Generative Engine Optimization: GEO vs SEO
Generative Engine Optimization is the practice of getting your business cited inside an AI-generated answer. The term comes from a 2023 research paper by Aggarwal and colleagues, accepted to KDD ‘24, which tested content changes against generative engines and found that adding citations, quotations and statistics lifted visibility by 30 to 40 per cent. In 2026 the word is used more or less interchangeably with AEO and LLM SEO, and Google’s own documentation says that optimising for generative AI search is optimising for the search experience, “and thus still SEO”. The useful question is not what to call it. It is which parts genuinely differ from ordinary SEO, and there are only three.
The rest of this sets out where the term came from, what the research actually measured, and which three differences survive scrutiny.
What is Generative Engine Optimization?
A generative engine is any system that answers a question by writing a paragraph rather than returning a list of links: ChatGPT, Google’s AI Overviews and AI Mode, Perplexity, Claude, Copilot. Generative engine optimisation is the work of being one of the sources such a system draws on, and of being named when it does.
The unit of success is a sentence, not a position. If the answer describes your category, names three businesses and yours is not among them, you did not come fourth. You were absent. There is no equivalent of page two, and no partial credit for having been the eleventh most relevant result.
One note on spelling before it becomes distracting. We write optimisation the Australian way in most of our copy. The term itself arrived from an American research paper and kept its American spelling, which is why the phrase people search for is “Generative Engine Optimization” and the thing we do all day is optimisation. Both spellings mean the same thing and neither affects whether a model quotes you.
The practical content of GEO overlaps with SEO by most of its surface area. Crawlable pages, accurate facts, clear structure, real expertise, pages that load. The overlap is not a disappointment. It is the finding. What follows is an attempt to be precise about the part that is left over.
Where the term came from
The phrase was introduced in “GEO: Generative Engine Optimization” by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, first posted to arXiv as 2311.09735 in November 2023 and later accepted to KDD ‘24. The authors built a benchmark of queries, ran candidate websites through generative engines, and tested nine content modifications to see which ones changed how much of the generated answer a given source accounted for.
The headline result is the one everyone quotes: adding citations, quotations and statistics lifted visibility by 30 to 40 per cent. The less quoted result is what failed. Keyword stuffing performed poorly. The tactic that defined a decade of bad SEO did not transfer.
The paper is worth knowing about for a practical reason rather than a bibliographic one. Most writing on GEO describes the term as having emerged with AI search generally, which leaves the impression that there is nothing underneath it to check. There is. The methods were tested, the results were published, and some of them are counterintuitive enough to be worth reading in the original.
GEO and SEO: what actually differs
Most GEO-versus-SEO comparisons you will find are written by people selling a GEO product, which tends to widen the gap. Forrester analyst Nikhil Lai has put the counter-case bluntly: the terms are significantly but not fundamentally different from SEO, and their advocates have an incentive to overstate the distance, because the distance is where a new category of tooling gets sold.
That is our read too. Here is the comparison as we would draw it, with the overlaps left visible rather than tidied away.
| Dimension | Classic SEO | Generative engine optimisation |
|---|---|---|
| What you compete for | A position in a list of links | A sentence inside a written answer |
| What the reader sees | Ten results, yours among them | One paragraph, sometimes with citations |
| What reads your page | A crawler and a ranking system | A crawler, a retrieval system, and a model that rewrites what it retrieved |
| What second place looks like | Position two, still clickable | Nothing. Absence is the only alternative to citation |
| What you change on the site | Titles, headings, structure, internal links, speed, content | The same list, in much the same order |
| Where off-site weight sits | Links from other domains | Being written about by other domains, in prose a model can read |
| What feedback you get | Impressions, clicks and positions in Search Console | Sampled answers. Most assistants report nothing |
| How fast feedback arrives | Days to weeks | Weeks to months. Indexes and models refresh on their own schedule |
| Whether accuracy is a metric | It is not. Nobody asks if Google described you correctly | It is. Being described wrongly is its own failure |
Five of those nine rows are differences of reporting and speed rather than differences of work. The row about on-site changes is the important one: the things you actually edit are the same things. If a page is slow, thin, badly structured or wrong, it fails in both worlds for overlapping reasons.
GEO, AEO and LLM SEO
Three names, one job, and no settled boundary between them. Wikipedia’s article puts it accurately: the academic literature has not settled on a definition that separates these terms, and in practice they are used interchangeably.
We use all three in writing and none of them in the product. If you want the distinctions as far as anyone draws them, they are in our glossary entry for GEO and the one for AEO, with answer engine optimisation covered at length here and the LLM-facing half here. This section exists so those pages do not have to compete with this one, and so you can stop wondering whether you are buying three things. You are buying one.
What Google actually says
Google’s search documentation, updated July 2026, states its position without much room for interpretation: from Google Search’s perspective, optimising for generative AI search is optimising for the search experience, “and thus still SEO”.
The same documentation advises prioritising effective SEO strategies over “AEO/GEO hacks”, and states that Google Search ignores llms.txt-style files.
That last point has been tested. Ahrefs looked at 137,210 domains in May 2026 and found that 28 per cent publish an llms.txt file, and that 97 per cent of those files received zero requests. A file that nothing fetches is not a tactic. It is a text file. If you have one it costs nothing to leave in place, but it is not the reason anything happens.
Take Google’s line seriously without taking it as the whole picture. Google is describing its own surfaces, and it has a commercial interest in publishers not treating AI answers as a separate discipline requiring separate tooling. The statement is still the most authoritative thing on the record, and it agrees with what the independent research shows about where the overlap sits.
The three things that genuinely are not SEO
Everything above is the case for GEO and SEO being nearly the same job. Here is the honest remainder. Three findings do not reduce to ordinary SEO, and they are the reason this article exists.
Generative visibility is not monotonic in search rank
This is the strangest result in the original paper and the least discussed. Aggarwal and colleagues found that the Cite Sources method produced a 115.1 per cent visibility increase for websites ranked fifth, while the top-ranked website’s visibility fell by 30.3 per cent. The same change helped the site in fifth and hurt the site in first.
Being first is not the same as being quoted. A generative engine assembling an answer is not ranking sources against each other. It is composing a paragraph, and it takes the sentences that fit the paragraph. A page that ranks first and buries its point is less quotable than a page that ranks fifth and states its point in the opening line.
The corroborating evidence from outside the lab points the same way. Ahrefs studied 863,000 keyword SERPs and 4 million AI Overview URLs in March 2026 and found that 38 per cent of AI Overview citations rank in the top 10, down from roughly 76 per cent in July 2025. About 31 per cent rank beyond position 100. Roughly a third of what gets cited is not ranking in any sense a rank tracker would notice.
Read those two numbers together and the trend matters more than the level. Citation and ranking were closely coupled in mid-2025 and are markedly less coupled now. If that continues, a rank report becomes a weaker and weaker proxy for whether you are being quoted, which is a problem for everyone who reports on rankings, ourselves included.
The retrieval mix inverts on-site-first SEO logic
Chen, Wang, Chen and Koudas, in arXiv 2509.08919 from September 2025, found that AI Search shows a systematic and overwhelming bias toward earned media, meaning third-party authoritative sources, over brand-owned and social content. They contrast this with Google’s more balanced mix. They also report a big brand bias that niche players have to overcome.
This is uncomfortable, and we would rather say so than dress it up. Classic SEO is on-site-first: the page you control is the asset, and links are the supporting cast. On this evidence, generative answers invert that. What other people wrote about you carries more weight than what you wrote about yourself, and large brands start ahead.
For a small business that is bad news twice over. Being written about elsewhere is slower, less controllable and less automatable than fixing your own pages. We build software that works on a site continuously, and we will say plainly that this is the part it cannot do for you. Earned media comes from doing something worth writing about and then telling people who write. A crawler cannot manufacture that, and any tool claiming otherwise is describing either a press-release blast or a link scheme.
What is automatable is the groundwork that makes coverage usable when it arrives: consistent facts everywhere a model might read them, pages that answer the question a journalist or a model is actually asking, and a record of what changed and when. That is real, and it is not the same as earning the coverage.
Accuracy is a metric with no real SEO analogue
In classic SEO nobody asks whether the search engine described your business correctly, because it does not describe your business. It shows your title tag and a snippet you largely wrote.
A generative engine writes its own sentence about you. It can get your services, your location, your pricing model, your founding date or your specialism wrong, and it will do so fluently. Being described wrongly is a distinct failure from not being mentioned, and it is arguably worse: an absent business is invisible, a misdescribed business is actively working against itself in front of a buyer who trusts the answer.
This needs its own measurement, and almost nothing in the standard SEO toolkit provides one. Rank tracking cannot see it. Search Console cannot see it. The only way to know is to ask the assistants questions about your own business on a schedule and read what comes back for factual errors, not just for whether your name appears. That is a different check from citation counting, and it is the one most likely to find something you can fix this week.
How to do GEO
The list below is ordered by evidence, not by novelty. The first four items are just good SEO, which is the point.
- Put the answer in the first paragraph of every commercially important page. This is the single change most consistent with the research, and it costs nothing but nerve.
- Write self-contained sentences. A sentence beginning “as noted above” cannot be lifted into an answer. One that names its subject can.
- Add citations, quotations and statistics where they are real. This is the modification the original paper measured at a 30 to 40 per cent visibility lift. The qualifier matters: a fabricated statistic is worse than none, and models increasingly cross-check.
- Make your facts identical everywhere. Site, profiles, directories, listings. A model with three versions of your founding date usually declines to state one.
- Fix the boring technical layer. If a page cannot be fetched, parsed or rendered, none of the above applies to it. Crawlability is a precondition, not a tactic.
- Do the earned-media work, knowing it is manual. Be useful to someone who publishes. This is the slowest item and, on the September 2025 evidence, the one with the most weight behind it.
- Skip the hacks. Keyword stuffing performed poorly in the paper that coined the term. llms.txt is ignored by Google Search and, per Ahrefs, mostly unfetched by anything else. Hidden instructions addressed to an assistant are treated by vendors as an attack to defend against rather than a preference to honour, so anything that works today is a bug on its way to being closed.
If that list looks like ordinary SEO with two additions, that is the correct impression.
How to measure it
Measurement here means putting a fixed list of questions to the assistants on a repeating schedule and recording two things: which businesses get named, and what is said about them. That is citation tracking, and it is sampling rather than measurement. The same question can name you on Tuesday and not on Thursday, with nothing having changed on your site.
So treat the trend as the signal and any single answer as noise. Three things are worth recording each time:
- Whether you were named at all.
- Which source the assistant cited, since it is often not your site.
- Whether what it said about you was true.
The third column is the one nobody builds, and per the section above it is the one with no SEO equivalent. The most useful output of the whole exercise is usually the list of questions where you never appear, because that list is short, specific and actionable in a way that a visibility percentage is not.
We do not guarantee rankings. We guarantee the work: the checks run, the log is readable, and what fails the standard does not ship. Nobody can guarantee a citation from a system that rewrites its answer every time it is asked, and you should treat a promise of one as a description of the seller rather than of the market.
Where to start
- Write down one description of your business and put it everywhere, unchanged.
- Move the answer to the top of your five most commercially important pages.
- List the ten questions a customer asks before buying from you, and ask an assistant each of them. Record who gets named.
- Ask three assistants to describe your business, and mark every factual error.
- Pick one thing you know that your competitors have not published, and publish it with the numbers attached.
Step four is the one most people skip and the one most likely to return something surprising. Step five is the hardest, the slowest, and the only item on the list that a search engine, a language model and a human reader all reward for the same reason.