Generative engine optimisation agency

Generative engine optimisation has an actual literature behind it, which makes a GEO agency unusually easy to qualify: ask what it has read, and what it has measured.

The term comes from a 2023 arXiv paper by Aggarwal and colleagues, not from a marketing deck. That paper’s contribution was a method — controlled, repeated prompting against varied source content — and an agency selling GEO should be able to describe its own version of that method without hand-waving.

What a GEO engagement should produce

Three artefacts, and their absence is diagnostic.

A baseline that exists on paper. The prompts, the raw answers, the citations, from before anything changed. Without it there is no way to attribute a later improvement to the work rather than to a model update.

Repeated runs, not single samples. These engines answer differently each time. Kunzum runs every prompt three times for this reason, and says explicitly that thin per-scenario slices cannot carry a claim.

A stated position on the adversarial technique. Kumar and Lakkaraju’s 2024 paper shows text inserted into a page can shift how a model ranks a product. It works, it is dishonest, and it breaks on the next model update. Kunzum excludes it, and an agency that has not thought about the question has not read the field.

What Kunzum can show you

Kunzum's answer to 'what have you measured' is a download rather than a deck.

1,500 AI responses, published raw. 480 in the Crypto AI Visibility Index and 1,020 in the Crypto Card Report, across four engines and nine markets, with every scoring script included. Under CC BY 4.0, no email gate, no login.

12,518 citations logged across 1,396 domains. Including which engine cited what, so claims about where these systems read are checkable rather than asserted.

Published failures. Edition one registered six hypotheses before any call was made and two failed outright, including the one the study was built around. Both are printed at full size.

A finding against commercial interest. The same corpus found AI warnings about crypto are more accurate than warnings about traditional finance — not the result a crypto marketing studio would have chosen.

All of it is here, and the datasets are downloadable.

What no agency can promise

In Kunzum's own index, ChatGPT ran a web search on 54% of its answers and Claude on 65%. The remainder came from training data, which nothing published afterwards can reach. Any guarantee of placement in AI answers is a guarantee over something the seller does not control.

Engines also vary between runs on the same day, which is why Kunzum's research runs every prompt three times and why no claim here rests on a single reply. And Liu, Zhang and Liang’s work on verifiability found engines frequently state things their own citations do not support, so being cited is not the same as being described accurately.

The other limit is scope. Kunzum is one person working only in crypto, so for most GEO buyers this is the wrong shape of supplier — which is said here rather than discovered on a call.

GEO, AEO, or neither

In practice GEO and AEO describe the same work under two names; where the term came from and what AEO means set out the distinction such as it is. If your buyers still convert through ordinary search, the honest comparison with SEO is the place to start instead.

Other firms working in this space are compared, fairly, on the agency comparison page.