Plain answers on how AI search works, what the terms mean, and what changes for a company that sells in Europe.
Most writing on this subject is either vendor marketing or American. Neither helps a European company decide what to do.
This is what we have written down: definitions that hold up, mechanics we have tested on client accounts, and the questions we are asked in every first meeting. If you are trying to work out whether any of this applies to you, start here rather than with a sales call.
The terms, defined
You will see four acronyms used interchangeably by people selling this. They are not the same thing.
AEO — Answer Engine Optimization Making a brand appear in the answers produced by AI systems: ChatGPT, Perplexity, Google’s AI Overviews, Copilot. The unit of success is a mention or a citation inside the answer, not a position on a results page.
GEO — Generative Engine Optimization Used by most practitioners as a synonym for AEO. Where a distinction is drawn, GEO refers specifically to generative systems that compose an answer from multiple sources, as opposed to systems that retrieve and rank.
LLMO — Large Language Model Optimization The narrower technical layer: how a brand is represented inside a model’s own parameters and retrieval index, rather than in what any single answer happens to say on a given day.
AI visibility The measurable outcome. Whether your brand is mentioned, whether your domain is cited, and how often, across a fixed set of prompts run under controlled conditions.
The distinctions matter less than the measurement. Anyone who cannot tell you how they measure the outcome is selling you the acronym.
→ What is AEO — the full explanation
How the systems actually work
Answer engines do not rank pages. They assemble an answer and, sometimes, cite where parts of it came from. That single difference undoes most of what a decade of SEO teaches.
Three consequences worth understanding before you spend anything:
- Being cited is not the same as being visible. A model can describe your category accurately, recommend a competitor, and cite you as the source of the definition. This happens more than you would expect.
- Answers are not stable. The same prompt produces different answers across sessions, accounts and dates. Any measurement taken from a single run is noise.
- Most conversations end without a click. Research from Bocconi University found that only 5.2% of ChatGPT sessions produce an external click, against 31.1% for Google searches. If the answer does not mention you, nothing downstream saves you.
→ How LLMO works — mechanics and measurement
Common questions
The questions we are asked in almost every first conversation, answered without a sales frame: whether AI visibility can be bought, whether schema markup helps, how long results take, whether existing Google rankings transfer, and what happens when a model gets your business wrong.
Written work
Analysis on AI search, answer engine behaviour, and what changes for European companies. Where we publish measurement data, we publish the method with it.
A note on how we measure
Everything on this site that looks like a benchmark comes from one protocol, and we would rather you know it than trust it.
We run cold prompts: logged out, no session history, no personalisation. The query set is frozen before measurement begins so it cannot be adjusted toward a flattering result. We distinguish between a mention, a citation, and a self-citation, because they are worth different amounts. Runs that return errors or refusals are voided rather than counted as absences.
If a supplier shows you AI visibility data without describing conditions like these, the data does not mean what they are telling you it means.
Want to know where you actually stand?
The audit runs your category’s real prompts and reports what the answer engines say about you today — mentions, citations, and who they recommend instead.