What is generative engine optimisation, and do you need a tool for it?

When a buyer asks an assistant what to buy, it produces a shortlist from what it has read about the category, plus whatever it retrieves at that moment. So the thing that decides whether you appear is the body of writing about you: comparison articles, community threads, retailer listings, documentation, reviews, and your own pages as one voice among them.

That is why the work often sits outside your own website, which is the part that surprises teams who arrive from search. You can have a technically excellent site and be absent from every answer, because nothing an assistant reads describes you.

Answer engine optimisation predates the current wave and covered being the source of a direct answer, including featured snippets and voice assistants. Generative engine optimisation is the term that stuck for the assistant era. Large language model optimisation and AI SEO are the same idea again.

It is worth knowing they are interchangeable mostly so that a vendor using a different acronym from yours does not read as a different category. If somebody insists on a sharp distinction between them, that is usually positioning rather than substance.

The unit changes. There is no position eleven in an answer. You are named or you are not, and if you are, either you were recommended or somebody else was while you were mentioned. Those are different outcomes needing different work, and a single visibility figure hides which one you have.

The target changes. A lot of the work is getting described correctly on pages you do not own, which is closer to communications than to technical optimisation, and it moves on a different clock. It is also more tractable than it sounds, because a few dozen sources usually account for most of what gets quoted in a category.

And the measurement changes. Ask an assistant the same question twice and you can get two different shortlists with nothing having changed. So a reading is a rate rather than a position, and any figure without a margin of error should be treated as one sample.

The first useful thing costs nothing. Ten questions a buyer would type, asked of two assistants, with the answers written down: which brands came back, in what order, whether you were among them, and whether anything said about you was wrong. That is an afternoon and it is genuinely diagnostic.

What that cannot do is tell you whether anything moved. Because the same question gives different answers, the variation between two hand checks is usually larger than a real improvement, so you can do a quarter of good work and read a fall. Knowing whether something worked needs repetition, a denominator and a margin of error, and that is the point at which measuring becomes a job rather than a task.

It is a reasonable order for most teams: check by hand, find out whether the problem is real, do the obvious work, and buy a measurement at the point where somebody starts asking whether it is working.

Check that you are not blocking the assistant crawlers in your robots file, which is a two minute job and is a surprisingly common own goal. Then ask the assistants about your category and read what is said about you specifically, watching for confident claims that are wrong, because those repeat and they cost more than absence.

After that, look at which sources keep coming up in the answers where a competitor is recommended. That list is where the work goes, and it is usually shorter and more approachable than a team expects.