The study that named GEO, and what it proved works
The structure and markup of this piece were refreshed for current answer engines; the original analysis is preserved as written.
For a year, "optimising for AI" was mostly intuition. At KDD 2024 in Barcelona, a Princeton-led paper changed that: it coined the term Generative Engine Optimization and ran the first controlled experiment on what actually moves a source’s visibility in a generated answer. The findings are clear, and a little counter-intuitive.
the short answer
At KDD 2024, Aggarwal et al. presented the paper that coined Generative Engine Optimization and tested it on GEO-bench (10,000 queries). Adding statistics, citations and quotations lifted a source’s visibility by up to ~40%; keyword stuffing fell below baseline. The best method depends on the domain — there is no universal trick, but credibility signals consistently win.
key takeaways
- At KDD 2024 (Barcelona, August), Aggarwal et al. (Princeton) presented the paper that coined "Generative Engine Optimization" and gave the first controlled evidence of what works.
- Using GEO-bench (10,000 queries), it measured that adding statistics, citations and quotations lifts a source’s visibility by up to ~40 percent (statistics +41%; quotations and cite-sources ~+28%).
- Citing external credible sources helped lower-ranked content most (up to +115%): pointing outward to others, paradoxically, makes your own page more citable.
- Keyword stuffing — a classic SEO trick — scored below baseline: traditional search tactics do not transfer to generative engines.
- Effectiveness is domain-dependent (legal/factual favours data; historical favours quotations; opinion favours a confident tone): there is no universal tactic.
visibility lift by method (GEO-bench, position-adjusted)
The shape of the result is the message. The credibility methods — statistics, citations, quotations, even just clearer writing — all push a source’s visibility up, with statistics leading at around +41%. Keyword stuffing, the one old-SEO trick in the set, lands below the baseline: it makes you less visible than doing nothing. The engines reward substance and punish manipulation.
Why a measured baseline changes the conversation
Up to this point, anyone could claim anything about optimising for AI, and no one could be checked. The field ran on plausible assertions — add an FAQ, sound authoritative, structure your headings — none of which had been tested against an outcome. The value of this paper is less any single tactic and more that it set a baseline: a way to say "this change produced this measured effect on this benchmark," which is the difference between a field and a folklore. Once a measured baseline exists, claims can be compared to it, weak tactics can be retired, and the strong ones can be invested in with more than hope. For a buyer trying to tell real GEO advice from confident guessing, the existence of controlled evidence is the first solid ground to stand on.
It also reframes what credibility means in practice. The paper’s strongest, most consistent finding is that the markers of careful sourcing — real statistics, named citations, direct quotations — are what generative engines reward, and that the effect is largest for content not already winning. That is an unusually encouraging result, because it means the route to AI visibility is not a new bag of tricks but the old virtue of being genuinely well-sourced, applied with the new engines in mind. The teams that were already rigorous about evidence have a head start; the teams that leaned on keyword mechanics have to unlearn a habit that now costs them. Either way, the paper points the same direction the AC Group has always argued: earn visibility by being the more credible source, not the more optimised one.
The findings, in three parts
What the study actually measured, which methods worked and which backfired, and why the best method depends on your domain. Open each layer for the part that changes how you optimise content.
01 What the paper measured
Before this work, optimising for generative engines was guesswork — sensible-sounding rules of thumb with nothing behind them. The paper’s contribution was to make the question measurable. The team built GEO-bench, a benchmark of ten thousand diverse queries, and defined visibility metrics suited to generative answers, where sources are not listed in rank order but woven into a single synthesised response. Their core metric, a position-adjusted impression score, captures not just whether your source appears in the answer but how prominently — early, central, and substantial counts for more than a passing mention at the end. With that instrument they could take a passage, apply one specific content change, and measure whether the change made the source more or less visible in the generated answer. That is the quiet importance of the paper: not any single tactic, but turning GEO from assertion into something you can run an experiment on.
02 What worked, and what backfired
The headline result is that credibility pays. The methods that lifted visibility most were the ones that make a passage more verifiable and authoritative: adding relevant statistics gave the largest gain, and adding citations and quotations from credible sources followed close behind, each pushing source visibility up by a meaningful margin and, together, by as much as around forty percent on some queries. Strikingly, simply improving the fluency of the writing — making it clearer, without adding any new information — produced a comparable lift, which says something about how much clean, parseable prose matters to a model summarising you. The mirror-image finding is just as useful: keyword stuffing, the old habit of repeating query terms, scored below the untouched baseline. The engines read for meaning, not term frequency, so the manipulation that once moved rankings now actively hurts. Credibility up, trickery down — a clean, encouraging pattern.
03 Why it depends on the domain
The paper’s most practical nuance is that there is no single winning tactic — what helps most depends on the kind of content. Their results showed the best method shifting by domain: factual and legal questions benefited most from hard data and cited sources, where verifiability is everything; historical and cultural topics gained more from expert quotations and a more persuasive style, the registers that carry weight in those debates; and opinion-oriented content did best with a confident, evidence-backed tone. The lesson for a content team is to stop looking for the universal GEO trick and start matching the method to the material. A statistics-heavy treatment that wins for a technical comparison page may do little for a narrative piece, and a quotation-rich approach that lifts a thought-leadership essay may be wasted on a specifications table. The evidence rewards judgment, not a checklist applied blindly — which is exactly the kind of editorial judgment the AC Group has built around for 27 years.
From the benchmark to your pages
A study is only useful if it changes what you do on Monday, and this one translates cleanly. Start by auditing your most important pages against what the paper rewards: do your claims carry real statistics, are they backed by named citations and direct quotations from credible sources, or do they assert things bare? Most business content fails this test not because it is wrong but because it is unsupported — confident sentences with nothing for an engine to verify. Adding the evidence that was always implicit, making the sourcing visible on the page, is the single change the study suggests will move your visibility most, and it needs no new tools at all to do.
Then apply the domain lesson rather than a blanket recipe: a technical or factual page wants hard data and cited sources, a narrative piece wants quotations and clear prose, an opinionated stance wants a confident, evidence-backed voice. And do the cheap negative work — remove the keyword-stuffed passages a previous era of SEO left behind, since the study suggests they now lower your visibility. Three moves, all drawn from the evidence rather than a hunch.
The GEO study: quick answers
Is a ~40 percent lift realistic for my content?
Treat the headline number as a ceiling observed under controlled conditions, not a promise for any one page. The study measured the lift on a benchmark of ten thousand queries, with a specific visibility metric, comparing a modified passage against its unmodified self — a clean experiment that isolates the effect of one change. Your content is messier: it competes with different sources, on different topics, in engines that keep changing. So the honest expectation is not "every page gains forty percent," it is "credibility-oriented edits move source visibility in a real, measured direction, and the size depends on your starting point and domain." Notably, the paper found the biggest gains from citing sources accrued to content that was not already well-positioned, which suggests the lift is largest exactly where you have the most to gain. The right reading is directional confidence, not a guaranteed multiplier: these changes help, often substantially, and they help most when you are behind.
Why would citing other sources make ME more citable?
It feels backwards, and it is one of the paper’s more striking results, so it is worth sitting with. A generative engine is trying to assemble a trustworthy answer, and a passage that backs its claims with citations and quotations reads as more careful and verifiable than one that asserts things bare. The model is, in effect, rewarding the markers of good sourcing — the same markers a careful human reader uses to decide whether to trust a page. Citing credible sources signals that your content is doing the work of grounding itself, which makes the engine more willing to lean on it and pass it through into the answer. The effect was strongest for content that was not already prominent, which fits: a lesser-known page that is visibly well-sourced gives the engine a reason to trust it that its ranking alone would not. So the counter-intuitive move — pointing outward to others — is part of what earns you the citation, because it is evidence of the very credibility the engine is looking for.
Does this mean my SEO skills are useless for GEO?
No, but it means some of the oldest reflexes actively backfire, which is the part worth internalising. The study tested keyword stuffing — the classic tactic of loading a page with repeated query terms — and found it scored below the untouched baseline, worse than doing nothing. That is because generative engines run on language models that read for meaning, not for term frequency, so cramming keywords signals manipulation rather than relevance. The skills that transfer are the ones that were always about genuine quality: understanding search intent, structuring content clearly, covering a topic thoroughly. The habits that do not transfer are the mechanical ones that treated the engine as a counter to be gamed. So GEO is not a repudiation of everything you know; it is a sharp reminder that the gameable parts of old SEO were always the weakest, and the parts about real substance are exactly what the new engines reward.
Is one study enough to act on?
One study is never the last word, and the responsible stance is to act on its direction while holding its precise numbers loosely. What makes this paper worth acting on is not certainty but that it is the first careful, controlled measurement in a field that had been running on pure intuition — it replaced "we think citations help" with "citations measurably helped, by roughly this much, under these conditions." That is a real upgrade in evidence, even if later work refines the figures or finds limits the first study missed. The practical move is to treat its findings as well-supported hypotheses: add statistics and citations and clear sourcing because the best evidence available says they help and the cost of doing so is low, while watching your own results rather than assuming the benchmark transfers exactly. Acting on the best current evidence, and staying ready to update, is simply how you operate in a field this young — which is the stance the AC Group has taken for 27 years.
A note on sources and timing
This is written in August 2024, as the paper is presented at KDD in Barcelona. Everything above comes from the study itself — the GEO-bench benchmark, the visibility metrics, the lift from statistics, citations and quotations, the failure of keyword stuffing, and the domain-dependence of the best method. We have not reached for the later commentary and follow-up measurements that built on this work, nor the larger citation studies that came afterward, because they did not exist as we wrote, and importing them backward would misrepresent what was known in August 2024. We have also tried not to overstate a single study: its numbers are the best evidence available, not the final word, and the responsible reading is to act on their direction while holding the exact figures loosely. What is solid is the shift the paper marks — from optimising for AI by intuition to optimising by evidence — and that shift is one the AC Group has been preparing clients for across 27 years of watching authority, not tricks, win out.