What Actually Gets You Cited in AI Answers (What the Research Says)

Ranking on Google, I understand. Getting an assistant to lift your exact page into its answer is a newer and stranger thing, and most of the advice about it is guesswork wearing a confident face. So I went looking for the people who actually measured it, because a controlled study beats my hunches every time.

The short version: the research points at a few unglamorous things, and none of them are the tricks being sold.

The one peer-reviewed study, and what it found

Princeton’s GEO paper (Aggarwal and colleagues, accepted at KDD 2024) is still the only academic work I know of that tested this properly. They rewrote source content in different ways and measured how often generative engines then cited it. The changes that worked lifted visibility by up to 40 percent, and they were boring on purpose: add relevant statistics, add direct quotations, cite your own outside sources, and write in clear, authoritative language. The change that did nothing was keyword stuffing, the tactic most people still reach for first.

Read that list again. Every winning move makes a page easier to quote and harder to argue with. The model is not rewarding effort. It is rewarding text it can lift cleanly into an answer and stand behind.

Citation is not the same as ranking

Here is the part that reorganized my head. Ahrefs compared AI citations against normal search results and found that, across ChatGPT, Gemini and Copilot, only 12 percent of the URLs those tools cited also appeared in Google’s top 10 for the same prompt. So the page winning the AI answer is usually not the page winning the blue links. Two different games, running at once.

Freshness is another lever people underrate. In a study of 17 million citations, Ahrefs found assistants lean toward newer content, with ChatGPT citing pages that were on average hundreds of days newer than the ones ranking organically. A page nobody has touched in three years is quietly at a disadvantage even when it is correct.

What I take from it, cautiously

If I were changing a page to get cited, the research says to do the dull things: put a clear, quotable answer near the top, back claims with a named statistic and a real source, keep the language plain and confident, and keep the thing current. Not to game an entity graph or stuff a keyword.

I hold this loosely, because these tools rewrite their own rules every few months, and a finding from this spring may not survive the autumn. But the pattern under all of it feels stable in a way the specifics do not: a machine cites what it can safely quote. Make the true thing about you quotable, and you have done most of the work the studies can actually see.