Generative engine optimisation is the practice of improving how often a brand appears in answers produced by AI assistants. Where answer engine optimisation asks whether one page can be quoted, GEO asks a market question: for a given prompt, who gets named, how often, and in what order.
How is GEO different from AEO?
They answer different questions about different things. AEO is a property of a page: can a model extract a clean answer from it. GEO is a property of a market: when someone asks about this category, whose name comes up.
You can do everything right on the page and still have no share of voice, because the model is drawing on sources you do not control. The reverse is also true — a brand can be named often because it is widely cited elsewhere, even with an unremarkable site.
Why does the same question give different answers?
Because generative models are not deterministic, and because they retrieve different sources on different runs. Asking once tells you almost nothing.
This is the practical difference from rank tracking. A search ranking is a position that can be read off. A generative answer is a sample from a distribution, so the only honest measurement is repeated sampling.
- Fix the prompt set, and change it deliberately rather than casually.
- Ask each prompt many times, not once.
- Report a proportion across runs, not a single observation.
- Re-measure on a schedule, because the market moves without you.
Frequently asked questions
- Is GEO just SEO for AI?
- Partly. Being well-ranked and widely cited helps in both. The difference is that GEO has no ranked list to occupy — the outcome is whether you are mentioned in prose, which is harder to influence directly and has to be sampled rather than read.
- How often should share of voice be measured?
- Often enough to see a trend and no more. Monthly suits most brands. Measuring daily produces noise, because the variation between runs is larger than most real week-to-week movement.
- Can share of voice be measured for free?
- You can ask an assistant the same questions manually and tally the answers. That works for a handful of prompts and stops being practical once you want a stable proportion across many runs and several competitors.