What Is Generative Engine Optimisation?

The short answer

Generative engine optimisation, or GEO, is the work of making your site one of the handful of sources an AI answer engine draws on when it writes a reply, and of being named when it does.

The engines in scope are ChatGPT, Google AI Overviews and AI Mode, Perplexity, Gemini and Copilot. The outcome you are buying is a citation and a mention, not a rank.

Why does it need a separate name?

Because the thing being competed for is different. A results page shows ten links and the reader chooses. A generative answer shows a paragraph and names three or four sources. Everything outside that paragraph is invisible to the person reading it, however well it ranks.

That changes the unit of optimisation. In classic SEO you optimise a page. In GEO you optimise a passage, because retrieval pulls a chunk of text rather than a whole document, and the quality of that chunk decides whether your wording survives into the answer.

How a generative answer is actually assembled

Simplified, but accurate enough to work from:

1. The engine interprets the question and decides whether it needs to search.
2. It runs one or more searches and gets back candidate pages.
3. It fetches a small number of them.
4. It extracts passages that look relevant.
5. It composes an answer from those passages, and may cite some of the sources.

Each step is a place to lose. You can fail at step three because a crawler cannot fetch your page, at step four because your answer is buried under an introduction, or at step five because another source stated the same thing more cleanly.

The two levers that matter

Almost everything useful in GEO falls into one of two categories.

Retrieval readiness is engineering: server-rendered content, clean headings, self-contained passages, schema that states what the page is, and crawler access. It is unglamorous and it is usually where the quick gains are.

Corroboration is the slow one. Models weigh what other sites say about you more heavily than what you say about yourself, which is why a directory listing frequently beats a better-written page. Closing that gap takes real coverage over time.

What about GEO vs SEO vs AEO?

The terms overlap and the industry uses them loosely. Our working definitions:

There is a longer comparison in AEO vs GEO vs SEO, and the service side is on our GEO agency page.

Is GEO the same as LLM SEO?

In practice, yes. GEO names the surface, a generative engine. LLM SEO names the system underneath, a large language model. Agencies use them interchangeably and the work is identical: be retrievable, be quotable, be corroborated.

If you want the model-side view of the same problem, see what is LLM SEO.

What GEO work looks like in practice

A realistic first engagement is mostly correction rather than creation.

• Rewrite the pages that should be answering so the answer leads each section.
• Fix the facts about your business wherever they disagree across the web.
• Make sure server-rendered content exists for crawlers that do not run JavaScript.
• Write schema that states plainly what the organisation and the page are.
• Add the pages that answer questions nobody has covered yet.
• Set up a repeated prompt set so the before and after are comparable.

Note what is absent: publishing volume, keyword density and link buying. None of those address the two levers.

What makes a passage quotable?

Since retrieval works on chunks rather than whole pages, it is worth being precise about what a good chunk looks like.

• Self-contained. It makes sense without the paragraph before it. No references to something mentioned above, no pronouns pointing off-screen.
• Specific. Contains the nouns and numbers a model needs to answer. Vague assurance is unquotable.
• Attributable. Says who is making the claim where that matters, so the model can cite it correctly.
• Short enough to lift. Two to four sentences is the sweet spot; a six-hundred-word block forces the model to paraphrase and your wording is lost.
• Under a heading that matches the question. Headings are a strong retrieval signal and a cheap one to get right.

If you only change one thing, break long sections into shorter ones with honest headings and make each one answer its heading immediately.

Why corroboration decides contested categories

Here is the uncomfortable part of GEO: in a competitive category, the better-written page often loses to the worse-written one that more sources agree with.

Models treat a claim repeated across independent sources as more reliable than the same claim made once, on the site of the company that benefits from it. That is reasonable behaviour, and it is why directory listings and trade coverage routinely outrank a company's own page in generated answers.

Closing that gap is slow work: accurate listings, genuine coverage, consistent facts everywhere, and time. There is no technical shortcut, and anyone selling one is selling links.

The practical consequence for planning: budget the retrieval fixes as a short project and the corroboration as an ongoing programme. Expecting the second to behave like the first is how GEO engagements end in disappointment.

Freshness, variance and other things nobody mentions

Three realities that shape what is achievable.

That last row deserves attention when setting expectations. A generated answer can name you without the reader ever visiting your site. The value is real but it does not land in your analytics as a session, which makes it easy to under-count and easy for an agency to over-claim.

Where does GEO actually fit in a marketing budget?

This is the question that decides whether a GEO programme survives its first quarter, and it is usually answered badly.

GEO does not replace a working channel. If paid search is producing enquiries at a cost you are happy with, moving that budget into generative visibility on the strength of a trend piece is a poor trade. What GEO competes with sensibly is the slow half of your SEO budget: the content and link spend aimed at terms you will not realistically win for two years.

The argument for shifting some of it is that the entry cost is lower. Being cited depends far less on domain authority than ranking does, and far more on whether your facts are consistent and your pages are fetchable. Those are engineering problems, and engineering problems have a floor.

A reasonable first allocation, for a business with no AI visibility work at all:

Note that none of it is new content in the first pass. That is deliberate. Publishing before the entity picture is clean adds noise to a signal a model is already struggling to resolve.

What a GEO report should contain

Since the field is new, reporting conventions are not settled, and that gap is being filled with screenshots. A report worth paying for contains five things.

1. The prompt set, written down, so you can see it has not changed between months.
2. Run counts, because a single run is anecdote. Several runs per prompt, per engine.
3. Named or not, per prompt, per engine, as a rate rather than a yes or no.
4. What was said, including anything inaccurate, because confidently wrong claims cost more than absence.
5. Which URLs were cited, yours and your competitors', because that is where the next piece of work comes from.

If a report shows you a rise without showing the prompt set, the rise is unverifiable. The prompts can be changed between months and the number will move on its own.

How do you measure GEO?

With a fixed prompt set, run on a schedule across the engines, recording whether you are named, what is claimed and which URL is cited.

The variance is real: the same question can give different answers minutes apart. That is why a single screenshot is worthless as evidence in either direction, and why honest reporting is a trend across repeated runs rather than a snapshot.

If you want a baseline before deciding anything, our free AI visibility audit records exactly this and sends you the findings in writing.