How AI Assistants Decide Who Is 'the Best'

When someone asks an assistant for the best villa manager or wedding photographer on an island, the machine does not weigh the work. It cannot see the work. It reads text about the work, then decides who to name. I keep coming back to that gap, because the business getting recommended is rarely the one doing the best job. It is the one the model can find a reason to trust in writing.

I cannot run a lab, but a few teams with real data have, and what they found is more useful than any afternoon of me typing queries.

What the citation studies actually measured

Ahrefs looked at 1.4 million ChatGPT prompts to see why it cites one page and skips another. Their analysis found the biggest factor was not quality at all. It was the retrieval channel: pages pulled from the search index got cited far more than anything else. After that came how closely a page title matched the question, cited pages averaged a cosine similarity of 0.602 to the prompt against 0.484 for pages that were skipped, and even something as dull as a clean, readable URL slug moved the citation rate up several points.

None of those measure who is good at the job. They measure who is legible to a machine.

Why “the best” is really “the most quotable”

Princeton’s GEO paper, the first peer-reviewed work on this and accepted at KDD 2024, tested changes to source content across generative engines. It found that adding statistics, adding direct quotations, citing outside sources, and writing in a clear, authoritative register could lift a source’s visibility in AI answers by up to 40 percent. Keyword stuffing, the old trick, did nothing.

So the levers that get you named are not “be better.” They are “be easier to quote and harder to misread.” A business with a clear, sourced, well-structured page can out-appear a better business that only exists as a phone number and a good reputation nobody wrote down.

There is a second bias stacked on top. Ahrefs’ work on the most-cited domains across ChatGPT, Perplexity and AI Overviews found the same handful of sites appearing over and over, with Wikipedia and YouTube near the top of nearly every list. Assistants lean on a small set of sources they already treat as safe, which means being mentioned on those sources often matters more than anything on your own site.

What I take from it

If the studies are right, being the named answer is a writing and reputation problem wearing a technology costume. The machine is not judging the work. It is judging how findable, quotable and corroborated the work is in text it can reach.

I find that clarifying and slightly grim. The honest move is not to trick the model. It is to make the true thing about you easy to read, and to get it said in the places the model already trusts. I might be wrong about the exact weightings, but every dataset I can find points the same way: legibility beats excellence when a machine is doing the choosing.