What the research actually shows about being cited by AI

The KDD study reveals something that very few people mention: the techniques increased the number of pages in 5th place by 115 per cent — and reduced the number of top rankings by 30 per cent. Here are the figures, with the techniques they correspond to.

The most widely cited research on AI-based citation analysis reveals something that very few people mention: the techniques were most beneficial to pages ranking far down the list — and harmed those in the top spot. Citing sources increased visibility by 115.1 per cent for a page in fifth place and reduced it by 30.3 per cent for the top-ranked page (Aggarwal et al., KDD 2024). Here is what the study actually measured, which figures correspond to which technique, and where our own Danish measurements agree and disagree.

What does the research actually show?

That targeted changes to the text can affect visibility in generative search engines — but not to the same extent, not for everyone, and not for all topics. The source is Aggarwal et al., “GEO: Generative Engine Optimisation” (KDD 2024), the first academic attempt to measure this issue systematically.

The scope of the study: a benchmark of 10,000 searches across multiple domains and datasets. In the abstract, the authors conclude that their methods can boost visibility by up to 40 per cent, and that the effect on Perplexity.ai — a real engine, not just the benchmark — reached 37 per cent. They also note that the effect varies considerably across domains.

It is worth bearing two things in mind when quoting these figures. Firstly, ‘up to 40 per cent’ is a maximum for one of two metrics (Position-Adjusted Word Count), not an average. Secondly, the study is from 2024 and was based on English-language content. It is the best measurement available — it is no guarantee for your Danish website.

Which techniques worked — and by how much?

Three stood out, and they all involve incorporating verifiable material into the text: statistics, quotes and source references. The figures below are the study’s total score for Position-Adjusted Word Count, where the baseline without optimisation is 19.3.

Technique	What it is	Score (baseline 19.3)
Quotation Addition	Insert relevant quotations	27.2
Statistics Addition	Insert relevant statistics	25.2
Fluency Optimization	Make the text easier to read	24.7
Cite Sources	Cite reliable sources	24.6
Authoritative	A more authoritative tone	21.3
Keyword Stuffing	More keywords in the text	17.7

The best methods raised the baseline by 41% for one metric and 28% for the other. Note the order: the three techniques that work best all require you to provide something concrete — a figure, a quote, a source. It’s difficult to cheat your way through.

Why did the techniques help most if you’re in fifth place?

Because the effect isn’t the same for everyone. The study calls this the ‘equaliser effect’: the lower-ranked pages gained significantly, whilst the top-ranked page lost out. It’s the most surprising table in the paper, and it’s almost never reproduced.

Technique	Rank 1	Rank 3	Rank 5
Cite Sources	−30.3%	+20.4%	+115.1%
Quotation Addition	−22.9%	+3.5%	+99.7%
Statistics Addition	−20.6%	+8.1%	+97.9%

The authors’ own explanation: generative engines base their results on content, not on backlinks or domain age. Consequently, small publishers are not penalised for factors on which they could never compete in traditional search. They call this an opportunity to democratise the field — and for a Danish SME, it is the most encouraging line in the entire paper.

But do bear in mind the downsides too. All three techniques reduced the visibility of the page that was ranked number 1. So if you’re already at the top of your niche, the advice isn’t the same as for someone ranked number five. This is an important nuance that “add statistics and get cited more” doesn’t capture.

What didn’t work?

Two things, and both are worth knowing because they’re recommended anyway. Keyword stuffing — cramming in more keywords — scored 17.7 compared to the baseline of 19.3. In other words, it performed worse than doing nothing. The authors state explicitly that techniques which work in classic SEO do not necessarily translate to generative engines.

A more authoritative tone also failed to yield any significant improvement. The paper puts it as follows: generative models are already reasonably robust against such stylistic features. This echoes a widely held piece of advice: writing in a more assertive manner does not make you more citable. Having substance to back it up does.

What do our own Danish figures show?

They point in the same direction on one point and in the opposite direction on another. We asked four AI engines 576 customer questions about 465 Danish companies. The pages cited as sources had a technical score of 77.9 compared with 72.7 for the others — a highly significant difference (p=0.000). (Only Perplexity of the four conducted its own search; ChatGPT, Gemini and Claude obtained their sources from a shared search plugin — so, in their case, ‘cited’ refers to the search layer in front of the model.) This confirms the basic premise: what is on the page and how it is structured influences whether it is used as a source.

However, we also found something the study does not measure: technical setup predicted nothing about whether the company was recommended by name (74.9 versus 74.9, p=0.988). Citation and recommendation are two distinct outcomes, and the GEO research focuses solely on the former. The entire analysis centres on what determines whether AI mentions your company.

And on one point, our data directly contradicts the standard GEO recommendation: llms.txt. Among the plumbing firms visible in search results, 6 per cent of those mentioned had the file, compared with 13 per cent of those not mentioned; among accountants, the figures were 10 per cent versus 16 per cent. The figures are small and do not prove that the file is harmful — but we cannot support the claim that it makes AI mention you. The background to this is our measurement of the prevalence of llms.txt in Denmark.

How can you apply this in practice?

Start with what is best supported by evidence and what you can do without help: include specific figures, quotes and source references in the pages that are to be cited. These are the three techniques that scored highest in the study, and they require no developer.

The actual writing process — how a paragraph is formulated, where the answer is placed, how an FAQ is structured — is covered step by step in the guide to writing content that AI can cite. This page is the evidence base; it is the recipe.

Frequently asked questions

How do I get cited by AI?
The best-documented approach is to include verifiable material in the text. In the KDD study, the three techniques – Quotation Addition (quotes), Statistics Addition (statistics) and Cite Sources (source references) – scored highest, and the best methods raised the baseline by 41 per cent on one of the study’s two metrics. Our own Danish figures point in the same direction: the pages that AI cited as sources had a technical score of 77.9 compared with 72.7 (p=0.000).

Is it true that GEO can boost visibility by 40 per cent?
Up to that, yes — but that is a maximum, not an average. The figure comes from Aggarwal et al. (KDD 2024), was measured on a benchmark of 10,000 English-language searches, and the authors themselves state that the effect varies considerably across domains. On Perplexity.ai, they achieved up to 37 per cent.

Why do these techniques work best for low-ranking pages?
The study refers to this as an ‘equaliser effect’. ‘Cite Sources’ increased visibility by 115.1 per cent for a page ranked 5th and reduced it by 30.3 per cent for the top ranking. The explanation is that generative engines base their response on the content itself rather than backlinks and domain age, so that smaller publishers are not penalised for factors they could never compete on.

Does cramming in more keywords help?
No. Keyword stuffing scored 17.7 compared to the baseline of 19.3 — in other words, worse than doing nothing. A more authoritative tone also failed to yield any significant improvement. Techniques from classic SEO do not automatically translate to generative engines.

Does that mean AI will then recommend my business?
No, and that’s an important distinction. The research is about being used as a SOURCE. In our own survey of 465 Danish companies, technical setup showed no measurable correlation with being RECOMMENDED by name (74.9 versus 74.9, p=0.988). It is perfectly possible to be the page that AI reads without being the company that AI points to.