Retrieval-augmented generationRAG

How does retrieval-augmented generation work?#

Retrieval-augmented generation works in two steps: a retriever searches an index for passages relevant to the question, then a large language model writes the answer with those passages placed in its prompt, citing the ones it used.

The technique was named in a 2020 paper by researchers at Facebook AI Research, University College London and New York University (Lewis et al.), which paired a text generator with “a dense vector index of Wikipedia, accessed with a pre-trained neural retriever.” The paper called the two halves parametric memory — what the model learned — and non-parametric memory, the index it can search. Every AI answer engine now runs a scaled-up version of that design, with the web as the index.

1. 1

   The question is rewritten into searches

   Engines rarely search the exact prompt. They write narrower sub-queries — query fan-out — and run them in parallel.

   Your lever: Title pages and sections the way a buyer’s follow-up question would be searched.

2. 2

   The retriever pulls candidates from an index

   Keyword matching and semantic search run over Google’s, Bing’s, Brave’s or the engine’s own index. Perplexity runs both at once and merges the results.

   Your lever: Be crawlable by the engine’s search bot and indexed where it looks.

3. 3

   Pages are split into passages and reranked

   Documents are broken into chunks, and a more precise model reranks them against the question. Only the top few survive.

   Your lever: Make each section answer one question in its first sentence.

4. 4

   The model writes from the survivors

   The surviving passages go into the model’s context, and it writes an answer grounded in them, attaching citations to the claims they support.

Google describes the same pattern for AI Overviews: RAG is used “to improve the quality, accuracy, and freshness of AI responses by relying on our core Search ranking systems to retrieve relevant, up-to-date web pages” (Google Search Central).

RAG vs fine-tuning: what's the difference?#

RAG gives a model new information at answer time by retrieving documents, while fine-tuning changes the model’s weights with extra training — so RAG handles fresh, citable facts, and fine-tuning shapes style and behavior.

The two aren’t rivals inside an AI product. A model is fine-tuned once to behave like a helpful assistant, and RAG then supplies the facts question by question. For marketers the distinction that matters is control: you’ll never fine-tune ChatGPT, and you can’t edit what it memorized. RAG is the part of the system that’s open to you — every engine that cites sources is retrieving them, and retrieval can be earned page by page.

Why does retrieval-augmented generation matter for SEO?#

Retrieval-augmented generation matters for SEO because the retrieval step runs on search indexes: a page that isn’t crawlable, indexed and matched to the engine’s sub-queries can’t be retrieved, and a page that isn’t retrieved can’t be cited.

That makes classic SEO the entry ticket, not the whole game. Retrieval works on sub-queries and passages, so a page can be cited for a follow-up question it answers well while ranking nowhere for the prompt the user typed.

Each engine retrieves from a different place — Google’s index for AI Overviews and Gemini, Bing plus OpenAI’s own index for ChatGPT, Brave for Claude, and Perplexity’s own 200-billion-URL index — so being findable means being indexed in all of them. The AI SEO guides cover each engine’s source and crawler.

How do you optimize content for retrieval-augmented generation?#

Optimize for retrieval-augmented generation by removing anything that blocks retrieval, then writing passages a retriever can match and a model can quote: descriptive headings, the answer in the first sentence, and specific names, numbers and dates.

1. Let the search bots in, and check your CDN, whose AI-bot toggles can override robots.txt. See AI crawlers.
2. Put the text in the HTML. Only Google’s pipeline renders JavaScript; OpenAI’s, Anthropic’s and Perplexity’s fetchers read raw HTML.
3. Match the sub-question in titles, H2s and URL slugs. Ahrefs found the titles of pages ChatGPT cited were semantically closer to the prompt than uncited titles (0.602 vs 0.484), and closer still to its fan-out queries.
4. Answer first, then expand. Growth Memo found 44.2% of ChatGPT citations came from the first 30% of a page. See answer-first content.
5. Make sections stand alone. A retrieved chunk arrives without the paragraphs around it, so name the subject instead of writing “it” or “as above.”
6. Stay current. Perplexity filters stale pages before ranking, and Ahrefs found assistants cite newer pages than classic search does. See content freshness.

Common mistakes with retrieval-augmented generation#

The most common RAG mistakes are assuming a top ranking for the head term is enough, hiding the answer behind JavaScript or a long intro, and writing sections that only make sense in context — each one stops a page at retrieval or extraction.

Related terms#

• How LLMs answerGroundingThe process of tying an AI model’s answer to specific, verifiable sources — usually web pages retrieved at the moment of the question — so the claims in the response can be checked and cited, rather than generated from the model’s memory alone.Read the entry
• How LLMs answerQuery fan-outAn AI search technique in which the engine rewrites one user question into several narrower sub-queries, runs them in parallel and builds its answer from the combined results — which is why a page can be cited for a prompt it doesn’t rank for.Read the entry
• How LLMs answerRerankingA second retrieval stage in which a more precise model re-scores the top results of a first, faster search against the query and reorders them — the step that decides which few passages an AI answer engine actually reads and cites.Read the entry
• How LLMs answerContent chunkingThe splitting of a page into smaller passages — often a heading and the text beneath it — that an AI retrieval system indexes, scores and quotes individually, which makes the section, not the whole page, the unit that competes for a citation.Read the entry
• How LLMs answerSemantic searchA retrieval method that matches a query to content by meaning and intent — typically by comparing vector embeddings — instead of by the exact words both contain, so a page can be found for phrasings it never uses.Read the entry
• LLMHow LLMs answerLarge language modelA neural network trained on vast amounts of text to predict the next word, which lets it write fluent answers — the technology behind ChatGPT, Claude and Gemini, whose knowledge is fixed at training time unless it retrieves fresh sources.Read the entry