How an AI assistant decides whom to name

What happens after the client presses Enter?

Roughly three things, in this order.

The question is split. "Which accountant in Maribor works with small online shops?" becomes several narrower searches: accountants in Maribor, accounting for e-commerce, reviews, prices. Google describes this as query fan-out. A page can be found through a sub-question the client never typed.

Pages are read. The assistant fetches the pages it found and writes the answer from them. This is retrieval. It is why a current, readable page counts for more than what the model happened to learn in training — the model's knowledge stops at its training cut-off.

Claims are tied to sources. Many assistants show where each part of the answer came from. Microsoft calls this grounding. A claim that only appears on your own site is weaker than one that a register, a directory or a trade publication confirms.

Why can a well-known company be left out?

Because two different questions are being answered. Asked "who is Company X?", the assistant describes you from whatever it can find. Asked "who should I hire for this?", it compares options and names the ones it can justify for this particular need.

The Slovenian games studio Outfit7 found exactly this: engines described the company correctly and still recommended smaller competitors for the questions its players asked. We wrote up what they did about it. We call the gap known vs chosen, and it is the first thing we measure.

What makes a page easy to use in an answer?

The only controlled study of generative engine optimisation we know of found that pages with citations, statistics and quotations were used more often in generated answers, and that keyword stuffing made things worse. In plain terms, a page that looks like evidence is more useful to the assistant than a page that looks like an advert.

That matches what the platforms say. Google's guidance is to follow ordinary search best practice, keep important content in text, and make sure structured data matches the visible page.

What to do with this

• Write down the three or four questions your best clients would ask an assistant, in their words.
• For each, ask yourself: could a stranger confirm from public pages that we fit this client?
• Where the answer is no, that is the gap the next chapters close.