Imagine you send a very fast, very literal research assistant into a library. You ask: "Who should I use to clear customs at Port Klang?"
The assistant does not read your question and recite from memory. They do four things. They break your question into smaller questions — who operates there, what does clearance cost, who is licensed, who do people rate. They run to the shelves and pull down whatever the catalogue points them at. They photocopy a few paragraphs — not whole books, paragraphs. Then they come back and write you one page in their own words, with little footnotes saying where each bit came from.
Your job, as a business, is not to be a good book. It is to be a paragraph that gets photocopied.
The mechanism, step by step
Nearly every AI answer that cites sources is produced by a pattern called retrieval-augmented generation, or RAG. Google confirms in its own developer documentation that its AI features work this way — they ground responses in indexed search results rather than in the model's memory. TIER A Here is the sequence.
A classifier inside the product decides if this question needs live information. "What is 12 times 8" does not. "Best freight forwarder in Klang" does.
Your one question is rewritten into a set of narrower ones. Google calls this query fan-out. Reported estimates put it in the range of eight to twelve sub-queries for a typical AI Mode question. TIER B You are not competing for one query any more. You are competing for a dozen you never saw.
Each sub-query runs against an index. What comes back is not a list of ten websites — it is a pile of passages, pulled from many pages, scored for how well they match that specific sub-question.
The passages are re-scored for quality, authority and agreement with each other. Perplexity, which has published more about its stack than most, runs candidates through several reranking layers and will discard the whole set and re-query rather than answer from weak sources. TIER B
The model writes one short answer constrained to the evidence it was handed, attaching citations to individual claims. In the Pew Research study of real Google usage, the median AI summary was 67 words and 88% of them cited three or more sources. TIER A
You are not trying to rank. You are trying to be one of roughly three paragraphs that survive a filter you cannot see, for a question that was never typed.
Why this reframes everything
Three consequences fall straight out of the mechanism, and they drive most of this book.
First: the unit of competition is the passage, not the page. A 400-word answer on a small site can beat a 5,000-word guide from a large one, if that specific paragraph is a sharper match for a specific sub-question. This is genuinely new, and it is the best news small businesses have had from search in fifteen years. Chapter 3 is entirely about this.
Second: you are being judged on agreement, not assertion. The engine pulls passages from many places at once. If your site says you are the leading customs broker in Selangor and nothing else on the internet says anything about you, you are a single unsupported claim. If four other sources describe what you do in roughly the same terms, you become a fact. This is why Chapter 5 — off-site mentions — is the biggest lever in the book, ahead of anything you can do to your own website.
Third: retrieval happens under time pressure. These systems fetch at enormous scale with tight timeouts. They do not wait for slow pages and they do not run your JavaScript. Chapter 6 covers this and it is the single most common way a good business is invisible for a purely technical reason.
Open ChatGPT, Gemini and Perplexity side by side. Ask each the exact question your best customer would ask before they knew your name — not your company name, the problem. Note three things: are you mentioned at all; who is mentioned; and which websites are cited underneath. That citation list is your real competitor set. It is usually not the list you expected.
Write down the ten questions a customer asks before they know you exist, and run all ten through two AI tools this week. Everything else in this book is a response to what you find.