RAG: definition, and why retrieval decides who gets cited

Most explanations of RAG stop at the acronym, which is exactly where they stop being useful. Something changed in 2024. ChatGPT, Claude, Gemini and Google's AI features all started searching the live web before answering. RAG stopped being a machine-learning detail and became the thing that decides whether your page is in the room when an answer gets written. This page gives you the plain definition and walks the five steps of the pipeline. Then it does the part almost nobody does, and shows you where in those five steps your visibility is won or lost. For the wider picture, start with our GEO agency pillar, or read the neighbouring GEO definition.

RAG: the one-sentence definition

RAG (Retrieval-Augmented Generation) is an architecture in which a language model retrieves relevant documents first, from a web index or a knowledge base, then writes its answer from what it retrieved rather than from memory.

Two verbs, in that order. Retrieve, then generate. Everything interesting about RAG lives in the comma between them, because that is where a decision gets made about which documents deserve to be read. A model without RAG answers from what it absorbed during training, the way you would answer a question from memory. A model with RAG goes and looks something up first, the way you would open a file before speaking. The second model is more useful, more current and easier to check. It is also, from where you sit, far more competitive, because a document has to be selected before it can be used.

Where the term comes from

RAG was named in a 2020 paper by Patrick Lewis and colleagues at Facebook AI Research, University College London and New York University. They paired a retriever with a generative model and showed the pair beat the model alone on knowledge-heavy tasks.

The paper is worth knowing about because it explains why the architecture exists. The authors observed that language models store a great deal of knowledge in their weights. But that knowledge is hard to update, hard to trace back to a source, and prone to being stated with more confidence than it deserves (Lewis et al., arXiv, 2020). Their fix was elegant. Stop asking the model to be the encyclopedia, and let it consult one instead. Six years on, that idea runs underneath the assistants you use every day. IBM Research describes RAG in exactly those terms in its public write-up, and Google documents how its own AI features retrieve from, and link back to, pages on the open web.

The five steps of a RAG pipeline

A RAG pipeline runs five steps: index the documents, chunk them into passages, embed each passage as a vector, retrieve the closest passages to the question, then generate the answer from those passages alone.

Read this as a funnel that gets narrower at every step, because that is what it is. Millions of documents at the start, a handful of passages at the end, and only those few passages reach the model.

Step 2 is the one people underestimate, so let me put it bluntly. The model does not read your beautiful 2000-word article. It reads a chunk of it, torn out of context, sitting next to chunks from four other sites. Every editorial decision you make should assume that fate. We went deeper into that in our piece on why chunking makes so much content useless to AI, and it remains the most common structural fault we see.

Why RAG became the default

RAG became the default because it fixes three problems retraining cannot fix cheaply. It grounds answers in real documents. It reaches information published after the training cutoff. And it lets a company plug its own knowledge into a general model.

Grounding is the headline benefit. A model that must cite the passage it used has less room to invent, and it can be checked. Freshness is the second. Training a frontier model takes months and costs a fortune, while a web index updates continuously, so retrieval is the only affordable route to yesterday's news. The third benefit is the commercial one that pulled RAG into every enterprise stack. A law firm or a manufacturer can point a general model at its own private documents and get grounded answers, without owning a training pipeline.

The engineering has kept moving too, and the numbers are instructive. Anthropic's engineering team reports that adding context to each chunk before embedding it, and pairing that with a keyword-search signal, cut the retrieval failure rate by 49 percent, rising to 67 percent once a reranking step is added (Anthropic, 2024). Read that from the outside and it says something simple: the industry is spending serious effort making retrieval pick better passages. Which means the passages that are easy to select keep getting selected.

What RAG changes for your visibility

RAG moves the gate. In classic search you compete for a rank in a list of links. Under RAG you compete to be one of a handful of passages retrieved before the answer is written. If your passage is not retrieved, the rest of your page is invisible, because the model never sees it.

This is the shift that catches teams off guard. The click data explains why it stings. Pew Research Center measured US Google users in March 2025 and found that when an AI summary appeared, only 8 percent clicked a traditional search result, against 15 percent on pages without a summary (Pew Research Center, 2025). The answer is now the destination. Being on page one of a list nobody scrolls is not the same as being in the passage the model quoted, and we unpacked the traffic side of that in our piece on how AI Overviews cut search traffic.

There is a flip side worth knowing, and it is good news. Retrieval does not respect the top ten. A passage can be pulled into an answer from a page that ranks well below the first screen, which is why we keep seeing AI citations coming from outside Google's top 10. Rank helps, but it is no longer the ticket. Clarity at the passage level is.

The stakes concentrate on precise questions, and those are more of the market than people assume. Of the 4836 French keywords Cicero Studio has analyzed, 34% draw fewer than 100 searches a month, a sign that the long tail dominates (Cicero Studio internal data). Across the 4253 French keywords Cicero Studio measured volume for, the median volume is 260 searches/month (Cicero Studio internal data). Sit with that median for a second. The typical query worth winning is not a head term with a five-figure volume. It is a narrow question asked a couple of hundred times a month, by people who already know what they want.

Those low-volume queries are not the throwaways. They are the specific, high-intent questions a buyer asks the moment before deciding. They are exactly the ones an assistant answers in two sentences with one or two citations. There is no page two to fall back on there. You are the citation, or you are absent.

How to make your content retrievable

Write so a single passage can stand alone. Open every section with a direct answer of one or two sentences. Keep one idea per heading. Back key claims with named sources and figures. And make sure AI crawlers can reach the page.

Start with the crawler, because it is the step that silently voids all the others. A page that AI crawlers cannot fetch, or whose content only appears after JavaScript runs, is not in the index and therefore cannot be retrieved by anyone, ever. Our guide on AI crawlers and invisible websites covers the traps we find most often, and the mechanics of how an engine grounds an answer in a retrieved page are laid out in our piece on how AI grounding rewrites the rules.

Then write for the chunk. A section that opens with a self-contained answer survives being torn out of its page; a section that opens with "as we saw above" does not. Add the credibility layer on top, because it is measurable: the 2024 GEO research from a team including Princeton and the Allen Institute for AI found that adding cited statistics and credible sources lifted visibility inside generative answers by roughly 40 percent, while keyword stuffing did nothing at all (Aggarwal et al., 2024). Structured data helps a machine reader parse who wrote what, a point our review of the schema markup and AI citations study examines with the data. And if you want the practical sequence rather than the theory, our method for appearing in ChatGPT and Google AI is the natural next read.

Here is the pattern I keep meeting. Across the 1207 SEO and GEO audits Cicero Studio has produced (Cicero Studio internal data), the recurring gap is almost never a shortage of content. It is content that a retrieval step cannot cleanly select: no direct answer at the top of a section, one section covering four topics at once, claims with nothing behind them. The business publishes constantly and stays unquotable. So the honest question is rarely "should we write more". It is "why can nothing we wrote be retrieved", and that is the question an AI-visibility audit is built to answer.

How I actually test this, by hand

The method is deliberately unglamorous, and I run it before touching a single page. I take the ten or so questions a real buyer would type, in their words rather than in keyword form, and I put each one to ChatGPT, Perplexity, Claude and Google's AI features in turn. Then I write down what actually comes back: which domains got cited, in what order, and whether the client appears anywhere at all. Most of the time they do not. The instructive part is what does appear, because the winning passage is usually not the best-written page on the topic. It is the one that answered the question in its first two sentences under a heading that named the question. That is a retrieval outcome, not a writing outcome, and it is repeatable.

The reverse test is just as telling. Take one of your own pages, cut it into 300-word slices, and read a slice at random with no memory of the rest. If that slice does not stand up on its own, the retriever will not rescue it. This is the cheapest diagnostic I know. It takes ten minutes and needs no tooling. The step-by-step version is in our GEO audit checklist, and the discipline it belongs to is set out in our LLMO definition and AEO definition.

The three shapes of an unretrievable page

Run that test across enough sites and the failures stop being varied. They fall into three shapes, and I now recognise all three within about a minute of opening a page.

• The buried answer. The page does answer the question, in paragraph six, after a warm-up about the company's history and values. The chunk containing the answer has no heading naming the question, so nothing signals what it answers. The chunk that does carry the heading contains the warm-up.
• The everything-section. One heading, often called something like "Our approach", sitting over 900 words that cover pricing, method, tooling and three case mentions. Chunked, it becomes four passages that are each about nothing in particular. A retriever matching a precise question finds no precise passage.
• The table with no sentence. Specifications, comparisons or prices laid out in a clean table, with no sentence anywhere stating the conclusion the table supports. A human reads the table fine. A retrieved passage of stripped table cells says almost nothing, so the model reaches for a competitor who wrote the sentence.

None of these are writing-quality problems. Each of those pages could be well researched, well designed and ranking respectably. They fail at step 4, and they fail silently, which is what makes the pattern worth naming.

The research explains why all three fail the same way. A chunk pulled out of a document loses the context that gave it meaning. Anthropic's engineering team had to inject that context back into every chunk before embedding it, and doing so cut the retrieval failure rate by 49 percent (Anthropic, 2024). The buried answer, the everything-section and the naked table are three ways of shipping chunks that cannot survive the separation. The fix for the third shape was measured directly by the GEO researchers. Write the sentence that states the conclusion, attach a figure and a named source, and you add the exact ingredient they found lifted visibility by roughly 40 percent (Aggarwal et al., 2024).

Alexis Dollé

CEO & Founder of Cicero Studio

I test AI visibility the slow way, by hand, one real business question at a time, across hundreds of sites. RAG is the part of the machine I keep coming back to, because it is where the decision is made and where almost nobody optimises. If a passage of yours cannot be lifted out and understood on its own, it does not matter how good the page around it is. That is the bar we hold every page to at Cicero Studio.

LinkedIn →

What RAG does not fix

A definition is only as honest as its edges, so here is what RAG is not. These are the four misreadings that quietly waste budget when nobody checks them.

Let me also be precise about what we can and cannot verify here, because the gap matters. We can observe what an assistant cites, we can repeat the test, and we can see whether a page is reachable and structured for extraction. We cannot see the retrieval stack itself: not the chunk sizes, not the embedding model, not the reranking weights. Anyone who tells you they know those numbers for ChatGPT or Google is guessing. Everything on this page is built either on published research and vendor documentation, cited below, or on what we measured ourselves and labelled as such.

This page also stays deliberately on the search side of RAG. It does not cover how to build an internal RAG system over your own documents, which is an engineering topic with its own trade-offs around vector databases, chunk sizes and reranking. What it covers is the side that decides whether the wider web, and the assistants reading it, ever reach your pages. If you would rather read all of this in French, the same definition is available on our page RAG définition.

Going further

We document our approach in the open, because that is our best proof. Across the 502 articles published on cicero.studio (264 in French, 238 in English), each resource below takes one angle of this definition further, from the retrieval mechanics to the audit method. Pick whichever matches your next question.