Where AI answers actually come from

Everyone selling AI visibility tells you the same thing: get into directories, get on lists, get mentioned. It is not wrong. It is just not the whole picture — and the part that gets left out is the part a small business can actually act on.

When someone asks ChatGPT for a wedding planner on Lake Como or an energy consultant near Ingolstadt, the model does not invent an answer out of thin air. It searches, reads a handful of pages, and writes a short answer naming a few businesses. Which pages it reads decides who gets named. So the only question that matters is: which pages does it read?

Last year the standard answer hardened into dogma: it reads third parties. Directories, listicles, Reddit, review sites. Your own website barely counts. Then people started measuring.

The number that breaks the dogma

Read that again, because it contradicts almost every GEO checklist in circulation. The model's most common source was not a directory. It was the business's own site.

A separate study of 282 million citations found the same instinct pointing in different directions depending on the industry: the share of citations going to listicles swings between 14% and 45% depending on the sector (Writesonic). There is no universal format that wins. There is only what wins in your category.

And a third dataset found that 88% of AI citations land on brand and product pages rather than on classic article formats (Qvery). Again: your own pages, not somebody else's blog post about you.

So why does everyone say "get mentioned"?

Because they are also right — just about a different thing. Being cited and being recommended are not the same event.

Ahrefs looked at 75,000 brands and correlated visibility in Google's AI Overviews against dozens of signals. The strongest correlation was not backlinks and not domain authority. It was branded mentions — how often the web talks about you by name, whether or not it links to you (Ahrefs).

What this means if you have five employees

The dogma sends you chasing things you cannot get. A German company needs €100 million in revenue or 1,000 employees to clear Wikipedia's notability bar (Wikipedia) — for a five-person firm that door is closed, permanently. Chasing it is a waste of a quarter.

The measurable part is closer to home. In the audits we run, the most common finding is not a missing directory entry. It is a page that already answers the question perfectly — and that no crawler can find, because it is in no sitemap and nothing links to it. The content exists. It is simply invisible. Fixing that costs an afternoon and moves the score more than three months of outreach.

The honest order of operations, then:

First, be quotable. One page per real question a customer asks, answering it directly in the first two sentences, with a number or a date in it. Reachable, linked, in the sitemap.

Then, be known. Get named in the places your category actually gets cited from — which, per the citation data, differs wildly by industry. For B2B software, review sites: 51% of buyers now begin their research inside an AI chatbot, and for 45% a review-site citation is the strongest trust signal (G2). For a wedding planner, it is curated lists. For a trade, it is the regional press. Find out which, before you spend a euro.

That is the whole method, and it is not glamorous: ask the AI the questions your customers ask, write down who it names and which sources it opens, and go be present in exactly those places. Everything else is guessing with confidence.