How Do LLMs Cite Sources? Retrieval, RAG, and What Gets Picked

LLMs cite sources when they answer from live retrieval: the engine searches the web or an index at question time, pulls a set of pages, writes its answer from them, and attributes the pages it leaned on. When a model answers purely from training data, there is nothing to attribute, which is why some [AI answers](/blog/patterns-that-keep-startups-out-of-ai-answers) carry citations and some arrive bare. Working that machinery deliberately, so the picked page is yours, is called generative engine optimization (GEO), and the same discipline also travels as answer engine optimization (AEO). One quick dispatch before we start: if you searched this phrase to learn how to reference ChatGPT in an essay, that is a style-guide question and the final FAQ above handles it in three sentences.

I have written elsewhere about [how a brand builds the mention base that citations draw on](/blog/how-ai-citations-get-built). This piece sits one level down, inside the engines themselves: the two layers an answer can come from, why one layer cites and the other cannot, and the selection rules that decide which pages get named.

The two layers every AI answer comes from

An answer from ChatGPT, Perplexity, or Google's AI features is built from some blend of two layers, and the citation behaviour of each layer is opposite.

The first layer is training data. Per IBM's explainer on retrieval augmented generation, models are trained on large but finite datasets of publicly accessible text, and each model carries a knowledge cutoff, the point after which it learned nothing new. Crucially, what the model keeps from training is a set of statistical patterns rather than a filing cabinet of documents. When it answers from this layer it is reconstructing, and there is no specific document it can honestly point at. That is why a purely from-memory answer ships without sources, and why a model pressed to produce sources anyway can invent plausible-looking ones. Grounding exists precisely because of this failure mode: IBM notes that retrieval lowers the risk of hallucinated output without eliminating it.

The second layer is live retrieval. Here the engine behaves less like a memory and more like a researcher: it runs searches at question time, fetches real pages, and generates its answer with those passages in front of it. The engine consulted genuine pages, so it can credit them. Every citation you have ever seen inside an AI answer came through this layer.

How retrieval turns into a citation

The architecture underneath is retrieval augmented generation, or RAG. IBM describes it as connecting a model to external knowledge so it can produce current, domain-specific answers without retraining, and it highlights the part that matters for this article: RAG systems can include citations to their retrieved sources, so a reader can check the claim against the material behind it. The citation is a verification feature of the architecture. It exists so the human can audit the machine.

Google has published detail on its own version of this loop. In its Search Central documentation on AI features, it explains that AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and data sources while the response is being generated, and that this surfaces a wider and more diverse set of supporting links than a classic results page. One buyer question becomes a small cluster of machine-issued searches, and every page that wins one of those hidden searches is a candidate for attribution.

I walked the full four-stage pipeline from crawl to generated answer, with a worked case, in the [how AI citations get built](/blog/how-ai-citations-get-built) piece, so I will spare you the embeddings recap here. The short version that matters for this article: retrieval operates on passages and on meaning, so a self-contained paragraph that answers a question cleanly can be lifted and credited on its own, whatever page it lives on.

Why some answers carry citations and some arrive bare

Three gates decide this, and none of them is random.

**Whether retrieval runs at all.** Stable general knowledge tends to be answered from the training layer, with no sources because none were consulted. Fresh, specific, local, and comparison-shaped questions are where retrieval earns its cost, and those are exactly the questions buyers ask. Google adds a product gate on top: AI Overviews are only shown when its systems judge the answer additive beyond classic results, which is why they often do not trigger.

**Which product you are in.** Perplexity is retrieval-first by design, so nearly every answer arrives cited. ChatGPT retrieves selectively, when the question or the user pushes it to search. Google's AI features are wired into a search engine, so supporting links are the default. Same underlying idea, three different citation temperaments, which is why measuring your visibility in one engine tells you little about the others.

**Whether the model can verify what it wants to say.** Even with retrieval running, the generation step decides which fetched pages get named. Claims that appear consistently across independent sources are safer to assert and easier to attribute. A claim that exists only on your own site gives the model one uncorroborated voice, and every business describes itself in glowing terms, so that voice is discounted.

What gets picked, and what the evidence says

Start with eligibility, because it is refreshingly boring. Google states that to appear as a supporting link in AI Overviews or AI Mode, a page needs to be indexed and eligible to show in Search with a snippet, and that there are no additional technical requirements. Its listed best practices open with making sure crawling is allowed in robots.txt, a check I would put first for any engine, since a frequent finding in our audits is AI crawlers blocked by a years-old blanket rule.

Beyond eligibility, the strongest public evidence on selection comes from the paper that named the field. The GEO study (arXiv 2311.09735, Aggarwal et al., accepted to KDD 2024) tested content changes against generative engines and measured visibility improvements of up to 40%, with the useful detail sitting in which tactics did the work: adding citations, quotations, and statistics helped, while traditional keyword stuffing performed poorly. The authors caveat that effectiveness varies by domain, and I would treat the 40% as the paper's ceiling rather than a promise. The transferable lesson is the ranking of tactics: engines reward material that looks like evidence.

The commercial stakes are the same ones we publish everywhere: organic search traffic is down roughly 30% across the board, while AI referrals convert at 3 to 6x traditional organic visitors. And the compounding is observable in our own case studies. [Ambiance Creations](/blog/ambiance-33x-breakdown), a kitchen and bath client, went from 939 clicks to over 31,000 in a year as the same citable content base fed both classic rankings and AI answers, one production effort serving both surfaces. That is also my standing answer to the GEO-versus-SEO framing: search expanded, and your surface area has to expand with it, because the engines citing you and the engine ranking you are reading the same pages.

IBM Technology has a clear seven-minute explainer on why language models hallucinate, which is the failure mode this whole citation apparatus exists to contain, and it pairs well with the two-layer model above.

<YouTube id='cfqtFvWOfg0' />

The honest caveats

These systems are probabilistic. Two people asking the same question can get differently sourced answers, retrieval triggers are opaque and shift with product updates, and nobody outside the engine companies can trace why one candidate passage beat another on a given day. So treat citation work as odds-shifting rather than placement-buying, and be suspicious of anyone selling guaranteed inclusion. The work also compounds on a timescale of months, and in full transparency I would rather tell you that here than have you auditing week-two dashboards. Finally, engine mix is audience-specific: where your buyers genuinely ask their questions should decide which citation temperaments you optimise for.

If you want to know whether your pages are even eligible to be picked, our free 46-point AI visibility checklist covers all seven audit categories, crawlability included. [Download the checklist](/checklist) and score your site this week.