A prospect types "which accountants in Bristol work with cafes" into ChatGPT and gets three firms back, each with a link. None of them is the firm that ranks first on Google for "accountant Bristol". That isn't luck, and it isn't a different kind of magic. An assistant doesn't know your business. It runs a search, reads a few passages from what comes back, and names the sources it leaned on. Getting cited is a retrieval problem first and a writing problem second, and this lesson follows the path from the question to the citation.
The assistant searches before it answers
A language model on its own has only what it absorbed during training, which is out of date and vague about small businesses. So every assistant that cites sources bolts a search step onto the model: the question goes to a search index, the best results come back, and the model writes its answer from those pages. Engineers call this retrieval-augmented generation. Think of it as "search, then summarise".
Which index depends on the assistant. Google's AI Overviews, AI Mode and Gemini draw on Google's own index. OpenAI's help documentation says ChatGPT's search uses third-party search providers as well as OpenAI's own crawler. Perplexity runs its own crawler and index. The lesson on llms.txt and AI crawlers names the crawlers involved.
The practical consequence is blunt. You can't be cited from a page that isn't in the index behind the assistant. Everything in How Google ranks a page still applies: the page has to be found, indexed, and rank well enough to land in the candidate set. A page Google can't see is invisible to AI Overviews, and a page that blocks OpenAI's search crawler is invisible to ChatGPT's search.
One question becomes many
Google describes AI Mode as using a "query fan-out" technique: it breaks the question into several related searches, runs them at once, and assembles an answer from what each one returns. The other assistants behave the same way in practice, whether or not they document it.
Take the cafe question. Behind the scenes it becomes something like: what accounting services a cafe needs, accountants in Bristol who work with hospitality, how those firms charge, VAT rules for food businesses, and what reviewers say about each firm. The answer stitches those pieces together, and a citation attaches to whichever page answered each piece.
This changes what "winning" means. Your page doesn't have to beat the national firms for the headline query. It has to be the clearest answer to one of the sub-questions. A specialist page on hospitality accounting in Bristol can sit in the same answer as a national brand, because it answered the part the national brand didn't.
It reads passages, not pages
Once the search step returns pages, they're split into chunks, usually a heading and the paragraphs under it. The chunks most relevant to each sub-question are handed to the model, and the model attaches a citation to the chunk it used. Google has documented that it can rank an individual passage from a page rather than the whole page; its passage ranking announcement described exactly that, and AI answers push the idea further.
What you can observe, without any company documenting it, is that the cited page is usually the one where a short passage states the answer on its own: a heading shaped like the question, a direct answer in the first sentence, then the detail. Long pages where the answer is spread across ten paragraphs get skipped even when they rank above the page that got cited.
So the unit of competition is the passage. Each section of a page should make sense if you lifted it out and pasted it into a chat window on its own. Writing for AI citations shows how to build pages that way.
It prefers names it already recognises
The model arrives with prior knowledge of brands and topics from its training, and assistants also weigh what several sources agree on. A business that a dozen independent pages describe the same way, with the same name, the same service and the same town, is an entity the model can place. A business that exists only on its own website is a stranger it has no reason to trust.
For "which should I use" questions in particular, assistants lean on third-party roundups, directories, forums and review sites rather than a business's own homepage. That is observed behaviour, not something the companies document, but it shows up consistently enough to plan around: the answer to "best accountants for cafes in Bristol" cites a list somebody else wrote, and the firms on that list get named.
Three things follow. Ask the assistant your own customer's question and note which sites it cites, then work out how to be listed on those sites. Keep your name and one-line description identical everywhere it appears, on your site, in directories and in profiles. Mark up your organisation with structured data so the name on your site matches the name the rest of the web uses. The authority chapter's Brand mentions and entity SEO lesson covers how to build this deliberately.
What to take away
- An assistant doesn't know your business; it searches an index, reads passages from the results and cites the ones it used, so you must be indexed before you can be cited.
- One question fans out into several searches, and being the clearest answer to a sub-question is enough to get named alongside bigger brands.
- Citations attach to self-contained passages, so every section of a page needs to make sense on its own.
- For "which should I use" questions, assistants cite third-party lists and reviews, which makes being mentioned elsewhere a visibility lever in its own right.
Next
Now that you know what an assistant is looking for, learn to write passages it can quote in Writing for AI citations.