Listen instead

There’s no shortage of articles telling you to “write clear, structured content” if you want AI to cite you. True, and useless — it doesn’t tell you why, so you can’t apply it to a case nobody wrote a tip for.

This post covers the mechanism instead. It’s the same pipeline behind ChatGPT search, Perplexity, and Google’s AI features, and it’s well documented because it’s the same architecture companies use internally for their own document search.

What Is LLM SEO?

LLM SEO is optimizing your content to be retrieved and cited by large language model systems — ChatGPT, Perplexity, Google’s AI features, Claude. The name is newer than the practice, and it overlaps heavily with what people call AEO and GEO.

The distinction that actually matters isn’t the acronym. It’s the unit of competition. Traditional search ranks whole pages against a query. LLM systems retrieve fragments of pages. Your page isn’t competing as a page — each section of it is competing on its own.

That single difference drives everything below.

How Retrieval Actually Works, Step by Step

When an AI system answers a question using live web content, it runs a pipeline usually called RAG — retrieval-augmented generation. Four stages matter for you.

Step 1: Your page gets chopped into chunks

The system splits your page into segments, typically a few hundred to about a thousand tokens each — roughly a paragraph to a few paragraphs. This isn’t a stylistic choice; embedding models and language models have hard token limits, so long documents have to be broken up to be processed at all.

Consequence: the system will end up holding a piece of your page with none of the rest of it attached.

Step 2: Each chunk becomes a vector

Every chunk gets converted into an embedding — a long list of numbers representing what that chunk means, not which words it contains. Chunks about similar ideas land near each other in that numeric space even when they share no vocabulary.

Consequence: keyword matching isn’t what’s happening. Repeating a phrase doesn’t move you closer. Meaning does.

Step 3: The question becomes a vector too, and the closest chunks win

The user’s question is embedded the same way, and the system retrieves the chunks whose vectors sit closest to it. Many production systems combine this with old-fashioned keyword search to catch exact terms — names, part numbers, addresses — that pure semantic matching can miss.

Consequence: your chunk has to be recognizably about the question, on its own, with no help from the heading three sections up.

Step 4: Top candidates get re-ranked, then cited

Retrieval is fast and approximate, so a second, more expensive model — a re-ranker — usually re-scores the top candidates for how well they actually answer the question. The survivors get inserted into the model’s working context, and the model writes an answer citing the fragments it drew from.

You can see how much this stage matters in Anthropic’s published numbers for its own retrieval work: adding context to chunks before embedding them cut top-20 retrieval failures by 35%, and adding a re-ranking step took the total reduction to 67%. Retrieval quality is not a solved problem even inside the companies building these systems.

// the whole thing in one line

Your page is split into fragments, each fragment is scored on meaning against the question, the best few are re-scored, and the winners get quoted. You are competing chunk by chunk.

The Orphaned Chunk Problem

Here’s the practical payoff, and it’s the thing almost no SEO checklist mentions.

Because chunks get retrieved alone, a chunk that only makes sense in place is close to worthless. Consider a paragraph from a real services page:

“This service is typically completed within two weeks and includes everything mentioned above, plus a follow-up visit.”

Retrieved on its own, that fragment answers nothing. Which service? Included with what? For whom, where? A human reading the full page has the answer three headings up. The retrieval system doesn’t — it has this fragment and nothing else. It won’t be selected, and if it somehow is, it can’t be cited usefully.

The same paragraph, written to survive extraction:

“Tile shower installation from RK Tile & Stone is typically completed within two weeks. It includes surface prep, waterproofing, tile setting, grouting, and a follow-up visit after the first month.”

Same information. But this one is self-contained — it names its subject, states its facts, and answers a real question without needing the rest of the page. That’s the entire difference between content that gets cited and content that doesn’t.

This is also why Anthropic’s contextual-retrieval result is worth knowing: their fix was to prepend explanatory context to each chunk before embedding it, because context-free chunks retrieve badly. They solved it on their side of the pipeline. You can solve it on yours, by writing chunks that already carry their own context.

Why the Platforms Disagree

Ask ChatGPT and Perplexity the same question and you’ll often get different sources. That isn’t randomness — it’s the compounding of different choices at every stage above: which index gets searched, how content is chunked, which embedding model is used, how aggressively results are re-ranked, and how many sources get pulled into context.

Independent audits consistently find surprisingly low overlap between the domains different AI platforms cite for the same queries. The specific overlap figures vary a lot between studies — more on why in a moment — but the direction is consistent and unsurprising given the architecture.

The practical read: don’t optimize for one platform. The mechanism is shared even where the implementations differ, so writing self-contained, clearly-attributed, factually specific content is the move that pays off across all of them. Chasing one platform’s quirks is how you end up with advice that expires.

A Word on the Statistics You’ll See

Search for AI citation data and you’ll find confident numbers: this platform cites brands X% of the time, that content type gets Y times more citations. Treat most of them carefully.

A large share of published “AI citation studies” come from companies selling AI-visibility tools. Many don’t disclose sample size, query selection, or methodology, and their headline figures often disagree with each other by wide margins — which is itself a signal that they’re measuring different things and calling it the same thing.

That’s not a reason to ignore the space. It’s a reason to separate two kinds of claim: the mechanics of retrieval are well documented and stable — chunking, embeddings, re-ranking are how these systems work, published by the people who build them. The precise percentages circulating in marketing content are much softer. Build on the first. Be skeptical of the second, including when it appears in this post — which is why the only hard numbers here come from a primary engineering source.

We wrote more about this failure mode in SEO advice that stopped working, including a case where we had to correct our own.

How to Write for Retrieval

Five things, all of which follow directly from the mechanism rather than from a list of tips.

1. Make every section survive extraction. Read any single section of your page as though it’s the only thing someone can see. Does it still answer something? If it depends on “as mentioned above” or “this service,” rewrite it to name its subject.

2. Name entities instead of using pronouns. “We” and “our practice” and “it” disappear the moment a fragment is lifted out. “RedLight Freedom” and “red light therapy in Fort Lee, Virginia” survive the trip.

3. One idea per section. A section covering three topics produces chunks that are strongly about nothing. Chunks that are unambiguously about one thing score higher against questions about that thing.

4. Put the answer in the first sentence. Re-rankers score how directly a passage answers the question. A passage that opens with the answer scores better than one that opens with throat-clearing.

5. State facts specifically enough to be quotable. “We serve the surrounding area” can’t be cited. “We serve Richmond, Chesterfield, Midlothian, and Colonial Heights, Virginia” can. Specificity is what makes a fragment useful to a system assembling an answer.

// takeaway

Write every section as if it will be read alone by someone who has never seen the rest of your site — because that is exactly what happens.

Frequently Asked Questions

What is LLM SEO?
LLM SEO is optimizing your content to be retrieved and cited by large language model systems like ChatGPT, Perplexity, and Google’s AI features. It differs from traditional SEO in the unit of competition: search engines rank whole pages, while LLM systems retrieve individual fragments of pages. You’re competing chunk by chunk, not page by page.
How does ChatGPT decide which sources to cite?
When ChatGPT searches the web, it doesn’t read whole pages. Content is split into chunks, each chunk is converted into a vector representing its meaning, and the system finds the chunks whose vectors are closest to the meaning of the question. Top matches get re-ranked for relevance, inserted into the model’s context, and the model cites the fragments it actually drew from.
Why do ChatGPT and Perplexity cite different sources for the same question?
Because they are different systems making different choices at every stage — which index they search, how they chunk content, which embedding model they use, how they re-rank results, and how many sources they pull. Independent audits consistently find low overlap between the domains each platform cites, which is why optimizing for one doesn’t guarantee visibility in another.
What is chunking and why does it matter for SEO?
Chunking is splitting a page into smaller segments — typically a few hundred to about a thousand tokens — so they fit the token limits of embedding and language models. It matters because a retrieved chunk arrives without the rest of your page around it. A paragraph that says “as mentioned above” or “this service” becomes meaningless on its own, and a meaningless fragment doesn’t get cited.
How do I write content that AI systems will cite?
Write so that any single section still makes sense pulled out of context. Name the subject explicitly instead of using pronouns or back-references, put the complete answer in the first sentence under a question-shaped heading, keep one idea per section, and state entity facts — who, what, where — in plain text rather than assuming the reader has the rest of the page.
Are AI citation statistics reliable?
Treat most of them carefully. A large share of published “AI citation studies” come from companies selling AI-visibility tools, often without disclosed methodology or sample sizes, and their numbers frequently disagree with each other by wide margins. The underlying mechanics of retrieval are well documented and stable; the precise percentages circulating in marketing content are much less so.
// related glossary terms AEO GEO Schema

// FURTHER READING

Mechanism described here was verified against primary engineering sources on August 3, 2026. Retrieval architecture is more stable than platform-specific behavior, but if you’re reading this long after that date, re-check anything you plan to act on.

Want to know whether your own pages survive being read in fragments? Our SEO & AI-search service in Richmond, VA looks at exactly that — starting with an honest read of where you stand.