The paper that named generative engine optimisation found that adding quotations, statistics and citations to a source made it noticeably more visible in AI-generated answers, while keyword stuffing made it less visible. The best methods improved visibility by roughly 40% in the authors' tests. The gains were largest for sources that started lower down the list of retrieved results.
Most GEO advice you'll read traces back to this one study, often loosely. This post covers what it actually tested and what it means for a business website.
The study in brief
GEO: Generative Engine Optimization was written by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, and accepted at KDD 2024. The team built GEO-bench, a set of 10,000 queries across many domains, split into training, validation and test sets.
For each query, a generative engine retrieved five sources and wrote an answer citing them. The researchers then rewrote one source using a specific method and measured how much of the answer came from it. Two main measures were used: a position-adjusted word count (how many words from the answer are attributed to the source, weighted by where they appear) and a subjective impression score judged by a model.
What worked and what didn't
On the position-adjusted word count, compared with an unmodified baseline, the results from the paper's main table work out roughly as follows:
- Quotation addition (adding relevant quotes from credible people or sources): about 41% better.
- Statistics addition (adding specific numbers): about 31% better.
- Fluency optimisation (cleaner, smoother writing): about 28% better.
- Citing sources: about 27% better.
- Technical terms: about 18% better.
- Easy-to-understand rewriting: about 14% better.
- Authoritative tone: about 10% better.
- Unique words: about 6% better.
- Keyword stuffing: about 8% worse.
The authors also tested on Perplexity, a real product. Quotation addition improved the word-count measure by about 22% there, and statistics addition improved the subjective score by about 37%. Keyword stuffing was again worse than doing nothing.
The finding most people miss
The paper broke results down by the source's original rank. For sources ranked fifth of five, citing sources improved visibility by 115%, quotation addition by nearly 100% and statistics by about 98%. For sources ranked first, the same methods reduced visibility, by around 20 to 30%.
In plain terms, these methods help most when you're not already the obvious source. A small South African agency competing with large US publishers for a citation stands to gain more from evidence-rich content than the incumbent does.
Limits worth knowing
- It's a 2023 setup. The main engine was a research system built on a language model of that period. Today's products use different models, retrieval and ranking.
- Retrieval was fixed. Sources were already selected before the rewrite. In the real world, you first have to be retrieved at all, which depends on crawling, indexing and relevance.
- Domain effects were large. The authors note that the best method varies by subject. What helps a history answer may not help a legal one.
- Model-judged scores. The subjective impression measure was scored by a model, which carries its own biases.
So treat the numbers as direction, not promises.
How to apply it without wrecking your writing
The worst reading of this paper is "stuff every page with statistics and quotes". The better reading is that specific, sourced writing gets used. Some practical rules we follow:
- One real number per claim that needs one. "Most projects take four to eight weeks" beats "projects are delivered quickly".
- Quote people who know. A short quote from your engineer, a client (with permission) or a cited expert gives a model something attributable.
- Link to the primary source. Google's documentation, a government statistic, the original study. Not a blog that summarised it.
- Write cleanly. Fluency helped on its own. Short sentences, clear subjects, no filler.
- Never fake it. Invented statistics and made-up quotes are a reputational risk and are increasingly easy for readers to check.
Where this fits in a GEO plan
The paper only covers one stage: making a retrieved page more likely to be used. A complete GEO programme also covers being crawlable by AI search bots, being retrieved for the right sub-questions and building corroboration on other sites. Our post on how ChatGPT search picks its sources covers the earlier stages.
The research gives a useful, slightly unglamorous lesson. The content that AI systems prefer looks a lot like content a good editor would prefer: precise, sourced and easy to read.



