I read the GEO research paper. Here's what most agencies in Australia miss.
I've been getting calls from prospects asking about "GEO services." Half the agencies in Australia now claim to offer it. Most of them haven't read the paper the term came from. The findings are not what the marketing departments are telling people.
Generative Engine Optimisation is now a real category. Australian agencies are pricing GEO retainers, running GEO audits, publishing GEO checklists. Some of that work is good. Some of it is SEO with a new label. And some of it is selling techniques the actual research showed make things worse, not better.
This article is what's in the paper, in plain language, with the specific findings most agencies are quietly not mentioning. If you're considering paying someone for GEO, read this first.
Where the term "GEO" actually came from
The paper is "GEO: Generative Engine Optimization" by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik R. Narasimhan, and Ameet Deshpande. It was first posted to arXiv in November 2023 and published at KDD '24, the ACM's premier conference on knowledge discovery and data mining, held in Barcelona in August 2024. The authors are affiliated with Princeton University, IIT Delhi, Georgia Tech, and the Allen Institute for AI — none of them digital marketing agencies.
That distinction matters. Most discussions of "GEO" in the agency world are downstream interpretations of this paper. The paper itself is a piece of computer-science research. It frames generative engines as a new class of system, defines a measurement framework, and reports experimental results from testing nine specific optimisation techniques across 10,000 queries.
It is also the source of the visibility numbers that get quoted on agency landing pages — the "boost AI visibility by 40%" claim that turns up everywhere. That number is from this paper. It's real. But the context in which it's true is narrower than most marketing copy suggests.
How generative engines actually work — the three-stage architecture
The paper's first contribution is a framework. Before you can optimise for something, you have to know what it does. The authors describe generative engines as three-stage pipelines:
- Retrieval. The engine fetches a set of candidate documents from an underlying index. This is the part that looks most like traditional search.
- Summarisation. Multiple sources are compressed into intermediate representations the language model can work with.
- Response generation. The model writes a final answer, choosing which sources to cite and how to weave them together.
To appear in an AI answer, your content needs to succeed at all three stages. Most SEO work targets only the first. Schema, link-building, and on-page optimisation help with retrieval — they get you into the candidate set. But getting retrieved is not getting cited. The summarisation and response-generation stages have their own selection criteria, and those criteria are different from Google's ranking signals.
Getting retrieved is not getting cited. Different stages, different signals.
This is why a business can rank well on Google and still be invisible in AI Overviews. The retrieval stage worked — the AI engine fetched your page among others. The synthesis stage chose to cite someone else.
The nine techniques the researchers tested
The core experimental contribution is testing nine specific techniques on a benchmark called GEO-bench: 10,000 queries spanning diverse domains, run through real generative engines including Perplexity. Each technique was applied to candidate sources, and visibility was measured before and after using two metrics introduced in the paper: position-adjusted word count and subjective impression score.
The nine techniques fall into three rough categories. Some are familiar from SEO. Some are specific to generative engines. One of them is the technique most SEO professionals reflexively reach for — and the paper's finding on it is the most interesting result in the whole study.
The headline numbers are real. The bottom number is the one that should worry you if you're paying for "GEO" today.
Which techniques worked, and which actually hurt
The three clear winners — citations, quotations, and statistics — improved visibility by more than 40% on the GEO-bench benchmark. That number is what every agency now quotes. What they don't tend to mention is that these are the only techniques that produced that scale of lift. The other six options ranged from moderately useful to actively counterproductive.
Fluency optimisation — making content easier to read — produced a smaller but real lift. Authoritative tone and accessibility helped, but moderately. Technical-term inclusion and unique-word optimisation came out close to neutral once domain effects were controlled for.
And then there's keyword stuffing.
The researchers tested keyword stuffing — the SEO technique of repeating target query terms throughout the content — as one of the nine GEO methods. It produced a measurable decrease in visibility of around 10% across the GEO-bench test set.
This is the single most consequential finding for anyone selling "GEO services" today. The technique the SEO industry has spent two decades refining doesn't just fail to help in generative engines. It makes things measurably worse.
The intuition behind the finding is straightforward. Traditional search engines use keyword matching as a primary ranking signal — repetition correlates with relevance. Generative engines don't rank pages; they synthesise answers from sources they assess as credible. Repetitive keyword usage looks, to a language model, like exactly the kind of low-quality content it has been trained to discount. The very signal that helps in classic SEO becomes a negative signal in GEO.
Most Australian agencies offering "GEO" services have not internalised this. Their content briefs still include keyword density targets. Their on-page optimisation still treats keyword repetition as a positive lever. Their reports still highlight keyword frequency as a win. They are, in effect, charging GEO fees to deliver the one technique the GEO research showed made things worse.
Why this matters for Australian businesses right now
If you're already paying for GEO services, three honest questions will tell you whether the work is grounded in the research or in repurposed SEO:
One. Ask your provider when the term "GEO" was coined and by whom. If they can't name Aggarwal et al. 2024 or the Princeton / IIT Delhi / Georgia Tech / Allen Institute collaboration, they're not basing the work on the research. They might still be doing useful work, but it isn't grounded in the evidence the discipline was built on.
Two. Ask them what the paper found about keyword stuffing. The answer should be: it reduced visibility by around 10%. If they don't know — or if their own content briefs still target keyword density — there's a mismatch between what they're selling and what the research supports.
Three. Ask what they're tracking. Real GEO measurement uses position-adjusted citation count or equivalent metrics. Lots of agencies report "AI visibility" using rank-tracking tools that watch for brand mentions in AI Overviews — which is fine as a proxy, but doesn't replace measurement of citation frequency, citation position, and citation share-of-voice. If the only metric is "we appeared in 3 AI Overviews this month", you're getting a tracking dashboard, not a measurement framework.
None of this means you shouldn't invest in GEO. The research is clear that the top techniques produce substantial visibility lift. It means you should know what you're paying for.
Want to know where your business actually stands on GEO?
Catnap is Australia's research-led GEO agency. We baseline your current citation rate across ChatGPT, Perplexity, Gemini, and Google AI Overviews using academic metrics from the original GEO paper. The 15-minute call is run by me personally — no slide deck, no pitch, just an honest read. See our full GEO agency services for Australian businesses.
Three things to do this month if you want to act on the research
If you're not yet engaged with an agency on GEO and you want to start applying what the paper actually shows, three concrete things move the needle most:
One: add real citations to your highest-value pages. Not link-building citations. Actual references to authoritative sources within the body text. Footnotes. "According to the Australian Bureau of Statistics" with the link. "The 2024 PWC report on AI adoption" with a link. Citation density is the single most impactful GEO lever the paper tested, and it's also one of the easiest things to retrofit onto existing content.
Two: add specific statistics, with sources. "We've worked with 250+ Australian businesses since 2016" is a real statistic. "We deliver leading AI marketing services" is a slogan. Generative engines treat them very differently. Quantified, attributable facts get cited; vague positioning language does not. Audit your service pages and find every claim that could be made specific. Make it specific.
Three: stop keyword-stuffing. Genuinely stop. If your SEO consultant is still asking you to hit "0.8% keyword density" on a target term, the request is producing negative GEO outcomes. The terms still need to be present — your pages need to be retrievable — but the relevant pattern is natural use, not repetition. A page that uses "generative engine optimisation" once in the H1, once in the meta, and twice in the body will out-perform a page that uses it twelve times.
And one thing to be patient about: the metrics. Even with the right techniques applied, GEO visibility lift compounds over months, not days. AI engines have to re-crawl, re-index, re-evaluate, and start preferring your content over alternatives. The paper's 40% figure is the result of optimisation followed by re-evaluation, not an overnight effect. If anyone promises measurable GEO results in a week, they're either lucky or lying.
The bottom line
The GEO research paper is real, peer-reviewed, and clear. Three techniques produce substantial visibility gains: citations, quotations from authoritative sources, and specific statistics. Several techniques produce modest gains. Two are roughly neutral. One — the one the SEO industry has spent twenty years refining — actually hurts.
Most agencies in Australia selling GEO today have not absorbed the third finding. They're still optimising for keyword density, still measuring keyword frequency, still treating GEO as SEO with new branding. The work isn't necessarily bad, but it's not what the research supports.
The good news is that the techniques that actually work are also the techniques that make your content genuinely more useful. Citing your sources, quoting experts, including specific data, writing fluently — none of this is exotic. It's just good content, applied systematically, measured against the right metrics. The research backs it. The agencies that read the research do it. The ones that didn't, mostly, do something else.
For more on how schema fits into all of this, see the schema tags AI engines actually read in 2026. For the full Catnap approach to citation engineering, see our GEO agency services for Australian businesses.