistently across the web — your own site, directories, press, reviews, reference sources — the model builds a stable statistical "impression" of who you are. This layer is slow to build and impossible to edit on demand.

Retrieval knowledge is what the engine fetches live at answer time. When a question needs current, local, or specific facts, the engine stops relying on memory, runs its own web searches, reads the returned pages, and writes an answer constrained to what it just read. This is retrieval augmented generation, and it updates in days rather than years. It is also where citations are actually won.

The strategic consequence is blunt: a business absent from training data can still get cited tomorrow through retrieval, and a business famous in training data can lose citations to a fresher, better-structured competitor. So the practical question — how do LLMs choose citations — is really a question about the retrieval layer, not the training layer. With ChatGPT alone reaching 800 million weekly users as of October 2025, that retrieval layer now shapes how millions of Americans research products, services, and local providers every week.

Retrieval augmented generation (RAG) is the architecture that lets a frozen model answer with live information. The steps are consistent across engines: the model interprets the question, writes one or more search queries, retrieves candidate documents, loads the most relevant passages into its working context, and generates an answer grounded in them. Every consumer AI search product — ChatGPT search, Perplexity, Gemini, Copilot — is a RAG system with a brand on top. Google's own version is covered in how AI Overviews work, and the largest standalone engine in how ChatGPT search works.

Why a business should care: RAG means the AI answer about your market is rebuilt from the live web every time someone asks. There is no permanent winner and no cached verdict. The recommendation your prospect hears today reflects whoever is most retrievable and quotable this month — a compounding advantage for businesses that maintain their content, and a fast leak for those that let it go stale.

Embeddings and vector retrieval: how the engine finds the right passage

The engine does not match your keywords letter-for-letter. It converts text — both the user's question and the candidate passages — into embeddings: long lists of numbers (vectors) that represent meaning. Passages about the same concept land near each other in that mathematical space even when they share no exact words, so "cheapest way to fix a leaking roof" can retrieve a passage titled "affordable roof repair options."

Retrieval, then, is a similarity search: the engine finds the passages whose vectors sit closest to the question's vector, often blended with traditional keyword and ranking signals. Three things follow directly for any page that wants to be found. Cover a topic's real sub-questions, because each one is a separate vector the engine might match. Write in plain, specific language, because vague marketing copy embeds into a fuzzy region that matches nothing precisely. And keep each idea in a self-contained passage, because the engine retrieves passages, not whole documents.

Grounding is the constraint that separates a search-grade answer from a confident guess. After retrieval, the engine is instructed to build its response from the retrieved passages and to attach citations to the specific claims those passages support. When grounding works, the sentence "the average roof replacement runs $9,000–$14,000" carries a link to the page that stated it. When a model answers without grounding — from training memory alone — that is exactly when it is most likely to invent a detail.

This is why citation is not a courtesy the engine extends to big brands. A citation is the engine showing its work: it marks the passage it leaned on to justify a claim. The page that supplied the cleanest supporting fact gets the credit, regardless of how large the brand behind it is. That single rule answers the underlying question — how do LLMs choose citations? — by following the supporting passage, not the reputation behind it.

Stage 4 — entity resolution — is the one classic SEO never had to solve. Before an engine can name your business, it has to be sure which business you are. If you share a name with three others, list inconsistent addresses across the web, or lack schema tying your pages to one identity, the model can retrieve you and still fail to attribute you confidently. Entity clarity is citation infrastructure.

Corroboration is the sibling signal. Engines trust a fact more when independent sources agree on it, which is why widely-referenced, community-vetted domains dominate AI answers: Semrush analyzed more than 100 million AI citations and found Reddit and Wikipedia the most-cited sources across ChatGPT, Google AI Mode, and Perplexity. The same logic governs whether an engine believes you exist and do what you claim: Ahrefs found that brand web mentions correlate with AI Overview visibility at 0.664 — far stronger than backlinks at 0.218. Being talked about across the web, consistently and by name, is now a leading input to being cited.

"LLMs don't cite authority — they cite utility. The page that hands the model a clean, sourced, current fact wins the citation over the household name that made it hunt. Corroboration is the tiebreaker: agree with the record and you get quoted; contradict it with no proof and you get dropped." — Thomas, Founder of AISEO USA, 16 years in digital marketing

A hallucination is a fluent, confident statement that is not true. It is a structural feature of how language models generate text — they predict plausible words, and plausible is not the same as verified. Grounding is the primary guardrail: by constraining the answer to retrieved passages and attaching citations, the engine gives itself a factual leash and gives the user a way to check the work.

Several guardrails stack on top. Corroboration filtering discounts claims that no retrieved source supports. Recency weighting favors current pages over stale ones. And increasingly, engines will refuse or hedge — "I don't have reliable information on that" — rather than invent, when retrieval comes back thin. The practical takeaway for a business is counterintuitive but clear: the way to survive the guardrails is to be the corroborating source. Publish specific, dated, attributable facts that agree with the wider record, and you become the safe passage the engine reaches for instead of the risky one it drops.

The recurring failure points, in rough order of frequency:

1. Crawler access. AI bots blocked by robots.txt or a bot-protection firewall — often without the business knowing. Blocked engines cannot retrieve you, so they answer your market's questions using competitors' pages instead.
2. No quotable substance. Pages full of adjectives and empty of facts embed into a fuzzy region and give grounding nothing to hold.
3. Entity ambiguity. Inconsistent names, missing schema, no third-party corroboration — the engine retrieves you but cannot confidently name you.
4. Staleness. Undated, un-maintained pages lose to fresher competitors in a system built to answer for now.

The fixes map to the two memories. For the retrieval layer (the fast lane): open access to AI crawlers, restructure money pages into direct answers built around quotable, sourced, dated statistics, and add the schema that makes entity resolution trivial. For the training layer (the slow lane): build consistent brand mentions across directories, press, and review platforms so every model's next training run knows exactly who you are. The peer-reviewed GEO study found that optimizing content for generative engines lifted source visibility by up to 40% — evidence that this is influenceable work, not luck.

This is the working definition of generative engine optimization, and it overlaps heavily with LLM SEO and broader AI SEO services. For the tactical checklist, how to get cited by AI walks through the moves in order. One standing rule: no one can guarantee AI citations — the systems are probabilistic by design, and anyone promising a placement is selling a fiction. The work raises probability, and it should be measured monthly.