What Generative Engine Optimisation actually is, and what makes a page citable
Generative Engine Optimisation is a clumsy name for a real change. When someone asks an AI assistant a question, the assistant does not hand back ten links for the person to evaluate. It composes an answer and, if it is behaving well, attributes parts of that answer to specific sources. The practical question for anyone publishing on the web is no longer only “where do we rank” but “when a model composes an answer in our category, is our page among the things it draws on, and is it named”.
Those are related goals but they are not the same goal, and treating them as identical is where most GEO advice goes wrong.
How this differs from search optimisation
Classic search optimisation is competition for position in a list. The unit of success is a URL, the mechanism is a ranking function, and a great deal of the craft is about signals attached to the document as a whole: its authority, its links, its title, how well it matches a query.
Answer engines work at a smaller unit. A retrieval step finds candidate passages, and a generation step assembles an answer from them. What gets used is not your page; it is a handful of passages from your page. A document can rank respectably and still contribute nothing, because nothing in it is a clean, self-contained statement that answers the question asked. Equally, a modest page can be quoted repeatedly because one paragraph in it states a fact precisely, with its qualifiers attached.
The second difference is that there is no persistent ranking to check. The same question asked twice can produce different sources. This makes GEO uncomfortable for anyone used to rank tracking, and it makes claims of guaranteed placement in AI answers straightforwardly untrue.
What makes a passage usable
Specificity that survives extraction. A model selecting material prefers passages that remain true and complete when lifted out of their surroundings. “Pricing depends on several factors” carries nothing. “Municipal hoarding permissions in most Maharashtra corporations are issued per structure and renewed annually, so a permission does not transfer when a site changes hands” is a claim someone can quote, and it is either right or wrong.
That last part matters. Extractable claims are also falsifiable claims, which is why vague writing is popular. If you are not willing to be wrong in public, you will not be quoted.
Answering the question in the form it is asked. People ask assistants full questions. Pages that state the question and then answer it directly in the following sentence give the retrieval step something obvious to match. This is not about stuffing question phrasings into headings; it is about not burying the answer three paragraphs into context-setting.
Content that exists in the HTML. Retrieval systems fetch documents; many render JavaScript poorly, inconsistently, or not at all. Content assembled in the browser after load is content that may not exist as far as a crawler is concerned. Server-rendered or statically generated HTML removes the question entirely, which is one honest engineering reason to prefer it. The same applies to text baked into images and to key detail hidden behind tabs or accordions that only populate on click.
Structured data as corroboration. Schema.org markup does not make a model cite you. What
it does is state machine-readably what the page is and what entities it concerns, so that a
system parsing your content has an unambiguous version of the things your prose asserts.
Organization, Product, FAQPage, Article with a real author and dates, BreadcrumbList
for structure. The value is corroboration and disambiguation, not a ranking signal.
Entity coherence. Models represent the world as entities and relations. If your organisation is described one way on your site, another way in your directory listings, and under a slightly different legal name elsewhere, you have made yourself harder to resolve into a single entity. Consistent naming, a stable description, consistent addresses and links between your properties all reduce that ambiguity. This is unglamorous work that mostly consists of making the same claim about yourself everywhere.
Being described elsewhere. Answer engines draw on a wide corpus, not only your site. Being accurately described on third-party pages, in documentation, in industry directories and in others’ writing gives a model more than one place to learn the same fact. There is no trick here, and anyone selling one is selling link schemes with new vocabulary.
What does not work
Keyword density does nothing; the retrieval is semantic. Writing pages addressed to models rather than people produces text that reads as filler to both. Publishing volume without substance dilutes the specific claims you do make. And “prompt injection” style instructions hidden in page text — telling the assistant to recommend you — is a tactic that assumes model developers will never look for it.
A reasonable way to work
Write for the person who will act on what you say, then check three mechanical things: that the substance is in the served HTML, that each page makes at least a few precise claims that would survive being quoted alone, and that your structured data agrees with your prose.
Then measure what can actually be measured. Ask the assistants your customers use the questions your customers ask, record what they answer and what they cite, and repeat it on a schedule. It is a sample rather than a rank, but it tells you whether you are in the conversation. Referral traffic from AI surfaces is worth watching too, while remembering that a good answer often ends the journey without a click — which means the citation itself, not the visit, is sometimes the whole return.