What Is Retrieval-Augmented Generation, and How Does It Pick Sources?

Updated: 10 hours ago

Key takeaways
|
Ask ChatGPT what happened in your industry last week and you'll often see it pause and say it's searching the web. That pause, a couple of seconds at most, is the most important moment in AI visibility. In those seconds, the assistant decides what to look up, pulls a handful of pages and picks a few passages to build its answer around. If one of yours makes the cut, you're in the answer.
That process is called retrieval-augmented generation, or RAG. Engineers explain it with vector math and architecture diagrams, and you don't need either. You need a clear picture of where your content can win and where it quietly loses. An earlier version of this post also credited some fan-out research to the wrong study, and I'll flag that correction where it lands.
What is retrieval-augmented generation?
Retrieval-augmented generation, or RAG, is a technique where an AI model retrieves relevant information from an outside source, like a search index or a document collection, before it writes an answer. It answers from what it found rather than from memory alone. That's why AI answers can include current information and source links. Learn how the lookup works in search grounding. |
The term comes from a 2020 research paper by Patrick Lewis and colleagues at Facebook AI Research, University College London and New York University. The problem it set out to solve is still the problem today. A language model only knows what was in its training data. That knowledge has a cutoff date and gaps, and when the model doesn't know something, it tends to produce a confident answer anyway.
RAG fixes that by letting the model look things up before it speaks. It retrieves relevant text, then writes its answer from that text. The result is fewer made-up answers, more current information and a way to cite sources. Every time you see a link inside an AI answer, retrieval is why it's there.
For marketers, that's the whole opportunity. If an AI tool is going to look something up, your page can be what it finds. If it answers from memory, you're relying on what the model learned months ago. RAG is the part of AI search you can influence this quarter.
Where does content win or lose in RAG?
I explain RAG as four moments, each a gate your content has to pass. Find is when the system decides what to search for, and Fetch is when it pulls candidate pages. Filter narrows those to the few it will use, and Frame is when the model writes the answer. Learn how the Filter step works in vector embeddings.

Most content teams only think about Fetch, because it looks like traditional SEO. You get indexed, you rank and you get found. But the biggest drop-off happens at Filter. That's where most content quietly disappears.
Kevin Indig's analysis of about 1.2 million ChatGPT responses found it retrieved roughly 6 times more pages than it cited. Getting fetched is table stakes. Surviving the filter is the actual game. The rest of this post walks through what each gate rewards.
The four gates also explain why the same page can win one question and lose another. Each question runs its own Find, Fetch, Filter and Frame. A page that survives for one phrasing can fail for the next. That's why coverage of a topic matters more than any single keyword.
What happens at Find and Fetch?
At Find, the system decides what to search for, and for complex questions it rarely runs just one search. At Fetch, it pulls pages from a search index. Your page has to be indexed, crawlable and readable to make it through. Learn how AI tools split questions in query fan-out.
An earlier version of this post credited the fan-out numbers below to Kevin Indig. They come from AirOps' March 2026 study of 548,534 retrieved pages across 15,000 prompts, which Indig cited. AirOps found 32.9% of cited pages appeared only in fan-out results, not in results for the original prompt. And 95% of those fan-out queries had zero measurable search volume.
The way those searches look keeps changing, too. Peec's analysis of more than 20 million ChatGPT fan-out queries found their average length roughly doubled between October 2025 and January 2026, from about 6 words to about 12. Later model updates changed the pattern again. Ahrefs' October 2026 study found only 14.4% of the ChatGPT responses it analyzed contained fan-out queries, at about 1.6 each.
Fetch is the most technical gate and the least forgiving. If your page isn't indexed, it often can't be fetched at all, and if you've blocked AI crawlers, they can't read it. If your key content only appears after scripts run, some crawlers see an empty page. None of these problems are visible from the front end of your site, which is why they persist.
Why is Filter where most content dies?
At Filter, the system compares fetched passages with the question, usually by meaning, and keeps only the best few. Trust signals probably play a part, though no platform publishes exactly how. What survives is remarkably consistent. Learn how to structure passages that survive in content chunking for AI.
The passages that survive answer the question directly. They stand on their own without needing the paragraphs around them. They include specific names, numbers and claims rather than generalities. And they come from sources the web already treats as credible.
That's why two pieces of work that seem unrelated end up mattering so much. Chunking makes your passages retrievable. Brand mentions across the web make your source trustworthy. One gets you considered, and the other gets you chosen.
Position on the page matters too. Indig's February 2026 study found 44.2% of ChatGPT citations came from the first 30% of a page. Lead with the answer, then support it. A great answer buried at the bottom is fighting the odds.
What happens at Frame?
At Frame, the model writes. It might quote your sentence, paraphrase it, blend it with others or cite you without naming you in the text. You have the least control here, but you're not powerless. Learn how to write quotable sentences in what is an AI snippet.
Short, definite, self-contained sentences tend to survive into the final wording. Hedged sentences and long, winding ones get paraphrased or dropped. The model is looking for something it can lift and stand behind. Give it that.
An earlier version of this post applied a 92.4% figure to every AI platform. Practical Ecommerce's summary of Daniel Shashko's 42,971-citation study shows it comes from Google AI Mode and Gemini, where citation links point to an exact sentence. There, sentences of 6 to 20 words made up 92.4% of citations. Every citation was a full sentence.
A distinctive, named idea helps too. A framework with your name attached is far more likely to be credited to you than a generic point anyone could have made. Generic points get blended into the answer without credit. Named ones come with a source.
What are the two speeds of AI visibility?
Retrieval answers draw on content fetched at the moment of the question, so they can change within weeks of new pages being indexed. Answers drawn from training data only change when a new model ships, which can take many months. That's two speeds. Learn the slower lane in LLM SEO.

This explains something that confuses a lot of teams. Some changes show up in AI answers almost immediately, while others seem to take forever. The fast ones are retrieval wins, and the slow ones are training wins. Knowing which lane a change lives in sets the right expectation.
I tell clients to work both lanes at once. Retrieval is where improvements show up first and where you can see whether the work is landing. Training data is where long-term brand association lives, built through broad, consistent presence across the web. Neglect the first and you'll wait a year to see anything.
Neglect the second and your gains stay fragile. A retrieval win can vanish when a competitor publishes a better page. A brand the model already knows is harder to displace. The two lanes protect each other.
Where does RAG fit in the Engine Optimization Matrix™?
RAG is the machinery behind the GEO row of the Engine Optimization Matrix™. The Distribution lever governs Fetch, the Content lever governs Filter and Frame and the Citation lever is where you see the result. Each cell answers a specific failure. Learn the full picture in what is generative engine optimization.
Once you understand the four gates, every cell in that row stops looking like a best practice. Distribution work keeps you fetchable. Content work keeps you through the filter. Citation tracking tells you which gate you're failing.
So diagnose before you fix. If you're never fetched, the problem is indexing or access. If you're fetched but never cited, the problem is your passages. The anatomy of an AI response shows what that looks like in a real answer.
What else do people ask about RAG?
What does RAG stand for?
Retrieval-augmented generation. It's a technique where an AI model looks up outside information before writing an answer.
Who invented RAG?
The term comes from a 2020 paper by Patrick Lewis and colleagues at Facebook AI Research, University College London and New York University. It was presented at NeurIPS 2020.
Does ChatGPT use RAG?
Yes, when it searches the web before answering. That's how it cites sources and covers recent events.
Why doesn't my page get cited even when it ranks?
Ranking gets you fetched, but most fetched pages are filtered out. Indig found ChatGPT retrieves about 6 times more pages than it cites.
What is query fan-out?
It's when an AI tool splits one question into several smaller searches. AirOps found 95% of ChatGPT's fan-out queries had no measurable search volume.
How long does it take for changes to show up in AI answers?
Retrieval changes can show up within weeks of a page being indexed. Training data changes only arrive with a new model version.
What makes a passage more likely to be cited?
Direct answers, self-contained passages, specific claims and a credible source. Placing the answer near the top of the page helps too.
Is RAG the same as search grounding?
They're closely related. Grounding is when the model connects its answer to search results, which is one form of retrieval.
Related guides
About the author. Harold Bell is the founder of MQL Magnet and has spent about 16 years in enterprise technology marketing, working with brands like Cisco, Rubrik and Nutanix in different capacities. He's a member of the Forbes Communications Council, the creator of the Engine Optimization Matrix™ and the author of Partner Over Product. |




