What is generative engine optimisation (GEO)?
Generative engine optimisation (GEO) is the practice of making your content more likely to be used and cited when an AI system writes an answer, rather than only ranking as a link. The term comes from a 2023 research paper, and it now covers ChatGPT, Google AI Overviews and AI Mode, Perplexity, Gemini, Claude and Microsoft Copilot. It's one part of the wider work I do as AI search optimisation. This guide covers what the paper actually found, what it can't tell you, what Google says, and what the numbers look like in New Zealand.
Where the term GEO comes from
GEO was named in "GEO: Generative Engine Optimization", a paper by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. It was first posted to arXiv in November 2023 and published at KDD 2024, a leading data mining conference. The authors, including researchers at Princeton University and IIT Delhi, proposed GEO as a way for website owners to improve their visibility inside AI-generated answers.
The paper calls AI answer systems "generative engines": search systems that retrieve sources and then use a large language model to write a single answer from them, with citations. To test ideas fairly, the authors built GEO-bench, a benchmark of 10,000 queries across many subject areas, about 80% of them informational. You can read the paper on arXiv.
Is GEO a real thing, then? Yes. It was named in peer-reviewed research, the industry now uses it widely, and Google's own documentation uses the term while describing the work as part of SEO. What isn't real is any promise that GEO guarantees a citation.
What the GEO research found
In the paper's tests, rewriting a source to cite credible references, add quotations or add statistics raised its visibility in AI answers by up to about 40%. Keyword stuffing, the old SEO habit, did worse than making no change at all. On Perplexity, a live AI search engine, the best methods improved visibility by up to 37%. Results varied a lot by topic and by where a source started.
The authors tested several ways of rewriting a web page. The broad pattern is below, using the paper's own method names where they are clear.
| Method | What it does | Result in the paper's tests |
|---|---|---|
| Cite Sources | Adds references to credible sources | Among the strongest |
| Quotation Addition | Adds relevant quotations from credible sources | Among the strongest |
| Statistics Addition | Replaces vague claims with numbers | Among the strongest |
| Fluency and easy-to-understand rewrites | Makes the text clearer and simpler | Moderate gains |
| Authoritative tone | Makes the wording more confident | Small gains |
| Keyword Stuffing | Repeats the query's keywords | Worse than no change |
The finding almost nobody repeats is about who gains. Sources ranked lower in the underlying search results gained the most, and the top-ranked source often lost visibility. With Cite Sources, for example, the fifth-ranked source's visibility rose by 115.1%, while the top-ranked source's fell by 30.3%. For a smaller New Zealand business, that's encouraging: in these tests, being first in the search results was not what decided whether a source was used.
What the research can't tell you
The GEO paper's main experiments used a simulated AI search system built on GPT-3.5 Turbo, answering from the top five Google results, in 2023. It measured how much of each answer drew on a source, not clicks, enquiries or sales. Today's assistants use different models and retrieval, and they change often. Read it as evidence about what makes content useful to quote, not as a recipe.
The tests also rewrote pages in a controlled setting, which is not the same as publishing a page and waiting for real assistants to find it, trust it and use it. The durable lesson is general: specific, well-sourced, quotable content gets used, and padded, keyword-stuffed content doesn't. Anyone who tells you "GEO boosts visibility by 40%" without those caveats is quoting the headline, not the paper.
What Google says about GEO
Google's guide to generative AI search says optimising for its AI features is still SEO, and it treats AEO and GEO as names for work focused on AI visibility. There are no extra technical requirements: a page must be indexed and eligible to show with a snippet. Google also says you don't need llms.txt files, special schema or content cut into small chunks to appear in its AI features.
Google explains how the features work in plain terms. AI Overviews and AI Mode use retrieval-augmented generation, which Google also calls grounding: its ranking systems find relevant pages in the index, and the model writes an answer based on them, with links for people who want to go deeper. They also use query fan-out, running several related searches on subtopics at once, which lets an answer draw on a wider set of pages than a single search would return. I define these and other AI search terms in my SEO and AI search glossary.
Two more points are worth knowing. Search Console now has a generative AI performance report, available to all sites since 31 August 2026, so you can see how often your pages appear in AI features. And Google says mentions that aren't genuine don't help. This site publishes an llms.txt file for other AI systems that may read it; Google says it ignores it, and I'd rather say so. The source for all of this is Google's guide to optimising for generative AI features.
GEO and SEO: how they fit together
GEO isn't a replacement for SEO. It depends on it, because most AI answers are built from pages that search systems retrieve first. What GEO adds is attention to how quotable, specific and corroborated your content is, and whether you're mentioned beyond your own site. The overlap is large, and I lay it out side by side in AEO vs GEO vs SEO.
So is SEO being replaced by AI? No. The job has widened rather than disappeared, and I've set out the evidence in will AI replace SEO. The short version: if search systems can't find and trust your pages, no amount of GEO will get them into an answer.
GEO in practice: what I'd do first
Start with access, then substance. Make sure AI crawlers and Google can read your site, then rewrite the pages that matter most so they answer real buyer questions with specific facts, figures and named sources. After that, tidy your business identity across the web, earn genuine mentions, and measure repeatedly. That order follows the evidence: nothing else matters if your pages can't be read.
-
1
Confirm AI crawlers and Google can read the site
Check robots.txt as it is actually served, not just the file on your server. OpenAI's OAI-SearchBot fetches pages for ChatGPT search and PerplexityBot does the same for Perplexity; both are controlled in robots.txt, separately from GPTBot, which OpenAI uses for model training. In Google, make sure your important pages are indexed.
-
2
Answer real questions with facts, figures and sources
These were the paper's strongest methods. Replace general claims with specific ones, give numbers where you have them, and name the standard, regulator or study you rely on. Write the answer in the first few lines under a heading that asks the question.
-
3
Put facts in text, not PDFs or images
Services, prices, areas served, credentials and timeframes should be readable on the page. A brochure PDF or a price list baked into an image is much harder for any system to use.
-
4
Make your business entity consistent
The same name, description and details on your site, Google Business Profile and industry profiles, with accurate schema markup that matches the visible page. It won't make thin content worth citing, but it removes doubt about who you are.
-
5
Earn genuine corroboration
Mentions in industry publications, trade bodies and real reviews. Google says inauthentic mentions don't help, so skip anything that looks like a mention scheme.
-
6
Measure repeatedly
Ask the same buying questions across several assistants over several weeks, and use Search Console for Google. The measuring section below sets out how.
Two illustrations
These are invented examples to show the idea. They are illustrations, not clients.
A commercial cleaning company. Before: "We offer fast, reliable commercial cleaning." After: "We clean offices, medical clinics and retail stores across Auckland on nightly, weekly or one-off contracts. Most new sites get their first clean within 48 hours of signing, and after-hours callouts are charged at a fixed hourly rate listed on our pricing page." The second version gives an AI system facts it can repeat with confidence.
A building inspection firm. Before: "Our reports are thorough and fast." After: "Our pre-purchase reports follow NZS 4306:2005, the New Zealand standard for residential property inspection, and 9 in 10 are delivered within two working days." That one sentence adds a named standard and a statistic, two of the qualities the GEO research found were most used.
What GEO looks like in New Zealand
In my study of 1,000 New Zealand commercial searches, Google showed an AI Overview on 73.9% of them, and 80% of the sources those answers cited were .nz websites. Consumer-advice and comparison publishers were cited more often than the businesses they compare. For a New Zealand business, that means local, specific, well-sourced pages have a real chance of being used.
Intent made the biggest difference. Advice searches showed an AI Overview 96.5% of the time and cost searches 96.0%, against 16.5% for searches for a local provider, where the map pack still dominates. The study is a balanced sample of commercial searches captured on desktop on 8 September 2026, and it measured whether an AI Overview appeared, not who clicked. The full data is in my NZ Google AI Overviews Study 2026.
Business-to-business markets look different again. In The Thin Market Playbook, an AI assistant with live web search answered business buying questions mostly from providers' own websites, and the answers hadn't settled: asked the same question five times, six business industries produced an average of 16.3 websites, and none appeared in all five answers. In a market where the answer isn't fixed yet, a provider with clear, specific pages can become part of it.
How to measure GEO
Measure GEO in three ways: Search Console's generative AI performance report for Google, repeated tests of real buying questions in ChatGPT, Perplexity and Gemini, and enquiries tracked in your analytics. Test each question several times over several weeks, because AI answers vary between asks. One check is one draw, and a single screenshot proves very little either way.
Plenty of tools now offer a GEO checker or an AI visibility score, and some are useful for running many prompts at once. But Google says no third-party tool has access to its internal ranking or AI systems, so treat any score as an estimate and judge it against Search Console data and real enquiries.
If you'd rather have it checked for you, the free AI Visibility Audit tests your buying questions across the main assistants, and you can see how I run AI search optimisation for clients in New Zealand and overseas.
GEO questions
Is generative engine optimisation a real thing?
Yes. The term comes from peer-reviewed research published at KDD 2024, the industry uses it, and Google's own documentation names it while describing it as part of SEO. What isn't real is any promise that GEO guarantees citations. It improves the odds by making your content more specific, better sourced and easier for AI systems to use.
Who invented generative engine optimisation?
The term was introduced in the paper "GEO: Generative Engine Optimization" by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, with authors at Princeton University and IIT Delhi among them. It was first posted to arXiv in November 2023 and published at the KDD 2024 conference. The paper also introduced GEO-bench, a set of 10,000 test queries.
Is GEO the same as AEO?
Mostly. Both describe making content usable by AI systems that answer questions directly. AEO (answer engine optimisation) is the older term, from featured snippets and voice assistants; GEO names generative AI specifically, where an answer is written from several sources. In practice the work overlaps almost completely, and Google treats both as part of SEO.
Does schema markup help GEO?
It can help AI systems and search engines identify your business and what a page is about, which is why I use it. But Google says no special schema is needed for its AI features, and schema won't make thin content worth citing. Accurate markup that matches the visible page is the rule, and it works alongside good content, not instead of it.
Do I need an llms.txt file?
Not for Google: its guide says Google Search ignores llms.txt. Some other AI systems may read one, and it does no harm, so this site has one. Treat it as optional housekeeping, not a ranking or citation lever, and spend the time on the pages themselves, which is where AI systems find what they quote.
How long does GEO take?
Technical fixes can be picked up as soon as pages are recrawled, which can take days or weeks. Content and corroboration take longer, typically months, similar to SEO. Because AI answers vary between asks, judge progress over weeks of repeated tests and Search Console data, not a single screenshot, and be wary of anyone who promises a date.
Sources
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande: GEO: Generative Engine Optimization (arXiv 2311.09735; KDD 2024, doi:10.1145/3637528.3671900)
- Google Search Central: Optimizing your website for generative AI features on Google Search
- Google Search Central: AI features and your website
- Google Search Central Blog: Search generative AI performance reports (3 June 2026)
- OpenAI: Overview of OpenAI crawlers
- Perplexity: Perplexity crawlers
Want to know whether AI assistants use your site?
The free AI Visibility Audit tests the questions your buyers ask across ChatGPT, Google AI Overviews and Perplexity, and shows where you appear today. Clear findings, no obligation.
Book Your Free Audit