How AI search works, and why it’s not that different

When someone asks an AI assistant a question, the assistant rarely answers from what it already knows. It searches the web, just like we all used to do. It takes the customer’s long-winded question, rewrites it into terms a search engine can handle, searches through an organized index of the web, and then chooses among the pages that come back what seems relevant. The answer we receive from an AI assistant is simply assembled from that shortlist of search results.

So most of the work of showing up in AI answers is the work of showing up in search. It’s one job, not two.

Trained on the past, searching the present

The large language models (LLMs) that power the AI assistants we know and love were all trained using copies of the web that stopped at some point in the past — snapshots of information in time. AI assistants know a great deal about the world in general but almost nothing about the local heating company with eleven staff. So when someone asks AI a question that is regional, specific, or about anything that might have changed recently, the assistant needs to go out and look for the information it needs.

But where AI looks for information varies. Every search engine keeps an index, a catalog of the pages it has visited, organized so it can be searched in a fraction of a second. As of September 2026, Google’s AI Overviews and AI Mode search Google’s own index, the same one behind ordinary Google Search. ChatGPT searches an index OpenAI built and maintains for quick responses, and when a paying user asks it to think longer, it draws on Google’s results as well. Other assistants use Bing, Brave, or indexes of their own. So while the engine behind each assistant is different, what’s common is that there’s an engine at all.

Lost in translation

A customer types: “my furnace keeps shutting off after a few minutes, is that dangerous, and who should I call in Belleville?” Nobody has ever searched those exact words, and no page was written for them. Specialists call this a long-tail search: long, specific, and worded the way people talk. This is how we’ve all learned to talk with AI assistants and it’s a fundamental shift from the old two- or three-word search we typed into Google.

Before the AI assistant looks anything up, it breaks the question into two or three searches an engine can handle. Something like: “furnace shuts off after a few minutes,” “furnace short cycling carbon monoxide risk,” and “furnace repair Belleville.” Google calls this query fan-out, and describes it in its own documentation as issuing several related searches across subtopics to build a response. This has two side effects worth knowing about:

1. The searches you can see in your own search reports don’t look like the questions your customers ask, because those questions are translated before they reach any search engine. You’re seeing the translation.
2. One question becomes several searches, and the answer is built from whichever pages covered the most of it. If your site comes up for “furnace repair Belleville” and nothing else, you’re in the running for the last third of the question. The pages that explained why the furnace keeps shutting off, and whether that’s dangerous, get to write the rest, and they may not be yours.

Being found and being cited are two different objectives

Once an AI assistant obtains a shortlist of potential sources, what happens next depends on how much effort the assistant is allowed to spend.

In its fast mode, ChatGPT mostly works from page titles and a couple of hundred characters of text, and opens almost nothing. In its slower mode it opens pages and reads them, and the difference is stark: one study this summer found that a page ChatGPT opened was cited about three times in four, while a page it retrieved but never opened was cited less than one time in ten. Google’s side is quieter about the mechanics but says the same thing about eligibility: a page needs to be indexed and eligible to appear in ordinary search results with a snippet, and nothing beyond that.

So there are two things you need to get right. The first is being found: making it onto the AI’s shortlist of sources, which is just search engine ranking under a different name. The second is being cited, which comes down to whether your page answers the question in text the assistant can see, near the top, without making it dig. A page that buries the answer under a menu, a slideshow, and three paragraphs of introduction gets found every time and cited almost never.

Where the assistants differ

The AI assistants don’t all do this the same way. Each one searches a different index, splits a question into a different number of searches, and reads more or less of each page before deciding which sources to name. Those details also shift inside each assistant over time, sometimes from one week to the next, and mostly without anyone announcing it.

What doesn’t change is the sequence. A question comes in, it’s rewritten into searches, the searches produce a shortlist, and the assistant picks its sources from that. As long as your pages hold up at each of those four steps, which assistant is asking matters a lot less.

One job, not two

If assistants find you by searching, then working on AI visibility separately from search means paying for the same thing twice. A lot of the tools built to watch AI answers work this way. They can tell you whether you were named, but they don’t have your search data, so they can’t tell you about the search that put you on the shortlist, or left you off it. And the shortlist is nothing more than the results of the searches the assistant ran. If your pages rank for those searches, you’re on it. That’s the same page-one work it has always been.

What the AI side adds is being cited, and that has less to do with search technique than with whether a page answers the question a customer would ask, in the words a customer would use. Google has been rewarding the same thing for years; the assistants have made it impossible to ignore.

Would you know?

None of this is visible from where you sit. A customer asks, the assistant searches, someone gets named, and the conversation leaves no trace in your reports. Google now shows when your pages appear in its own AI answers; nothing you already have shows you a ChatGPT answer that named a competitor, or one that named nobody at all.

The useful question isn’t whether AI assistants are answering questions about your line of work (they are), it’s whether you’d know what they said, and whether that’s measured alongside the search results it comes from rather than checked once and forgotten.