A buyer researching a purchase increasingly never sees a results page. They ask an assistant, get a synthesised answer with three or four sources cited inline, and either act on it or ask a follow-up. The ranked list of ten blue links — the thing two decades of SEO was built around — is not in that loop.
This does not make search visibility worthless. It moves where the value sits: from being ranked to being quoted.
What is generative engine optimisation?
Generative engine optimisation is the practice of structuring content so that AI answer engines cite or quote it when generating answers.
The term comes from academic work rather than the marketing industry. A 2024 paper presented at ACM SIGKDD introduced it as a formal problem, along with a framework for measuring visibility inside generated responses — which is harder than it sounds, since a citation can be prominent or buried, and counting citations alone misses that.1
How is GEO different from SEO?
The technical foundations are nearly identical. What changes is the content structure that wins, because the consuming system is different.
| Classic SEO | GEO | |
|---|---|---|
| Goal | Rank a link | Be quoted in an answer |
| Success looks like | A click | A citation, click optional |
| Unit of competition | A page per query | A passage per question |
| Rewards | Relevance, links, page experience | Extractability, specificity, attribution |
| JavaScript tolerance | Generally rendered | Frequently not executed |
| Author identity | Indirect quality signal | Direct input to source selection |
The most important row is the last two. Classic search crawlers render JavaScript; many answer-engine crawlers do not. And where classic ranking treats author expertise as one quality signal among many, a system deciding what to quote has a sharper reason to care who said it.
What actually makes content citable?
The mechanism is extraction. A model answering a question is looking for a passage that answers it — completely, briefly, and without needing the surrounding paragraphs to make sense. Everything below follows from that.
Put the answer first, and make it self-contained
Open with a direct answer to the question in the title, in roughly 40 to 60 words, written so it can be lifted whole. The test is simple: read the passage with no context and ask whether it still makes sense. If it begins "As we saw above" or "This approach," it is not liftable.
This is the single highest-leverage change, and most content fails it because the conventional article structure warms up for three paragraphs first.
Write headings as the questions people actually ask
"What is a good CAC payback period?" is a better H2 than "Payback benchmarks." Conversational queries are phrased as questions, and heading text that mirrors the question makes the match unambiguous.
Define terms in canonical sentences
One sentence, of the form "X is …", that stands alone. A glossary entry or a definition callout does this explicitly, which is why both formats get quoted disproportionately.
Quantify and attribute
The SIGKDD work found that adding relevant statistics and adding quotations from credible sources were among the most effective changes for improving visibility in generated answers — with the strongest methods reported to improve their visibility metrics by roughly 30–40% over an unoptimised baseline.1
That finding matches the intuition: a model choosing between a page that says "cart abandonment is high" and one that says "Baymard Institute puts documented average cart abandonment at roughly 70%" has an obvious reason to prefer the second. It is more useful to the person asking, and it carries its own attribution.
The obligation this creates is real. Statistics only work if they are true and sourced. Inventing a plausible number to look citable produces content that is confidently wrong, and being cited for a false claim is worse than not being cited.
Use tables for anything comparative
A comparison expressed as prose has to be parsed. The same comparison as a table is already structured. Tables are among the most readily extracted blocks on a page.
Give the content an identified author
Author identity is a citation input, not decoration. That means a real name, a real role, credentials that bear on the subject, and — where they exist — links to profiles that let a system resolve the person as an entity rather than a string.
Serve real HTML
If the content only exists after hydration, a crawler that does not execute JavaScript sees an empty shell. Static rendering or server-side rendering is not an optimisation here; it is the precondition.
What about structured data?
It removes ambiguity, which is worth more to a system deciding what to cite than to one deciding what to rank.
The types that carry weight for this purpose are unglamorous: Article with a
real author and dateModified, FAQPage for question blocks, HowTo for
procedures, BreadcrumbList for context, and a consistent Organization with
sameAs links to the profiles that identify you elsewhere.
That last one is quietly important. Entity consistency — the same name, the same description, the same profile links everywhere you appear — is what lets a system resolve scattered mentions into one recognisable entity. Inconsistent naming produces several vague entities instead of one solid one.
What about llms.txt?
It costs almost nothing, so it is worth doing, and it is not a ranking lever.
llms.txt is a proposed convention: a plain-text file at the root of a site
describing what the organisation is and mapping its important content in
Markdown. Adoption by the major answer engines is not universal or guaranteed,
and it should be treated as a low-cost bet rather than a technique with
demonstrated returns.
Anyone selling llms.txt as the core of a GEO programme is selling the
cheapest item on the list as though it were the most valuable.
How do you measure GEO?
Imperfectly, and that is worth saying out loud rather than papering over.
The honest position today is that there is no equivalent of rank tracking with comparable reliability. Answers are non-deterministic, personalised, and vary between engines and over time. What can be done:
- Prompt panels. Maintain a fixed set of questions a buyer would actually ask, run them across the major engines on a schedule, and record whether you appear and in what role. Directional, not precise — but a trend over months is readable.
- Referral traffic from assistants. A small and growing share of sessions arrive with AI assistant referrers. It undercounts influence badly, since the citation with no click is the common case, but it is real data.
- Branded search volume. When answer engines mention you, some proportion of people search your name afterwards. A rise with no other explanation is weak evidence that something is working.
None of these is a conversion-grade metric, and treating them as one would repeat the mistake this whole discipline is supposed to avoid.
Where to start
In order of return on effort:
- Rewrite the openings of your best existing pages so the first passage answers the title question and can stand alone. Cheapest change, largest effect.
- Check that your content exists without JavaScript. View source, not inspect element. If the body copy is missing, nothing else on this list matters.
- Add real authorship to anything substantive, with credentials that bear on the topic.
- Source every statistic with a link, and remove any you cannot source.
- Convert comparative prose into tables.
- Add structured data for articles, FAQs and your organisation, with consistent entity information.
The first two are where most sites lose before they start.