What is Generative Engine Optimisation (GEO)?

Generative engine optimisation (GEO) is the practice of improving how often, and how accurately, your brand appears in answers generated by AI systems such as ChatGPT, Google AI Overviews and Perplexity. It targets the answer, not the links below it.

That is the short version. The longer version matters, because most of what gets written about GEO either recycles SEO advice under a new name or treats generative engines as a black box you simply have to hope well of. Neither is much use if you are trying to decide whether this deserves budget.

Where the term comes from

GEO is not an agency invention. It was formalised in a 2023 research paper, GEO: Generative Engine Optimization, by Pranjal Aggarwal and colleagues at IIT Delhi, Princeton, Georgia Tech and the Allen Institute for AI, later published at KDD 2024. The paper coined “generative engines” as a category, built a benchmark of user queries, and tested which content changes moved the needle. Its headline finding was that targeted optimisation could improve visibility in generative responses by up to 40%.

Two things are worth taking from that. First, the term has an academic origin rather than a marketing one, which is unusual. Second, the paper’s premise was that content creators had lost control of how their work surfaced, and needed a way to get some of it back. That is still the honest framing of the discipline.

What a generative engine actually does

A traditional search engine matches a query to documents and ranks them. You compete for a position in a list.

A generative engine does something different. It reads your question, works out what information it needs, runs one or more searches in the background, pulls a small set of documents, and then writes an answer using them. You often never see the list. What you see is a paragraph, with a handful of source links attached.

Three consequences follow, and they shape everything GEO does:

• The result set is tiny. A ranking page shows ten results. An AI answer typically synthesises from a handful of sources. Being on page one is no longer close enough.
• One question becomes several searches. Engines commonly decompose a query into sub-questions and search each one. So the phrasing you optimise for and the phrasing the engine actually searches are rarely the same.
• The answer is written, not retrieved. The engine paraphrases. It can describe your product accurately, vaguely, or wrongly, and it can recommend a competitor in the same sentence that mentions you.

GEO works on retrieval, not training

This is the point most GEO explainers skip, and it is the one that determines what is actually possible.

A model can answer from two places. It can answer from what it absorbed during training, which is fixed, months out of date, and impossible for you to edit. Or it can answer from documents it fetches at the moment you ask, which is a live process you can influence.

For anything commercial, current, or specific, engines lean heavily on the second. They retrieve. That is why ChatGPT can tell you about a company founded last month, and why answers change when your website changes.

So GEO is not about getting into the training data. You cannot do that, and anyone selling it is selling you something else. GEO is about improving your odds in the retrieval step, and improving what happens to your content once it has been retrieved. In practice that means:

• Being accessible to the crawlers that feed these systems. OpenAI, Perplexity, Google and Anthropic all operate distinct crawlers. Blocking them in robots.txt, or hiding key content behind JavaScript that renders poorly, removes you from the running before anything else matters.
• Being easy to lift a sentence from. Engines quote and paraphrase. Content built as a clear claim followed by evidence gets used. Content that buries the answer three scrolls into a narrative usually does not.
• Being corroborated elsewhere. Engines rarely retrieve only your website. They pull in comparison articles, directories, review sites, forums and trade press. If those sources describe you differently to how you describe yourself, the engine tends to trust them. Our guide to the source ecosystem in AI search goes into how that works.
• Being factually pinned down. Pricing, locations, service scope, team, credentials. Vagueness on your own site becomes hedging or invention in the answer.

GEO vs SEO: what actually changes

Plenty carries over. Crawlability, site speed, clear information architecture, genuine subject authority and being cited by credible third parties all still count, because retrieval is still built on search infrastructure.

What changes is the objective. SEO optimises for a click. GEO optimises for a mention, a citation, and an accurate description, in an answer that may never produce a click at all. That reframes several habits:

• Keyword volume matters less than question coverage.Nobody types “GEO agency UK” into ChatGPT. They type a paragraph about their situation.
• Page length matters less than answer density. A 3,000 word guide with no extractable claim in it performs worse than a 900 word page that states things plainly.
• Rank tracking stops working entirely, which we will come to.

GEO vs AEO: same job, different label

Being straight about this is more useful than defending a taxonomy.

Answer engine optimisation (AEO) grew out of the SEO industry. It originally described work aimed at featured snippets, People Also Ask boxes and voice assistants, and was then stretched to cover AI answers. GEO came from the research side and was about generative systems from the start. Our full guide to answer engine optimisation covers the AEO side in the same depth as this page covers GEO.

You will see people argue a distinction: that AEO is about extractive answers, where the engine lifts a passage more or less intact, and GEO is about generative answers, where the engine synthesises something new. That distinction is real in theory. In practice it is thin, because the underlying work is identical. Be retrievable. Be quotable. Be corroborated. Be accurate. Nobody has yet described a GEO tactic that is not also an AEO tactic.

Our practical advice is to use whichever term your buyers and your board already use, and to be sceptical of anyone selling you two separate programmes. If an agency’s GEO offering and its AEO offering have different deliverables, ask what the second one does that the first does not. Tilio runs one programme under both names, which is why our GEO agency page and our AEO work describe the same delivery.

How you measure GEO

You cannot rank track a generative engine. There is no position one. Ask the same question twice and you may get two different answers, with different sources, and neither is wrong.

That variability is the measurement problem, and the only reliable way around it is repetition at scale. The approach that works looks like this:

1. Build a fixed set of tracked prompts. Real buyer questions, phrased the way buyers actually phrase them, grouped by intent: problem-aware, solution-comparison, vendor-selection. This set stays stable so results are comparable over time.
2. Run them repeatedly, across engines. ChatGPT, Google AI Overviews and Perplexity behave differently and disagree with each other. A single engine tells you very little.
3. Record mentions and citations separately. A mention means your brand was named. A citation means a page from your site was used as a source. They move independently, and the gap between them is often the most actionable number you have. Our guide to mentions vs citations explains why.
4. Record which pages get cited. Frequently the cited page is not the page you would have chosen, which tells you exactly where to work next.
5. Record how you are described, and who is recommended instead. Presence is not the same as a good outcome. Being mentioned as the expensive option in an answer that recommends a competitor is a result you need to see.
6. Treat referral traffic as a lagging, partial signal. GA4 will show some ChatGPT and Perplexity sessions, but most AI answers produce no click at all. If you judge GEO on referral traffic alone you will underrate it substantially.

That is the methodology behind how we measure AI visibility at Tilio, and it is deliberately unglamorous. Consistent prompt sets, run often, compared against competitors on the same questions.

What GEO cannot do

Worth stating plainly, because the category attracts overclaiming.

You cannot buy placement in an organic AI answer. You cannot edit a model’s training data. You cannot guarantee a mention for a specific prompt, because the same prompt does not reliably return the same answer. You cannot fix a weak market position with content alone: if the trade press, review sites and comparison articles in your sector do not mention you, engines have little to work with, and that is a PR and positioning problem before it is a content one.

You also cannot expect this to move in a fortnight. Retrieval indexes refresh on their own schedule, and third-party corroboration takes longer still. Meaningful movement on tracked prompts usually takes a few months.

Where to start

If you are early, the sequence that wastes the least effort is: measure first, then fix the pages that already matter, then work on third-party corroboration.

Measuring first is not a stalling tactic. Without a baseline you cannot tell whether you have a visibility problem, an accuracy problem, or a competitor problem, and the fixes for those three are completely different. Our pricing page sets out where a managed programme starts and what a one-off audit covers.

If you are weighing up outside help, we apply the tests above, including the two-programmes question, in our rundown of the top GEO agencies in the UK. It names competitors as well as us, because a list that only recommends its author is not a list.

FAQs