what is GEO, exactly?

when someone asks an assistant "who does penetration testing for small healthcare companies?", the reply is a short paragraph naming three or four firms. there is no page two. if you aren't named, you were never in the running, and nothing in your analytics will show you the loss. GEO is the work of getting into that paragraph.

how the answer gets assembled

three sources feed any answer. training data, which you can't edit and which lags reality by months. live retrieval, the search the assistant runs while answering, which is the part you can influence this quarter. and the model's own summarisation, which strongly favours text that already reads like a clear, factual answer. being retrievable and being quotable are two separate jobs, and most sites fail the second even when they pass the first.

what moves it

1. being crawlable by the AI crawlers specifically. they aren't googlebot; a default deny in a bot rule blocks them silently, and content that only exists after javascript runs often isn't read.
2. answering the question in the first two sentences, so the passage can be lifted as it stands.
3. publishing specifics you're willing to commit to: numbers, ranges, timelines, named trade-offs.
4. structure a machine can parse: proper headings, real tables, FAQ and service schema.
5. being cited somewhere that isn't your own site. three independent mentions become the consensus the model reproduces.

how we measure it

there's no ranking to track, so we track prompt coverage: the twenty or thirty questions a buyer would plausibly ask, run across the major assistants on a schedule, recording whether you're named, in what position, and which sources the answer cites. three named competitors get the same treatment, so you can see the gap move. prompt stuffing and mass-generated pages don't help, and neither does chasing every new assistant. we don't do any of it. the full working definition goes deeper, and this question explains where GEO sits next to paid.