Generative engine optimisation is the practice of getting your content cited inside AI generated answers. This guide covers the mechanics in more depth than our complete guide to AI search optimisation, which sets out the wider picture.
What GEO is, precisely
Watch: the short version
If you would rather have the argument in a minute than in three thousand words, start here.
GEO is optimising so that a language model retrieves your page, uses it while composing an answer, and names you as a source. Three distinct things, and you can fail at any one of them independently.
Retrieval fails when the crawler cannot read you. Usage fails when your prose does not survive being chopped into passages. Citation fails when the model synthesises your information without attributing it, which happens more often than anyone admits.
The fan out problem in detail
A model does not search the question it was asked. It decomposes that question into shorter sub queries and searches each separately. Everything else about GEO follows from this.
Work through it practically. Take the most commercially valuable question a customer could ask an assistant about your category. Write it out in full. Now write the three to five short searches a machine would plausibly derive from it, using the plain terminology it would reach for rather than your internal vocabulary.
Check whether you have a page that would be retrieved for each fragment. Most businesses discover they have one page targeting the whole elaborate question and nothing addressing the fragments. That gap is the entire content plan.
Writing for passage retrieval
Retrieval operates on chunks rather than documents, which changes how a page should be written.
Put the answer in the first forty to sixty words under each heading, then elaborate. Keep paragraphs to two or three sentences. Name the subject instead of writing “it” or “this approach”, because a pronoun loses its referent the moment a passage is lifted out of context.
The test is simple. Take any paragraph on your page, paste it somewhere with no surrounding text, and read it. If it no longer makes sense alone, a model cannot use it cleanly either.
Measuring GEO
There is no rank to track because language models are non deterministic. The metric is share of voice, meaning how often you appear across a fixed basket of prompts run repeatedly.
Build twenty prompts covering the questions that precede a purchase in your category. Run them across ChatGPT, Perplexity and Gemini once a month. Record whether you appear, how you are described, and who gets cited instead. That manual baseline costs nothing and tells you whether a paid tool is measuring anything you cannot see yourself.
One observation proves nothing. Somebody checking once and reporting that you are invisible has taken a single sample from a probabilistic system.
