What makes an AI cite your site, according to the study that founded GEO

More and more people search by asking an artificial intelligence instead of typing into a search box. The answer they get cites a handful of websites and leaves out every other one, and what decides that is not a ranking algorithm but a language model summarising five sources.

Plenty has been written about how to get onto that list and very little has been measured. This article goes through what actually was measured, in the academic work that gave the field its name, and separates its findings from what they get made to say.

What is GEO?

GEO — generative engine optimization — is the set of changes to a website's content aimed at getting an AI system to cite it when answering. It is also called AEO, answer engine optimization. What is being chased is no longer a slot in a list of links: it is appearing inside the text the model writes, early and with enough room.

The term and the framework were introduced by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande in GEO: Generative Engine Optimization, presented at the KDD 2024 conference.

What exactly was measured

This section matters more than the results, because it sets how far they can be stretched.

The authors built GEO-bench, a benchmark of 10,000 queries spread across 25 domains, drawn from nine different sources and following the real distribution of search intent: 80% informational, 10% transactional and 10% navigational.

The generative engine in the main experiment is not a commercial product. It is a two-step reconstruction: the top five sources are retrieved from Google search, and gpt-3.5-turbo is asked to write the answer from them. The authors also repeat the evaluation on Perplexity.ai, which is a deployed engine, to check the methods generalise.

The main metric is position-adjusted word count: how many words of the answer come from a given source and where in the text they appear. It is a measure of presence within the answer. It does not measure visits, clicks or sales.

What worked and what did not

The study tests nine content modifications against an unoptimised site. These are the results, calculated by us from the absolute figures in their table 1: unoptimised content scores 19.3, and each method is compared against that baseline.

Two readings stand out. The first is that what works is supplying verifiable material: attributed quotes, figures and references. The second is that keyword stuffing, the classic SEO tactic, scores below doing nothing at all. The authors say so explicitly: techniques that work in search engines do not carry over to generative ones.

It is also striking how little an authoritative tone contributes. Writing with more authority does not convince a model; giving it data it can attribute does.

The wrinkle almost nobody mentions

Here is the study's most awkward finding, and the one that rarely shows up in commercial writing about GEO.

The authors repeated the experiment splitting sites by the position they held in the search engine. The effect changes sign:

For a site that is already the top result, applying these methods reduced its visibility in the generated answer. For the one sitting at position five, it roughly doubled it.

The explanation the authors give is that traditional search rewards inbound links and domain presence, both hard for a small site to build, while the generative engine attends mostly to the content itself. That makes GEO a disproportionately useful lever for anyone who is not the leader of their niche, which is the situation of almost every small business.

What the study does not say

Three limits worth keeping in mind before making expensive decisions:

It does not measure traffic or business. It measures presence within an answer. That a company gets cited more often in ChatGPT and that this turns into customers are two different claims, and the second is not demonstrated in this work.

The experiments are from 2023 and 2024. The main engine is gpt-3.5-turbo over Google's top five results. Today's systems retrieve and reason differently. The direction of the findings is probably stable — verifiable material over filler — but the specific percentages should not be treated as a price list.

Effectiveness varies a lot by sector. The authors stress this themselves, and their per-domain table highlights fields such as law, science and history. Tourism is not among those analysed, so extrapolating the percentages to a hotel is an assumption, not a finding.

What we do with this

What follows from the study is fairly sober and requires no tooling: write content a model can cite without making anything up. In practice that means figures with their source linked, claims attributed to whoever made them, direct answers to real questions, and saying explicitly what is not known.

This article is written that way, and so are the others on this blog. We do it because it is the only recommendation in this whole field that arrives backed by a measurement.

To that you add the technical part, the part nobody sees: that the content is in the served HTML rather than depending on someone executing JavaScript, that the site declares structured data, that an llms.txt exists, and that robots.txt is not quietly blocking AI crawlers without anyone having noticed.

If you want to know whether your site can be cited today and what is keeping it out, write to us.

Sources

• Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, GEO: Generative Engine Optimization, arXiv:2311.09735, November 2023, latest version June 2024. Published in the proceedings of the 30th ACM SIGKDD conference, August 2024. The absolute scores come from their table 1 and the per-position figures from their table 2; the improvement percentages against the unoptimised baseline are our own calculation from those figures.