What actually gets a page cited by AI — the one study that measured it

If you’ve read anything about getting cited by AI — including in this series — you’ve met the claim: cited sources, statistics, and quotes lift AI visibility by up to 40%, and the lift is biggest for underdogs. That sentence has a single source, and it isn’t a marketing deck. It’s “GEO: Generative Engine Optimization” — Aggarwal, Murahari, and colleagues at Princeton, Georgia Tech, the Allen Institute for AI, and IIT Delhi — peer-reviewed and published at KDD 2024, one of computing’s top research venues. This is the piece that opens the box: what they actually did, what won and lost by how much, and — just as important — what one experiment can and cannot prove.

What the experiment actually was

The team built a benchmark of 10,000 real queries across 25 domains, drawn from nine existing research datasets — mostly informational questions, the kind a family types at midnight. For each query, they took real web sources, rewrote them nine different ways — adding statistics, adding quotations, citing sources, improving fluency, simplifying language, adding keywords, adding technical terms, sounding more authoritative, using rarer words — and measured whether the rewritten page earned a more prominent place in a generative engine’s answer.

“Visibility” here means something specific, and the precision matters: their metrics score how much of the AI’s answer a page’s citation earns and how early it appears — not clicks or traffic, and certainly not phone calls. The engines were the 2023 generation — GPT-3.5-turbo working over live search results, validated against the Perplexity of the day — a timestamp worth holding onto. And the benchmark and code are public, which matters for a reason we’ll get to.

What won, what lost, by how much

Three methods dominated: citing sources, adding quotations, and adding statistics — a relative visibility lift of 30 to 40% on the study’s harder metric, with the ~40% peak coming from quotation addition. A quieter result rode alongside: simply improving the fluency and readability of a page lifted visibility 15 to 30% — writing clearly is itself an AI-visibility lever, not just a kindness to the person reading.

One method lost: keyword stuffing — the oldest reflex in SEO — was the only approach to score below the do-nothing baseline, offering “little to no improvement” in the paper’s own words. The move every funeral home was taught a decade ago is dead weight in an AI answer, or slightly worse.

Then the finding this series quotes most, stated precisely: citing sources gave a page ranked fifth in search results a 115.1% visibility increase — while the average top-ranked page’s visibility fell 30.3%. The authors’ framing: lower-ranked websites, which typically struggle for visibility, benefit significantly more. Read that as a funeral home that will never out-rank a consolidator’s domain: the lever has the most travel exactly where you’re standing. Two honest footnotes ride with it — the paper warns that which method works best varies by domain, and the real-engine validation on Perplexity confirmed the direction at smaller magnitudes. No silver bullet; a strong bet.

What one study can and cannot prove

Here’s the section a vendor pitch will never include. It’s one study — influential, cited by dozens of follow-on papers (61 on Semantic Scholar as of this writing, 13 rated highly influential), but singular. The engines it tested are generations old now. Its metrics measure citation prominence, not business outcomes — a page can win the metric and never ring the phone. It was a controlled lab setup, not a study of real grievers using real assistants with location and memory in play. And the team that coined the term “generative engine optimization” is the team that validated it — mitigated by peer review, the public benchmark, and 10,000 queries, but worth saying out loud.

Nobody has cleanly re-run the experiment on today’s engines. The follow-on academic work extends the question rather than re-testing it — one 2025 analysis found AI search favoring earned media over brand-owned content, a complement to the original’s on-page finding, not a replication. What does exist is structural corroboration at scale: Ahrefs analyzed roughly four million AI Overview citations in March 2026 and found only 37.1% came from Google’s top ten organic results — the rest from pages ranked eleventh to beyond the hundredth. AI citation is demonstrably not a mirror of search rank, which is exactly the world the underdog finding predicted, observed live three years later. So the honest label is: the best evidence that exists — a very different thing from proven forever, and still far better than any competing claim’s evidence, which is typically none. The measurements this study sits among are kept, with methods, in what the AI-visibility numbers actually say.

What an owner does with it

The study’s translation costs writing effort, not ad spend, and it fits on an index card. Take the five pages that do the real work — homepage, pricing, obituaries, FAQ, about — and give each one the three features the experiment rewarded: a real statistic with a named source, a quotation from a named authority, a citation the reader could follow. Write plainly while you’re in there; fluency was a measurable lever. Delete the keyword-stuffed paragraph your last SEO vendor added; the evidence says it’s ballast. The whole job is an afternoon per page, and the audit that shows whether it’s working — or the four-assistant version — costs twenty minutes a month.

The study also hands you a vendor filter, which may be worth more than the lift. When someone sells your funeral home “guaranteed AI visibility,” ask for the mechanism. A legitimate provider can name what they’ll change — sources, statistics, quotes, structure — and tie it to the published finding you now know by name. A vendor who answers with mystique, or guarantees placement inside a system nobody controls, has told you everything. Nobody can promise what a generative engine will cite. Somebody can stack the features it measurably prefers.

Why this series looks the way it does

Forty-five percent of consumers used AI to find local businesses in the past year, on assistants used by hundreds of millions of people weekly — so which pages get cited stopped being a lab curiosity a while ago. What AEO means for a funeral home is where this series began; this study is why every post in it carries named sources, real numbers, and quotations — not as house style, but as the one strategy a controlled experiment supports. Including this post, which you may notice practices what it just explained. The receipt is attached. That’s the whole method — and it’s available to any funeral home with an afternoon and something true to say.