A shop owner asks ChatGPT which accountant understands online sellers in Manchester, and it names a few firms with links. She asks Google the same thing and an AI Overview sits above the results, citing a handful of pages. Neither answer came out of the model's memory. Both were assembled from searches that ran a moment before the answer appeared, and the pages those searches found are the ones that got named. Understand that retrieval step and you understand most of what it takes to be one of the sources.
Assistants search before they answer
A language model is trained on a snapshot of the web that ends at some point, so on its own it cannot know your new prices, this year's regulations, or that your clinic opened a second location. For anything current, specific or local, the assistant does what you would do: it runs searches, opens the results, reads them, and writes an answer from what it read. Only then does it cite the pages it leaned on.
ChatGPT's search mode, Gemini, Perplexity, Google's AI Overviews and Google's AI Mode all work this way, with different search indexes behind them. Perplexity shows the searches it runs while it works. Google's features draw on Google's own index, and its documentation for AI features and your website says there are no special requirements beyond the usual ones: the page must be indexed and eligible to be shown as a snippet, and the same robots and snippet controls apply.
That has a blunt consequence. If a page isn't in the index an assistant searches, it cannot be picked, no matter how well it's written. The crawl, index, rank pipeline from the last lesson is the entry ticket here too. Which crawlers feed which assistant, and what to allow, is its own lesson in llms.txt and AI crawlers.
One question becomes several searches
Assistants rarely search for the question as typed. Google's own description of AI Mode calls the method query fan-out: the question is broken into related sub-questions, each is searched separately, and the results are pulled together into one answer. Other assistants behave the same way, which you can watch happen in Perplexity's list of searches.
Take "which accountant should a small online store in Manchester use". The fan-out might search for e-commerce accountants in Manchester, what an e-commerce accountant does differently, typical fees for an online shop, and how VAT works for online sellers. The pages cited in the final answer are the ones that answered those sub-questions well. A firm's page that plainly explains fees for online shops can be named even though it has never ranked for the original phrase, and a page that ranks for the original phrase but answers none of the sub-questions can be skipped.
This is why "rank for the keyword" is the wrong target for AI visibility. The target is to be the clearest answer to the sub-questions an assistant asks about your subject.
Passages get cited, not pages
When the assistant reads a retrieved page, it isn't grading the whole thing. It is looking for the passage that answers the sub-question it searched for, so it can quote or paraphrase it. A passage is easy to lift when it sits under a heading that names the question, starts with the answer, and includes the specifics: a price range, a step, a named product, a figure with its source.
A passage is hard to lift when the answer is spread over several paragraphs, wrapped in "it depends" with nothing after it, or only makes sense alongside the rest of the page. Long unstructured prose can rank perfectly well on Google and still be almost impossible to quote.
The assistant also checks what it reads against the other pages it retrieved. When your page says one thing about your opening hours or your prices and a directory says another, the safest move for the model is to leave you out or hedge. Keeping your facts consistent everywhere you appear is part of being citable. How to write passages built for this is the subject of Writing for AI citations.
Why this overlaps with classic SEO
Three things the assistant needs from you are things Google already needed.
It needs to find and index your page, which is crawling and indexing, unchanged. It needs your page to come back for the sub-question searches, which is relevance, the same contest as ranking, only for more specific queries. And it needs reasons to trust you over the other retrieved pages, which is the authority Google reads from links and mentions on other sites, plus the reviews and directories an assistant can read directly.
The differences are in degree. An answer names a handful of sources where a results page lists many, so the contest is tighter. Retrieval works at passage level, so structure matters more than it does for a blue link. And because the model reads the whole web's account of you, not just your site, consistent facts and third-party mentions carry more weight than they used to. The AI search chapter goes deeper on each of these.
What to take away
- AI assistants search before they answer, so a page that isn't indexed by the search engine behind the assistant can never be cited.
- One question becomes several sub-question searches, and the pages that answer those sub-questions plainly are the ones that get named.
- Assistants quote passages, so a section with a question-shaped heading and a direct answer is worth more than a long page that covers everything.
- Being cited rests on the same three things as ranking: indexable pages, relevance to specific searches, and trust earned from the rest of the web.
Next
Every one of those sub-question searches has an intent behind it, and matching it decides which page can appear at all. Continue with Search intent: the one idea that decides which page can rank.