Do you need to rank on Google to get cited by AI?

Most GEO advice assumes AI visibility is a prize you collect after you win at search. Three datasets say the two have come apart — and the gap between them is where a small company gets its opening.

A plumber in Augsburg sits on page three for the one phrase he cares about, and has been told that until he cracks the top ten, ChatGPT will never mention him. It is a reasonable assumption. It is also measurably wrong, and the measurements are not subtle.

Ahrefs ran 15,000 long-tail queries through Google and Bing, put the same questions to ChatGPT, Gemini, Copilot and Perplexity, then compared what the assistants cited against what actually ranked (Ahrefs).

The average hides a split worth knowing about. Perplexity, which was built around citing its sources, landed 28.6% of its citations on top-10 pages. For ChatGPT, Gemini and Copilot the figure sits near 8%. There is no single "AI visibility" here; there are engines with very different habits.

The reason is mechanical, and Google documents it plainly. Both AI Overviews and AI Mode use a "query fan-out" — the system issues multiple related searches across subtopics rather than one search for the words you typed, which lets it show "a wider and more diverse set of helpful links" than a classic result page. Pages that keep surfacing across those sub-searches get cited. A page sitting sixth for three neighbouring questions can beat a page sitting first for one.

The exception is closing too

AI Overviews used to be the place where classic rankings still ruled. In July 2025, Ahrefs found 76% of AI Overview citations came from top-10 pages. They repeated the study this March across 863,000 SERPs and four million cited URLs.

The tempting reading is that Google has stopped caring about rankings. That is not what the data says. BrightEdge tracked AI Overview citations across nine industries for sixteen months and found the opposite trend at a coarser level: the share of citations coming from pages that rank organically at all grew from 32.3% to 54.5% (BrightEdge). What changed is where in the rankings those pages sit. Only 16.7% of citations came from the top ten. Nearly all the growth came from positions 21 to 100.

Industry decides how much of this applies to you. In the BrightEdge data, healthcare sat at 75.3% overlap with organic rankings and education at 72.6%, while restaurants sat at 19.2% and e-commerce at 22.9%. If you run a restaurant, roughly four out of five sources behind an AI answer in your category are not pages that rank organically for the query. Ranking work does not reach them. Knowing which of those two worlds you are in is worth more than any checklist.

What this changes for a company with five people

It does not mean search work is optional. Google's own documentation is blunt about the entry ticket: to appear as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, and there are no additional technical requirements. A page blocked in robots.txt or carrying a nosnippet directive is not in the running, and a page nothing on your site links to may never get indexed in the first place. Being nowhere in the index is still being nowhere.

What it changes is the target. Beating three national firms for a head term is a multi-year project with an uncertain ending. Ranking somewhere in the top fifty for forty specific sub-questions — what a thing costs in 2026, whether a rule applies to your case, how two options differ — is a few months of writing, and those are the pages fan-out actually retrieves.

It also changes what you measure. If your rank report stops at position 10, you are blind to the band where AI citations are growing fastest. Widen the tracking to 100, or skip the proxy and measure the citations themselves.

None of this makes the work smaller. It makes it different, and the different version happens to suit a small firm: many narrow answers instead of one contested keyword. You will not out-rank a competitor with thirty people writing content. You can out-answer them on the forty questions they consider too small to bother with.