When ChatGPT or Perplexity answers a question and cites a few sources, it is not simply reciting what it memorized during training. Most modern AI answers are built with a technique called retrieval-augmented generation, or RAG: the system searches an external source for relevant material, then writes an answer grounded in what it found. Understanding this matters, because it tells you exactly where your brand can win or lose. You cannot edit what a model learned in training, but you can influence what it retrieves and what it chooses to cite. This is that mechanism, explained without the jargon.
What is RAG?
Retrieval-augmented generation is a way of answering a question that combines two steps: retrieval and generation.
An older-style language model answered purely from its training, everything it had absorbed up to a cutoff date, with no ability to look anything up. That made it fluent but unreliable and out of date. RAG fixes this by giving the model a search step. Before it writes, it retrieves relevant, current passages from an external source, the live web or a search index, and then generates its answer using that material. The retrieved passages are what it cites.
The short version: training is what the model knows; retrieval is what it looks up. RAG is the bridge, and it is why AI answers can be current and cite sources at all. If you want the one-line definition, it is in our AI search glossary.
The stages of a RAG answer
It helps to see the pipeline as a sequence, because each stage is a different opportunity, or a different way to be excluded.
- Query interpretation. The system takes the user's question and works out intent. Often it rewrites or splits the query into several sub-queries, a process sometimes called query fan-out, so one question triggers several searches.
- Retrieval. The system searches its source, a search index, the live web, or a curated database, and pulls back a set of candidate passages that look relevant. This is the gate. Only pages that get retrieved here can possibly appear in the answer.
- Scoring. The candidates are ranked for how well they answer the query and how trustworthy they look. Passages that state a clear, direct answer and come from sources the system trusts rise to the top.
- Generation. The model writes a synthesized answer using the top-scoring passages, usually naming only a few brands and weaving together a handful of sources.
- Citation. The system attaches links to the sources it drew on. Those citations are your credit for being part of the answer.
Miss stage two and nothing else matters. Win stages two and three and you are in the answer.
Where you can actually influence it
Here is the practical map, stage by stage, because this is what turns a technical explainer into a strategy.
You cannot touch training. What a model absorbed about the world is fixed until the next model version, and you have no direct lever on it. This is why "just wait for the model to learn about us" is not a plan.
You can heavily influence retrieval. Being retrieved requires being in the source the system searches: indexed by Google for AI Overviews, crawlable by the retrieval bots for ChatGPT and Perplexity, present on the wider web the model searches. If you are not indexed, or you have blocked the retrieval crawlers, you are invisible at the gate. This is the technical floor of AI visibility.
You can influence scoring and selection. Once retrieved, whether your passage is chosen depends on how clearly it answers the question and how trustworthy it appears. Clear, self-contained, directly-stated answers get selected; buried or hedged ones do not. That is the entire idea behind answer engine optimization. Trust comes partly from how the wider web describes you, which is generative engine optimization.
The original 2023 GEO research found the biggest lever at the selection stage was "information gain", content that adds something new rather than repeating what other sources already say. Passages that added quotations, statistics and citations were more likely to be picked. In other words, be worth retrieving and easy to quote.
Why RAG changes the visibility game
RAG is the reason ranking and being cited have come apart.
In classic search, the index and the ranking were the whole system. In a RAG system, retrieval is only the first gate; a separate scoring-and-generation step decides who actually makes the answer. That is why a page can rank well and still not be cited, and why only 38% of AI Overview citations now come from pages in Google's top ten. Being retrievable gets you considered. It does not get you chosen.
It also explains why training-only presence is not enough. A model might "know" your brand from training, but if you are not retrievable at inference, you will not be cited with a live source. This is the same reason the training-versus-retrieval crawler distinction matters so much: blocking training crawlers costs you nothing, but blocking retrieval crawlers removes you from the one step that produces citations.
Why different platforms cite different sources
Because each runs its own retrieval, over its own sources.
ChatGPT, Perplexity and Google's AI features do not share one index. ChatGPT's sourcing has shifted toward Google's results over the past year but still pulls from its own index and elsewhere; Perplexity leans heavily on community content, with Reddit its single most-cited source; Google's AI Overviews retrieve from Google's index. Same question, different retrieval pools, different citations. It is why being visible on one AI platform tells you little about the others, and why measuring across all of them matters. We go deeper on that in AI Search vs Traditional Search.
What to do with this
The RAG pipeline gives you a clean checklist. Be retrievable: indexed, crawlable, retrieval bots allowed. Be selectable: answer directly, in self-contained passages, with information worth quoting. Be trustworthy: consistent, accurate, and well-described across the web so the scoring step has reason to pick you. And measure across platforms, because each retrieves differently.
You are not trying to trick the model. You are trying to be the easiest correct source for it to retrieve, trust and quote.
Frequently asked questions
Want to see whether AI systems are retrieving and citing your brand? Get a free AI Visibility Audit.
Sources
- Ahrefs, "Update: 38% of AI Overview Citations Pull From The Top 10" (March 2026) — https://ahrefs.com/blog/ai-overview-citations-top-10/
- Profound, "AI Platform Citation Patterns" (Reddit is Perplexity's most-cited source) — https://www.tryprofound.com/blog/ai-platform-citation-patterns
- Profound, "AI Search Shift" (ChatGPT's shift toward Google's index) — https://www.tryprofound.com/blog/ai-search-shift
- Aggarwal, Pranjal, et al. (2024), "GEO: Generative Engine Optimization" (information gain; citations, quotations and statistics as selection levers), arXiv:2311.09735 — https://arxiv.org/pdf/2311.09735
- OpenAI, "Overview of OpenAI Crawlers" (retrieval vs training crawlers) — https://developers.openai.com/api/docs/bots