What is generative engine optimisation (GEO)?

Generative engine optimisation, usually shortened to GEO, is the work of getting a business named and quoted inside answers written by generative systems: AI Overviews and AI Mode in Google, and assistants such as ChatGPT, Gemini, Perplexity and Copilot when they search before replying. The category is young, the proposals are already everywhere, and a fair number of them describe ordinary search work with a new cover sheet.

You will also see it written generative engine optimization. That is the American spelling of the same thing.

This article is written for the person deciding whether to pay for it. It sets out what the work contains, what moves the price, what cannot be promised, and what to ask on the call.

The name is new, the mechanism is not entirely

A generative answer is grounded. The system runs its own searches, retrieves a handful of documents, and writes from them. That means two things a buyer should hold onto. Retrieval is still an index problem, so the technical foundations are the same ones search has always needed. Generation is a writing problem, so clarity, consistency and structure decide who gets quoted out of the documents that were retrieved.

In our experience roughly the first half of any GEO engagement is technical search work. That is not a criticism. It is the reason a programme sold as entirely new work should be read carefully, because you may already be paying for that half somewhere else.

The systems are not one thing

GEO is not a single target, and a supplier should say which system they are working on. Google's AI Overviews and AI Mode are grounded in Google's own index, which is why eligibility runs back through Google Search Essentials and why the company's documentation on AI features in Search, at developers.google.com/search/docs/appearance/ai-features, is worth reading directly rather than in somebody's summary. An assistant such as ChatGPT or Perplexity generally issues its own queries through a search provider, retrieves a handful of pages, and writes from those. A model answering from memory with no retrieval step at all is a third case, and nothing you publish this quarter reaches it.

That distinction decides what can be promised. Work aimed at grounded retrieval can be measured within a quarter, because the system goes and fetches material each time it is asked. Work aimed at what a model has already absorbed cannot be scheduled by anyone, and a proposal that blurs the two is describing the measurable half and charging for both.

What a real programme contains

There are four components. A proposal missing one of them is usually missing it because it is the difficult one.

1. A measured baseline: the questions your buyers ask, written down, asked of named assistants on a stated schedule, with the answers recorded
2. Entity reconciliation: one agreed set of facts about the business, and a corrected version of every place that currently contradicts it
3. Structure: organisation, service, article and FAQ markup from schema.org, implemented on templates and validated after every change
4. Answer led content: pages that answer one buyer question properly, written by somebody who knows the subject

Everything else in a GEO proposal is either one of those four wearing a different name, or it is standard technical SEO, which is fine as long as the invoice says so.

The baseline is the part suppliers skip

Without a baseline, nothing that follows can be judged. With one, the conversation changes completely, because the first report usually tells you something more useful than a ranking ever did: how these systems currently describe your business to a stranger.

A baseline is worth having on your own terms. The prompts should be written down verbatim so they can be repeated, the assistants named, the date recorded, and the whole answer kept rather than a favourable fragment. Answers vary between runs for the same prompt, which is exactly why one screenshot proves nothing and a repeated observation proves something.

Entity reconciliation is dull and it carries the weight

A generative system assembling an answer about you reads your site, your profiles, your listings, your old press coverage and whatever an aggregator decided about you some years ago. Where those disagree, the cautious move is to quote a business whose facts agree with themselves. This is the least impressive slide in any proposal and the one we would fight to keep.

• One legal name, punctuated identically everywhere it appears
• One address and one phone number, with the contact page checked first because it is the page that most often disagrees with the rest of the site
• One sentence describing what the business does, short enough to be quoted whole
• One list of service names, matching between the site, the markup and every profile
• Retired services, old addresses and closed offices removed rather than left to be found

Content that deserves retrieving, not content at volume

The failure mode of this work is obvious once you have seen it. A supplier generates a page for every phrasing of every question, publishes them at speed, and calls the result answer coverage. Google's spam policies name that directly as scaled content abuse, and the demotion arrives without warning and at the worst possible time.

The version that works is slower. One page per genuine question, opened with an actual answer in the first two sentences, written with a subject matter expert who can say the thing a competitor's marketing department cannot. The test is whether two consecutive sentences pulled from the middle still mean something and are still true on their own. If they need the rest of the page, they will not be quoted.

Structured data, and what it does not do

Schema.org markup states facts instead of implying them, which removes the interpretation step. It is worth doing carefully, on templates rather than page by page, and validating after every release because structured data fails quietly and nobody files a bug about it.

It is not a ranking lever and it does not cause citations. Marking up content that is not visible on the page is against Google's structured data guidelines and is a good way to lose rich results entirely. Any proposal in which schema is the headline deliverable is thin.

Mentions without links still count for something

Search has spent two decades teaching marketers that a mention is only worth having if it carries a link. Generative retrieval is less attached to that idea, because what a system needs is corroboration: the same claims about you appearing in places it already trusts. Trade press, category directories, professional bodies, community answers, comparison sites and documentation all qualify, and the ones that matter are category specific.

The practical work is unromantic. Find the sources that keep appearing in the answers your baseline recorded, get your facts on them right, and give them something worth referring to. That is digital PR with a different acceptance criterion, not a new discipline.

How this is priced, and what moves it

We price this as a rolling monthly engagement rather than a project, because the systems change on their own schedule and a one off report ages within a quarter. What moves the figure is not a keyword count, and any proposal priced by keywords is describing a spreadsheet rather than a programme.

• How many buyer questions are tracked, and across how many assistants
• How badly the current facts, profiles and listings disagree with each other
• How much answer led content is written each month, and whether your experts are available to be interviewed
• Whether structured data has to be built from scratch or corrected
• Whether the site's technical foundations need work first, which is search work and should be invoiced as search work

Our own view is that this is usually bought alongside an SEO programme rather than instead of one. The foundations are shared, and paying two suppliers to fix the same crawl problem helps nobody but the suppliers.

Questions that expose a thin proposal

Ask these on the call, and listen for whether the answer contains a mechanism or a mood.

1. What is your baseline method, and will I get the prompts in writing so I can run them myself?
2. Which parts of this are technical SEO, and am I already paying for them elsewhere?
3. How do you handle the fact that answers differ between runs of the same prompt?
4. How much content will be published each month, who writes it, and who is the subject matter expert?
5. What will you not do, and what in Google's spam policies are you steering around?
6. What does month one hand over, as a document I can keep?

The answer to the last one is a fair test on its own. A supplier who cannot describe a first month deliverable in a sentence has not decided what the work is.

What nobody can sell you

Some of what is being promised in this category is not available to anyone, at any price.

• A guaranteed mention in any assistant, since these are not ranked lists and they change without notice
• Paid placement inside a generated answer, which is advertising if it exists at all and should be labelled as such
• A citation share figure, because there is no published denominator and any percentage quoted at you was produced by a tool making assumptions it does not show you
• A fix for having nothing distinctive to say, which is a positioning problem and not a markup problem

If a supplier guarantees a citation, ask them to put the remedy in the contract. What happens, specifically, in the month the citation does not appear? The answer to that question ends the conversation more reliably than any argument about method.

A reasonable expectation to hold

Here is what we think a well run programme actually produces. Within a few weeks you know how the major assistants describe your business, which is often the first time anyone has checked. Within a quarter the contradictions are gone, the markup is clean and validating, and the pages people arrive on answer the question they arrived with. Over two or three quarters the recorded answers start naming you more often, unevenly, with setbacks when a provider changes something without telling anyone.

That is a slower promise than the ones in circulation. It is the one we are willing to put in writing.

A sceptical buyer should be less worried about being left behind by AI search than about paying two suppliers to fix the same crawl error. Tech Conversion

Questions people ask about this

What is generative engine optimisation?

Generative engine optimisation is the work of being cited inside an answer written by a generative system, rather than ranked in a list of links beneath it. In practice it combines technical search foundations, one consistent set of facts about the business everywhere it appears, structured data, content that answers buyer questions directly, and repeated measurement of what the assistants actually say.

Is GEO worth paying for separately from SEO?

Often not, and we would rather say so. The technical foundations are shared, so buying both from different suppliers usually means paying twice for the same crawl and indexing work. What genuinely sits outside an SEO retainer is the baseline measurement, the entity reconciliation across profiles and listings, and the answer led writing. Those are worth scoping explicitly. Ideally one plan covers GEO, SEO and the content both depend on, so nothing is bought twice.

How long before it shows results?

A baseline takes a couple of weeks, because it is measurement rather than change. Shifts in how assistants describe a business usually appear over one or two quarters, unevenly, since these systems update on their own schedule. Anyone promising movement within weeks is either describing something they cannot control, or counting a single favourable answer as a result.

What should a GEO report contain?

The prompts used, written out in full so you can repeat them yourself. The assistants queried and the dates. The complete answers rather than favourable fragments. Whether you were named, in what company, and how you were described. Which of your pages was cited. Then a short written commentary on what changed since the previous run and what the next month does differently.