What is RAG? Retrieval-augmented generation for marketers

What is RAG? Retrieval-augmented generation for marketers

geoNovember 19, 2026
By Antonio Fernandez

TL;DR

  • RAG (Retrieval-Augmented Generation) was introduced in a 2020 Facebook AI Research paper and splits answering into a retrieval step and a writing step, so a model can answer from text newer than its training data.
  • Google's documentation for AI features describes query fan-out and says normal indexing and snippet eligibility apply; OpenAI documents OAI-SearchBot separately from GPTBot; chunk sizes, selection weights and index freshness are not public for any vendor.
  • A citation comes after being reachable, indexed, retrieved and selected, so a blocked bot, JavaScript-only text or a paragraph that only makes sense in context can end the chain before citation.
  • RAG reduces invented answers but does not remove them: the wrong passage, an outdated source or a citation that does not contain the claim can still produce a confident wrong answer.

RAG stands for Retrieval-Augmented Generation. It is a process in which a language model (LLM) first retrieves relevant documents and then writes its answer from that retrieved text, instead of answering only from what it memorised in training. For marketers the practical meaning is that your content has to be retrievable before it can be used or cited in an AI answer. Much of how each vendor selects text is not public, and this guide says so plainly. It covers the search and GEO angle only and is not a tutorial on building your own RAG system.

What is RAG (retrieval-augmented generation)?

The name comes from a 2020 research paper by Lewis and colleagues at Facebook AI Research, titled "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks". The idea has two parts. Retrieval means finding relevant passages in a collection of documents. Generation means having a language model write the answer with those passages in its context.

The reason this exists is that a language model's knowledge stops at its training date and does not include any organisation's private documents. Asked about something it never saw, or something that has since changed, it may guess in a confident tone, which is called hallucination. RAG addresses that by placing accurate, current text in front of the model before it answers.

The term is used in two contexts that are worth keeping apart. In the first, developers build a question-answering system over their own company documents. In the second, an AI search tool searches the web and summarises an answer. This guide is about the second, because that is the one that affects brand visibility.

How does RAG work: retrieve, then generate?

Each system is designed differently, but the general structure looks like this.

  1. Prepare the collection (indexing): documents are split into pieces (chunks), and each piece is converted into a numeric vector (an embedding) that represents its meaning. The vectors are stored in an index. Traditional search systems also keep a keyword index.
  2. Receive and reshape the question: the user's question may be rewritten or split into several sub-questions before searching.
  3. Retrieve: the system finds the pieces closest to the question by comparing meaning, matching keywords, or a mix of both (hybrid).
  4. Rerank: some systems add a step that moves the pieces that answer the question most directly to the top.
  5. Assemble the prompt: the selected pieces are placed in the model's prompt with the question.
  6. Generate: the model writes an answer based on the supplied text, and some systems attach links to the sources.

What are chunk, embedding and grounding?

  • Chunk: a piece of text cut from a document so it can be searched and placed in a prompt. Size and cutting method vary by system, for example by heading or by character count. There is no worldwide standard value.
  • Embedding: turning text into a list of numbers so that texts with similar meaning have vectors that sit close together. This allows search by meaning even when the words do not match.
  • Grounding: tying an answer to a source that can be identified. A grounded answer is one you can trace back to the original text.

The point that matters to content teams is that the model does not read your whole website while answering. It reads only the pieces the system chose to retrieve. The useful questions are whether a piece from your page will be retrieved, and whether it makes sense once it has been pulled out.

How is RAG different from an LLM answering from memory, and from fine-tuning?

The three approaches give a model knowledge in different ways. The table summarises the mechanical differences.

How is RAG different from an LLM answering from memory, and from fine-tuning?
ApproachWhere the knowledge comes fromWhen the information changes
LLM answering from memoryModel weights learned in trainingNeeds retraining or a new release; information after the cutoff is missing
Fine-tuningWeights adjusted with an extra datasetNeeds another round of training; often better for style and behaviour than for facts that change often
RAGText retrieved into the prompt at answer timeUpdate the documents in the index and answers can change without retraining the model

The three can be combined. Many real systems use a model that has been trained and tuned, then add RAG for information that has to be current. For marketers the key difference is that RAG gives newly published content a chance to appear in answers without waiting for a model to be retrained, once the content has been found and retrieved.

How do AI search tools use RAG-style retrieval?

AI search tools such as AI Overviews, ChatGPT search and Perplexity search first and answer second, which resembles RAG. Their internal architectures are not publicly detailed, though, so what can be said with confidence is limited to what each vendor has put in public documentation.

  • Google AI Overviews and AI Mode: Google Search Central's documentation on AI features says these features may use a technique called query fan-out, which issues several related searches across subtopics to find supporting pages. The same documentation says there are no extra requirements: a page needs to be indexed and eligible to show a snippet, under normal SEO practice. Read the current version, because it is updated often.
  • ChatGPT search: OpenAI's publisher documentation says OAI-SearchBot is used to show sites in ChatGPT search results and is separate from GPTBot, which relates to collecting data for model training. Site owners can control each bot separately in robots.txt.
  • Perplexity: Perplexity publishes information about PerplexityBot, its crawler for surfacing and linking sites in its results. Check its current documentation for what the bot does.

What is not known from the outside

  • The chunk sizes and cutting methods each vendor actually uses.
  • The weight of the factors that decide which pieces are retrieved and which are cited.
  • Which index each tool searches, whether it is entirely its own or combined with other sources, and how often it refreshes.
  • How retrieval quality in Thai compares with English for each tool.
  • The criteria for choosing a cited source when several pages answer equally well.

So treat any advice that claims to know the exact formula for being cited in AI search with suspicion. The honest approach is to start from the mechanics everyone shares and measure results by testing on the systems you care about.

How does RAG relate to GEO, and why must content be retrievable before it is cited?

GEO (Generative Engine Optimization) is the work of making content likely to appear in, and be cited by, AI-generated answers. A RAG-style process shows that a citation comes after several other steps, and if an earlier step fails the later ones never happen. The sequence is: reachable, indexed, retrieved, selected for use, cited.

Where content usually drops out of the sequence

  • Not reachable: the bot is blocked in robots.txt, at the CDN or firewall, or the page sits behind a login.
  • Text not in the HTML received: content that appears only after JavaScript runs may not be seen by some bots. Check each system's documentation.
  • Not indexed: the page is noindex, canonicalised elsewhere, or so thin that Google does not keep it.
  • A piece that is hard to interpret: a paragraph that opens with "this" or "it" without saying what it refers to has no meaning once it is cut out on its own.
  • The answer is buried: the fact that answers the question sits in the middle of a long paragraph mixed with other points.
  • Stale or conflicting information: dates, figures or definitions that disagree between pages.

The conclusion is that basic SEO remains the foundation, because being reachable and indexed is a condition for almost every system. Making content findable is already part of SEO in Thailand. Shaping content so AI can read and cite it correctly is the subject of the GEO service.

How should you write content so systems can retrieve and use it?

The advice below is reasoned from the retrieval mechanics explained above. It is not a ranking factor published by any platform, and there is no public evidence that confirms any format raises citation rates by a specific amount. Treat it as hypotheses to test.

  1. Let each section answer one question: use h2 and h3 headings that match what users actually type, and open the first paragraph with a direct answer in one or two sentences.
  2. Make each paragraph stand alone: name the subject inside the sentence, for example "A TLS certificate must cover every hostname..." instead of "It must cover...", because a chunk may be separated from the paragraph before it.
  3. Define terms where they are used: give the meaning and full name the first time a term appears in a section, instead of relying on the reader going to another page.
  4. Use tables and lists for comparisons: clear structure makes each fact easy to lift out.
  5. State dates and numbers explicitly: include the year, version and unit in the sentence so a clipped passage still has full context, and update the page when the facts change.
  6. Keep the key text in HTML: do not hide answers in images or in content that loads only through JavaScript.
  7. Do not block the bots you want to reach you: check robots.txt and CDN settings so the search bots you care about can get in, and make the decision about training crawlers separately.
  8. Keep brand, product and service names consistent across pages: shifting names make it harder for a system to connect information.

One caution: do not write for machines so heavily that people cannot read it. Content a human cannot follow is not something anyone wants to cite, and stuffing repeated question-and-answer blocks in the hope of being retrieved can fall under Google's spam policies.

How much does RAG reduce hallucination?

RAG reduces invented information when the system retrieves accurate text that matches the question, because the model has real text to rely on. It does not remove the problem, and there is no neutral figure to quote for how much it reduces, because results depend on the system, the data and each question.

  • The wrong piece is retrieved: the system may fetch text that is close but does not answer the question, and the model writes from it.
  • The source is old or wrong: if the retrieved text is wrong, the answer is wrong, now with a credible-looking link attached.
  • The model reads more into the text than it says: a summary can add detail that is not in the cited source.
  • The citation does not match the claim: the attached link may be a related page that does not contain the sentence the answer relies on.

For a brand there are two effects. On the positive side, clear and verifiable content lowers the chance that an AI describes the brand wrongly. On the watch-out side, wrong answers about a brand can still happen. Check real answers on the systems your customers use at intervals, and when retrieved information turns out to be wrong, fix it at the source, which is your own pages.

RAG in the Thai market: what changes with the Thai language

Thai is written continuously without spaces between words, which makes tokenization and chunking harder than in languages that use spaces. A system that cuts chunks by character count may cut through a word or a sentence. Keyword matching depends on how well the system segments Thai words. Embedding-based retrieval works by meaning, but its quality in Thai depends on each vendor's embedding model, and there is no reliable public comparison, so you should not conclude that any system is better at Thai.

Practices that help in this setting include writing short sentences that finish their meaning on their own, using punctuation and spacing between sentences consistently, giving proper nouns, technical terms and brand names in both Thai and English script when people search both ways, and using the same name on every page.

Another point is that a Thai speaker who asks an AI in Thai may receive an answer that the system built from English pages. If your site has two languages, make both reachable, use correct hreflang links, and keep the key facts consistent so the two languages do not contradict each other. Arranging content so AI assistants answer accurately across both languages also falls under AI SEO.

How do you measure results when you do not know the system's formula?

When the internal mechanics are not published, the usable method is systematic external testing. Collect thirty to fifty real questions your customers ask (these numbers are illustrative, not a standard), ask them in the AI systems your customers use in both Thai and English, and record whether the brand is mentioned, which page is cited, whether the facts are right, and which competitor or source is cited instead. Repeat monthly, because AI answers change over time and can differ even when the same question is asked twice in a day, so ask several times and read the trend, not a single result.

When a page that should be cited is not, work through the order explained above. Check first that it is reachable and indexed. Then check whether the passage that answers the question reads well when lifted out, look for conflicting information, and see whether another source answers more clearly. Testing in rounds like this is the core of GEO done systematically, and tracking mentions across systems builds on the site's existing SEO foundation.

Summary

RAG is the pattern in which an AI retrieves text first and answers from it. For marketers, content must be reachable, indexed, retrieved as a chunk, and still make sense when separated from its page before it can be cited. The selection formula of each system is not known from outside, so plan from the mechanics that are certain and confirm with your own tests. If you want a GEO plan for a bilingual site, start with the GEO service from Relevant Audience.

RAG FAQ: common questions

What is RAG in one sentence?

RAG is a method in which a language model first retrieves relevant text and then writes its answer from that text. The full name is Retrieval-Augmented Generation, and it separates a retrieval step from a writing step so answers can rely on information newer than what the model memorised in training.

How is RAG different from fine-tuning?

RAG puts information into the prompt at answer time, while fine-tuning adjusts the model's weights through extra training. When information changes often, RAG is easier to update because you change the documents in the index, whereas fine-tuning suits changing the model's style or behaviour better. The two can be used together.

Do ChatGPT, Perplexity and AI Overviews use RAG?

All three search before answering, which resembles RAG, but their internal architectures are not publicly detailed. Google documents query fan-out for its AI features, and OpenAI and Perplexity publish information about their search bots. How they choose passages and cited sources is not known from the outside.

If I already do SEO, do I need to do anything extra for RAG?

Basic SEO is still necessary because a page must be reachable and indexed first. What you add is making each section answer one question directly, making paragraphs readable when lifted out, and checking that the bots you want are not blocked. Then test real questions in AI systems to see whether your content is cited.

Does RAG stop AI from making things up?

No. RAG reduces invented information when the right text is retrieved, but answers can still be wrong if the wrong passage is retrieved, the source is outdated, or the model reads more into the text than it says. Check that a cited page really contains the claim before you trust the answer.

Antonio Fernandez

Antonio Fernandez

Founder and CEO of Relevant Audience. With over 15 years of experience in digital marketing strategy, he leads teams across southeast Asia in delivering exceptional results for clients through performance-focused digital solutions.

Share to:
Copy link:

Read us often? Add Relevant Audience as a preferred source so our articles surface more in your Google results.