You cannot optimise for a model. You can optimise the evidence it reads.

AI search optimisation is not a ranking technique aimed at a model. It is the discipline of making a brand’s evidence legible to systems that assemble answers rather than list results. It splits cleanly into two, and conflating them is the most common error in the category.

AEO — answer engine optimisation — asks whether a specific claim on a specific page can be lifted cleanly and attributed. That is an extraction problem, and it is largely solvable on your own site. GEO — generative engine optimisation — asks whether the brand is corroborated well enough elsewhere that a generating system will repeat it. That is an authority problem, and most of the work is off your site entirely.

A page that treats them as one thing will sell you schema markup and call it authority, or sell you links and call it answer readiness. They move different numbers.

A ranked result list is a single ordering, produced once, shown to everyone who asked the same thing. An assembled answer is built per query from passages retrieved across several sources, then written. There is no position to occupy, and there is frequently no list to appear on.

What that changes in practice: the unit of competition moves from the page to the passage. A page can be the best document on a subject and contribute nothing, because no single stretch of it states a complete, attributable claim. Conversely a modest page with one clean, well-scoped statement gets used repeatedly.

This is the whole reason answer-first writing is a technical requirement rather than a style preference. The claim has to survive being removed from its page.

Before extraction is a question, reachability is. AI crawlers are a distinct set of user agents from Googlebot, they are blocked by a distinct set of defaults, and a robots.txt written in 2021 has no opinion about them — which is usually fine and occasionally fatal, when a CDN bot rule or a WAF silently returns 403 to everything unfamiliar.

On the one engagement this site can name, 15 of 15 AI crawlers were served full HTML. That was the good news, and it was worth establishing before anything downstream was discussed, because every finding about extraction is void if the document never arrives.

The second half of the same check is render parity: whether what a crawler receives matches what a browser assembles. A page whose copy is fetched by script after load is not a page that assembles reliably.

Extraction rewards a specific shape: a direct answer in one to three sentences, immediately under the question it answers, followed by the reasoning rather than preceded by it. Everything that delays the claim — a scene-setting paragraph, a benefits list, a "in this article we will" — costs extraction, because the passage retrieved may not reach the claim at all.

Structured data helps only where it describes what the page visibly says. On the same engagement, 0 of 84 relevant URLs carried FAQPage or HowTo markup — a real and cheap gap. The failure mode in the other direction is worse: markup describing content the page does not show is the strongest signal available, aimed the wrong way.

A generating system that cannot resolve who you are will not assert things about you. Entity work — one consistent name, one resolvable organisation node, profiles that actually exist — is the floor, and it is unglamorous enough that it is routinely skipped in favour of content.

Above that floor sits corroboration: whether independent sources say the same thing about the brand. This is where link counts mislead most completely. On the Ambar engagement, 396 referring domains were found and 2 were not spam. A tool reported an authority profile; a manual review of all 396 found there was effectively none.

That figure is the honest shape of the problem. Corroboration is built slowly and measured by reading, not by a domain count in an export.

Nobody can guarantee a citation in ChatGPT, an appearance in an AI Overview, or a position in a Perplexity answer. The systems are non-deterministic, unversioned from the outside, and reweighted without notice. An agency that promises a named citation is either misunderstanding the mechanism or counting on you not to check.

What can be done is bounded and real: establish that the crawlers are served, that the claims are extractable, that the entity resolves, and that the corroboration is genuine rather than counted — then measure whether extraction improves. That is a system with an input you control and an output you can read.