Every AI answer you have seen this year runs on the same move. The summary at the top of Google, the reply from ChatGPT with little source chips under it, the paragraph Perplexity hands back. In each case the model does not answer from memory. It goes and fetches real pages first, reads them, and writes the answer from what it found.
That move has a name. Retrieval-augmented generation, or RAG.
I care about it for one blunt reason. If you publish anything you want found, RAG is the machine that now decides whether your page makes it into the answer or gets skipped. Ranking on the old page of ten blue links still matters, but it is no longer the whole story, because plenty of queries now get resolved by whatever the system retrieves and quotes before anyone scrolls down to a list. So it is worth knowing exactly what RAG does, and then what the data says it rewards.
What RAG actually is
Start with the problem it solves. A large language model, the kind behind ChatGPT, is trained once on a snapshot of text and then frozen. Ask it something and it predicts a plausible answer out of that frozen memory. This is called parametric memory, and it has two failure modes that matter for search. It goes stale, because the training snapshot has a cutoff date. And it invents things, confidently, when it hits a gap, because writing fluent text and telling the truth are not the same task.
RAG fixes both by refusing to let the model answer alone. Before it writes a word, the system runs a search, pulls back real documents, and hands them over as source material. The model answers from those documents instead of from memory. Google uses a plainer word for this. It calls it grounding: leaning on its core Search ranking systems to retrieve relevant, up-to-date pages, then building the response from what those pages actually say.1
That buys two things. The answer can be current, because the pages were pulled a second ago, and it can show its work, because the system knows which pages it read. Those little source chips under an AI answer are where that comes from.
How it actually works
Here’s the whole loop.
- 1Your question goes in PROMPT
Someone asks a real, messy question in plain language, the kind nobody types into a keyword box.
- 2The engine fans it out FAN-OUT
It quietly rewrites that one question into several related searches to cover more ground. Google calls this query fan-out.
- 3It retrieves and ranks chunks RETRIEVAL
It searches an index, breaks the matching pages into small chunks, and keeps only the passages closest in meaning to the question. It works with passages, not whole pages.
This is the step SEO now competes in. Get your passage into that shortlist, or you are not in the answer.
- 4The best chunks fill the context window CONTEXT
The winning passages load into the model's short-term memory alongside the original question. Everything else is left behind.
- 5It writes a grounded answer GENERATION
The model writes the reply from those passages, not from memory, and links the sources it leaned on.
Three of those steps break habits that SEO spent twenty years building. They’re worth slowing down on.
Query fan-out. The engine rarely takes your question at face value. It rewrites it into a handful of related searches and merges what comes back. Google confirmed this is how its AI features work, and gave the example of a lawn-weeds question spinning off separate searches for weed types, herbicide options, and prevention.1 So the page that gets used is often the one covering the surrounding sub-questions, not the one that matched the exact phrase typed.
It reads in chunks. The model never gets your whole article. The system breaks pages into small passages, turns each one into a string of numbers that stands in for its meaning, and keeps only the passages closest to the question. So a single strong paragraph can get pulled while the rest of the page sits unused, and a weak paragraph doesn’t drag the good ones down with it.
The context window. Only the winning chunks load into the model’s short-term memory, and there’s a hard limit on how much fits. Models also pay the most attention to what sits at the very start and end of that window, and the least to the middle. So being retrieved isn’t quite enough on its own. You want to sit near the front of what gets pulled in, where the model is actually reading closely.
Why this changes your job
For twenty years, SEO meant getting one page to rank on a list. That’s still a job, but it’s shrinking. The newer job is getting a single paragraph picked out of your page and quoted inside an answer you don’t control, which is a genuinely different thing to write for.
And the click is drying up while this happens. Pew Research looked at real browsing data and found that when an AI summary sits at the top of Google, only 8 percent of those searches end in a click to a site, against 15 percent when there’s no summary.2 So the answer keeps more people on the page than the old list ever did, and if your page isn’t in it, a lot of those queries never reach you at all. I traced how search got here in the history of AI search, and what it means for classic rankings in AI Overviews vs classic SEO.
What the data says gets retrieved
We can finally measure what these systems pick out, and the numbers are specific enough to change how you write.
Position matters more than almost anything else. Kevin Indig analyzed 1.2 million AI answers and just over 18,000 verified citations, and found a pattern his team named the ski ramp: 44.2 percent of all ChatGPT citations came from the first 30 percent of a page.3 Whatever your best line is, if it’s sitting halfway down the page, the retriever usually never gets to it.
The most-cited text was also thick with proper nouns, 20.6 percent of it, against the 5 to 8 percent you see in ordinary writing.3 A retriever has something to hold onto in a named brand, tool, person, or number, and nothing to hold onto in a phrase like “a leading platform.”
Questions help too. Cited passages were twice as likely to contain one, and most of those questions sat in headings.3 Ask the real question in an H2, answer it in the sentence right underneath, and you’ve matched how retrieval tends to read a page.
Freshness is the last one. Ahrefs pulled just under 17 million cited URLs and found the pages AI assistants cite run about 25.7 percent newer than the organic results below them, with ChatGPT skewing newest of all.4 An older page can still rank fine in the classic results while the answer sitting above it quotes something more recent.
What you can skip
I disagree with most of the RAG-SEO advice going around. A small industry has sprung up telling you to bolt on special files and markup so the robots can read you. Google’s own guidance says the opposite, and bluntly. To show up in its AI features, it says, you do not need machine-readable files, AI text files, markup, or Markdown. It names llms.txt directly. You don’t need structured data. You don’t need to chop your content into tiny pieces or write in some robot dialect.1
I wrote a whole guide on llms.txt and I still think the file is worth having. But be clear-eyed about what it does. Google says it doesn’t use it for this, and the retrieval studies above point somewhere completely different: front-loaded answers, named entities, real questions in headings, fresh pages, and content a crawler can actually load. A config file at your root isn’t going to move any of that.
What I would actually do
Most of that is just clear writing with the good stuff first, which you were meant to be doing anyway. A couple of things still get missed, though.
Write passages that stand on their own. The retriever lifts a single paragraph out and shows it with nothing around it, so anything that leans on “as we saw above” arrives broken. On the Consumer Cellular blog, we rebuilt a declining, AI-invisible site around getting retrieved and cited, and its AI Overview citations climbed from about 327 to 2,088 over nine months. It was mostly the same information as before, just shaped for how these systems read.
The other one is duller and easier to fix: make sure a crawler can reach the words at all. If your content only shows up after JavaScript runs, or the page is slow to respond, the fetch can time out and you’re dropped before you’re ever in the running.
Frequently asked questions
What does RAG stand for?
Retrieval-augmented generation. It is the method behind most AI search: instead of answering from memory, the system retrieves real web pages first and writes the answer from what they say, usually with citations.
Why does RAG matter for SEO?
Because AI answer engines use it to decide which pages to quote. Your content is broken into passages and judged one passage at a time, so the goal shifts from ranking a whole page to getting a single passage retrieved and cited.
Does RAG stop AI from hallucinating?
It reduces hallucination but does not remove it. Grounding the answer in retrieved sources means the model works from real text instead of guessing, though it can still misread or misattribute what it pulled.
Do I need an llms.txt file to get cited?
No. Google states plainly that its AI features do not require llms.txt, special markup, or structured data. The retrieval data points to writing quality instead: front-loaded answers, named entities, question headings, and fresh pages.
What is query fan-out?
It is when the engine takes one question and quietly rewrites it into several related searches, then merges the results. It means pages answering the surrounding sub-questions, not just the exact phrase typed, can end up in the answer.
How do I know if AI is citing my content?
Test it directly. Ask the assistants questions you should be the answer to and watch which sources they show. Tools that track AI citations exist, but manual spot-checks on your key topics are the honest baseline.
Keep going
More on how AI search works under the hood, and what it means for the pages you publish.
Sources
-
Google Search Central, Optimizing your website for generative AI features on Google Search. Source for the definition of grounding, the query fan-out example, and the guidance that llms.txt, special markup, Markdown, and structured data are not needed for AI features. ↩ ↩2 ↩3
-
Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results (July 22, 2025). ↩
-
Kevin Indig, “The science of how AI pays attention,” an analysis of 1.2 million AI answers and 18,012 verified citations, as reported by Search Engine Land, 44% of ChatGPT citations come from the first third of content. ↩ ↩2 ↩3
-
Ahrefs, Do AI Assistants Prefer to Cite “Fresher” Content?, analysis of 16.975 million cited URLs across ChatGPT, Perplexity, Gemini, Copilot, AI Overviews, and organic Google (2026). ↩
Working on this same shift?
I write about SEO, GEO, and getting found by AI search.
If this resonated, I'd love to compare notes.