If you want to know how to get cited by AI, you have to understand what happens before an AI writes a single word of its answer. Tools like ChatGPT and Google's AI Mode do not just pull from what they learned during training. They run searches, retrieve live pages, and stitch relevant passages into their responses. That retrieval process has a name: retrieval augmented generation, or RAG.
A recent explainer on the Ahrefs Blog by Louise Linehan breaks down exactly how RAG decides which pages get read and cited. The short version: getting into a model's base training data is largely outside your control, but getting into the retrieval layer is, in many ways, an extension of SEO. That is the part you can actually influence.
This guide explains how RAG works in plain terms, how AI search decides which sources make the cut, and what content signals raise your odds of being retrieved and cited. If you have been treating AI search as a black box, this is where it starts to make sense.
What retrieval augmented generation actually is
Retrieval augmented generation is a technique where a large language model queries an index (a search engine, knowledge base, or vector database) to find contextually relevant information before answering, instead of relying only on its training data. Models are trained on huge datasets, but that training has a cutoff date. Ask a model about last week's news and it is working from memory with no reference in front of it.
That is when models are most likely to hallucinate. RAG solves this by grounding the response in specific sources so the model is not freestyling from its training data. As the Ahrefs piece puts it, the LLM is either supplementing or overriding its internal knowledge (its parametric memory) to give a more reliable answer.
The name maps to three stages. Retrieval: the model runs a search to find relevant content. Augmented: it adds that content to its input. Generation: it uses the query plus the retrieved content to write the response. Most AI tools run RAG and trained knowledge together, with a base model generating language and a retrieval layer going looking for sources to attach.
Getting into the base model's knowledge means being part of its training data, and that isn't something you can easily control. But getting into the retrieval results is, in many ways, an extension of SEO.
Dorron Shapow, via the Ahrefs Blog
How AI search works, step by step
Every RAG-powered answer follows the same rough sequence: decide whether to search, run the search, then generate. Understanding each step tells you where your content can slip through the cracks and where it can win.
Step 1: The model decides whether to search at all
Before anything gets retrieved, the AI decides whether it even needs outside data. Simple queries like "what is a VPN?" can usually be answered from training knowledge alone, no retrieval required. In ChatGPT's case, research cited by Ahrefs (from David McSweeney) points to a smaller classifier model that assigns probability scores to decide whether a query needs no search, a simple search, or a complex multi-step search. Other tools handle this differently, but the logic is shared: not every query triggers a search.
Step 2: The model runs the search and fans out
When a question needs more context, the AI expands it into multiple related queries and sends them to an external index like Bing or Google. That expansion is known as query fan-out. Once pages are collected, on-page SEO factors like the title, meta description, and URL help determine which page gets read in full. Research by Dan Petrovic, referenced in the source, found that sources are shortlisted for scraping based on relevance, authority, recency, and diversity of perspective.
Step 3: Content gets chunked and the closest match wins
Before a page can be served in a response, the scraped content gets broken into smaller pieces called chunks. Think of it as tearing a book into individual chapters. The system then asks which chunk best answers the question. This is why a single tightly written passage can get you cited even when the rest of the page is only loosely related to the query.
Why some domains get a shortcut
Not every source is treated equally. According to research from Jérôme Salomon cited in the Ahrefs article, ChatGPT appears to build its own index of cached content, so it does not always retrieve from live search results. On top of that, separate research from Mark Williams-Cook, David McSweeney, and Suganthan Mohanadasan suggests ChatGPT feeds in content from a separate, licensed tier of authoritative sites and publishers, many with existing content deals, examples given include Reuters, the WSJ, and Wikipedia.
The takeaway for most brands is not that you need a licensing deal. It is that authority and trust still matter enormously in how AI search selects sources. If you are not a household-name publisher, your route into retrieval runs through the same fundamentals that have always driven organic visibility: crawlable pages, strong topical coverage, and earned authority.
How to get cited by AI: the content signals that matter
Because retrieval is an extension of SEO, most of the levers you already know still apply, with a few RAG-specific adjustments. Here is where to focus RAG content optimization.
- Answer the question in a single, self-contained passage. Chunking rewards content where one paragraph fully answers a specific query without requiring the surrounding context.
- Nail the title, meta description, and URL. These are the signals that decide which retrieved pages get read in full, so make them precise and query-aligned.
- Cover topics with genuine depth and multiple angles. Diversity of perspective is one of the shortlisting criteria, so thin, single-note pages get skipped.
- Keep content fresh. Recency is a stated retrieval factor, so update key pages and add publish and modified dates.
- Build authority through links and citations. Authority is a shortlisting criterion, and earned coverage still signals trust to both search engines and AI systems.
- Make sure pages are crawlable and fast. If your content cannot be reached or parsed cleanly, it cannot be chunked or retrieved.
The structural side of this work overlaps heavily with answer engine optimization, where the goal is to format content so machines can extract clean answers. Pairing that with a solid content strategy that maps real questions to dedicated passages is what moves you from occasionally retrieved to reliably cited.
How this changes the way you plan content
Query fan-out means a single user question can spawn several related searches. That rewards content that comprehensively covers a topic cluster rather than chasing one keyword at a time. Strong keyword research that surfaces the questions and sub-questions around a topic gives you a map of the passages you need to write.
It also raises the stakes on the technical foundation. If your pages are slow, blocked, or riddled with rendering issues, they never enter the retrieval pool. A thorough technical SEO baseline (crawlability, clean HTML, correct status codes, structured data) is what makes the rest of your AI visibility work possible.
Frequently asked questions
How does ChatGPT choose sources?
Based on the research cited in the Ahrefs article, ChatGPT first decides whether a query needs a search at all, then expands the query into related searches (query fan-out) and pulls results from an external index. On-page factors like title, meta description, and URL influence which pages get read in full, and sources are shortlisted for scraping based on relevance, authority, recency, and diversity of perspective.
What is RAG in SEO terms?
RAG SEO is the practice of optimizing content so it gets retrieved and cited by AI systems that use retrieval augmented generation. Since the retrieval layer behaves like a search engine, the same fundamentals apply: relevant, authoritative, well-structured pages that answer specific questions in self-contained passages.
Can I control whether I appear in an AI's training data?
Not really. Getting into a model's base training data is largely outside your control. What you can influence is the retrieval layer, which is where RAG content optimization comes in. Focus your effort there rather than trying to game the training set.
Do AI search citations replace traditional SEO?
No. AI search citations build on the same signals as organic search, they do not replace them. Crawlability, authority, freshness, and clear content structure serve both channels, which is why AI visibility and classic SEO are best treated as one integrated program.
RAG has quietly turned AI search into a retrieval game, and retrieval is something you can win with the right content and technical foundation. If you want a clear plan for how to get cited by AI across ChatGPT, AI Mode, and other engines, our team can help. Explore our AI / generative search optimization service or get in touch with LASEO for a proposal built around your topics and market.