The query that returned ten links now returns a paragraph. Three sources are cited beneath it, sometimes four. The visitor reads the answer and does not always click — but when they do, they already know who you are, because the machine told them.
That shift is what GEO names: generative engine optimisation. The term has been circulating since a 2024 research paper and is already being misused — usually as one more channel to buy, when it is neither sold nor bought.
What changes is not the channel, it is the output
A conventional search engine orders documents. Its promise is implicit: here are ten pages, sort it out. The work of selection stays with the visitor, and position in the list is what gets contested.
A generative engine does that work instead. It reads, synthesises, decides, and says where what it asserts comes from. The question is no longer “how do I climb” but “on what terms am I quoted”.
The difference is not cosmetic. In a list, ten results coexist. In a written answer, three sources are enough and the fourth does not exist. Concentration is higher, and the gap between cited and uncited is harsher.
What conventional search still governs
Nothing that made a site sound has become useless. A generative engine feeds on the same indexes, under the same constraints:
- a reachable page — served without blocking, without a consent wall hiding the text, without rendering that depends on a script
- content readable without execution: what the crawler receives in the HTML, not what the browser eventually displays
- structured data naming the entity, the author, the date — markup the machine does not have to guess
- plausible domain authority: a site cited elsewhere is a site the machine judges quotable
A site Google cannot crawl properly does not become visible because someone added an FAQ. The order of works has not changed.
GEO does not rescue an invisible site. It determines what you do with a site that is already legible.
What decides a citation
Three qualities recur, and none of them is technical. They come down to how the text is written.
- The clear statement. A machine repeats a sentence that stands alone. A paragraph that builds for six lines before concluding offers nothing to extract.
- The attributed figure. A number with its source gets cited; the same number without provenance gets avoided. It is also the only protection against being quoted wrongly.
- The date. Undated regulatory content is set aside out of caution — the machine cannot tell whether it still holds.
What the three have in common: they also make better texts for human readers. That is why GEO is hard to tell apart, at first glance, from ordinary editorial rigour.
Why now rather than in two years
The argument is not that traffic from AI engines is already substantial — it is not. It is that a machine reputation is being formed right now: what engines retain about a company, the sources they associate with it, the claims they attribute to it.
That reputation is hard to correct after the fact. A company whose pages never state clearly what it does, for whom, and what sets it apart will be described by what others say of it — directories, aggregators, competitors. Writing the answer costs less than denying it.
What it assumes you have already settled
A machine cannot cite a company on a promise the company has never made. It will repeat the wording available, and if the only wording available is vague, so will the citation be.
This is where GEO meets a question that is not technical at all: knowing what you claim, and writing it in a sentence that stands alone. Companies that go wrong here almost never stumble on the markup. They stumble on the fact that no one inside the company ever settled the sentence.