What the KDD 2024 GEO Research Found (And What It Means for You)

The most-cited research on getting content into AI answers is a single peer-reviewed paper: Aggarwal et al., “GEO: Generative Engine Optimization,” presented at KDD 2024. It tested nine optimisation methods across roughly 10,000 queries and 25 domains against a live generative engine. The short version: adding quotations, citations and statistics boosted visibility by up to 40%, keyword stuffing made things worse, and the biggest winners were pages that weren’t already ranked first. This post walks through what the study actually found and what each finding means for a page you own.

What is the KDD 2024 GEO study?

The KDD 2024 GEO study is a peer-reviewed paper by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, published in the proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining in August 2024. It was the first rigorous, large-scale attempt to answer a question the whole industry was guessing at: what actually makes a generative engine cite your content?

The team built a benchmark of around 10,000 queries drawn from real sources across 25 domains, applied nine different content optimisation methods, and measured each against a live generative engine using visibility metrics they designed for the purpose. Because it is peer-reviewed and tested at scale, it carries weight that the vendor blog posts flooding this topic do not.

What worked, in order of effect

The researchers found that content-level changes could, in their words, “boost the visibility of source content in generative engine responses by up to 40%.” The methods that produced those gains, ranked by effect:

• Adding quotations. Including relevant quotes from credible sources was the single strongest method tested. A well-placed, attributable quote gives the engine something concrete and trustworthy to lift.
• Citing sources. Linking out to authoritative references for your claims. Counterintuitively for SEO teams trained to hoard link equity, sending readers to good sources made content more likely to be cited.
• Adding statistics. Replacing qualitative claims with quantitative, attributable ones. “Improves conversions” is unquotable; “lifted conversions 18% over three months” is a sentence an engine can pull.
• Fluency and simplicity. Cleaner, more readable prose that is easier for both a reader and a model to parse.

Notice the pattern. Every winning method adds substance an engine can verify and attribute. None of them is a formatting trick or a density tweak.

What failed, and why it matters more

The failures are the more useful half of the paper, because they contradict a decade of SEO instinct.

Keyword stuffing scored below doing nothing. Content optimised with traditional keyword-density tactics was less visible in generative engines than the unoptimised original. This is the finding to sit with. The tactic that built the SEO industry doesn’t merely fail to transfer to GEO — it damages you.

Sounding authoritative without being authoritative did nothing. Cosmetic tone changes that made content read as more expert, without adding any real substance, produced no significant improvement. The engines are already robust to that. You cannot bluff a retrieval system with confident phrasing; you have to give it verifiable material.

The finding that changes strategy for underdogs

Buried in the results is the most commercially important finding for any brand that isn’t already dominant: the optimisations helped lower-ranked pages far more than top-ranked ones.

A source sitting around the fifth position roughly doubled its visibility after adding citations, quotations or statistics. The source ranked first actually lost share when everyone else optimised. In classic SEO, the gap between position one and position five is enormous and hard to close. In generative engines, position five is a genuinely winnable starting point — sometimes a better one, because the top page has the most to lose.

For a newer agency, a challenger brand, or a business in a crowded category, this inverts the usual math. You do not need to outrank the incumbent to be cited alongside or instead of them. You need to be more quotable than they are.

Applying the findings to a real page

Turn the research into a checklist you can run against any post:

• Does at least one section contain a named quotation from a credible source? If not, find one and add it.
• Does every factual claim that could be doubted link to an authoritative reference? Vague claims with no source are never quoted.
• Have you replaced qualitative claims (“many,” “significant,” “often”) with real numbers and dates wherever you honestly can?
• Is the prose clean and direct, or padded with keyword repetition that the research says is actively hurting you?

One honest caveat to carry into client conversations: the study used the generation and retrieval setup available in 2024, so absolute magnitudes will differ on today’s engines. Treat the direction as robust — substance-adding beats keyword-manipulating — and the specific percentages as indicative rather than a promise.

Frequently asked questions

The bottom line for your content

Strip away the marketing noise around GEO and one peer-reviewed result underpins almost all of it: generative engines reward content that gives them something real to quote, and punish content that tries to manipulate them with density. Quotations, citations and statistics work. Keyword stuffing backfires. And the pages best positioned to win are the ones currently sitting just outside the top spot.

We build AI visibility on exactly this evidence base — measured, not guessed. See how our AI Visibility service turns these findings into monthly citation gains.