Generative Engine Optimisation (GEO)

For twenty years the object of search was a ranking: be the first link and the click follows. That is quietly ending. A growing share of research now happens in a conversation with a model that reads the web on the person's behalf and returns a paragraph — with two or three sources named, and everything else invisible.

Position four on Google used to mean some traffic. Position four in a model's reading list means nothing at all. There is no page two to be on: either the answer is built from your words, or you were not in the room.

They overlap, and the foundations are shared — a fast, well-structured, crawlable site helps both. But the objective is not the same, and neither is the work.

A ranking is won by relevance and authority against a query. A citation is won by being the most usable source on a subject: specific, unambiguous, structured so a machine can lift a fact without misreading it, and consistent with what every other credible page says. Vagueness is fatal in a way it never was for search — a model cannot cite an adjective.

Structured facts

Schema.org across the estate — organisation, product, price, availability, review, FAQ — so a model reads facts rather than inferring them from prose. Stated once, stated identically everywhere, because contradiction is what gets a source dropped.

llms.txt

The emerging convention for telling a model what a site is and where its substance lives: a plain-text map at the root, and a full-text edition for the single fetch a model usually gets. We generate both from the content itself, so they cannot drift.

Content built to be quoted

Pages organised around the question a person actually asks, answered in the first paragraph, with the specifics — numbers, names, constraints — a model needs to reproduce the claim safely.

Entity consistency

The same description, category and detail across your site, your listings, your profiles and your press. Models resolve entities across sources; disagreement makes yours the one they leave out.

Crawl access

A deliberate decision, per crawler, about who may read the site — and robots and directives written so the ones you want are not blocked by an accident nobody has re-read since launch.

Measurement

Referral traffic from AI surfaces isolated in analytics, and periodic prompt testing across the major assistants to see whether you are named, what they say about you, and what they get wrong.

No agency controls what a model says, and anyone offering guaranteed placement in an AI answer is selling something that does not exist. There is no auction to enter and no ranking to buy.

What can be controlled is whether you are legible: whether the facts are structured, whether the claims are specific enough to be repeated, and whether the site agrees with itself. The properties that do well here are the ones that were already writing honestly and precisely. That is unusually good news — it means the work compounds instead of expiring with the next algorithm.

The goal is narrow and measurable: when someone asks an assistant for the best in your category, your brand is one of the two or three it names, and what it says about you is accurate. That is a smaller audience than a search results page and a far more decided one — the person asking has already delegated the shortlist.