What Is Generative Engine Optimization (GEO)?
Generative Engine Optimization is the practice of getting your content selected, quoted and cited inside AI-generated answers — the summaries produced by Google’s AI Overviews and AI Mode, ChatGPT, Perplexity, Gemini and Copilot — rather than simply ranking in a list of blue links.
That definition is unusually well-founded, because unlike most marketing acronyms GEO has a documented origin: a peer-reviewed paper presented at KDD 2024. What has happened since is that the term escaped the paper and became a sales category, and the version being sold rarely resembles the version that was tested. This guide covers what GEO actually is, the mechanism it operates on, what the controlled evidence supports, and what is genuinely new work as opposed to SEO with a new invoice line.
If you are mainly trying to work out how GEO relates to AEO, LLM SEO and AI SEO as terms, our comparison of GEO, AEO and LLM SEO settles the naming question. This piece goes deep on GEO itself.
Where the term came from, and what the paper actually claimed
GEO was introduced in “GEO: Generative Engine Optimization” by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, published at KDD 2024 (arXiv 2311.09735). The paper’s own framing is worth quoting because it is more careful than anything built on top of it. It introduces GEO as “the first novel paradigm to aid content creators in improving their content visibility in generative engine responses through a flexible black-box optimization framework”, and states that “GEO can boost visibility by up to 40% in generative engine responses”.
Three details matter enormously and are almost always dropped in the retelling:
- It measured visibility, not traffic. The authors built custom metrics — principally a position-adjusted word count, which weights how much of your content appears and how prominently, plus a subjective impression score. No click, session or revenue outcome was measured.
- It was evaluated on GEO-bench, a benchmark of roughly 10,000 queries drawn from nine sources including MS MARCO and Natural Questions, split into training, validation and test sets.
- The authors flag the core limitation themselves: generative engines have a “black-box and fast-moving nature”, and “content creators have little to no control over when and how their content is displayed”. The evaluation stack is also two years old now and not the production systems you are trying to appear in.
So GEO is a real, published research direction with a measured effect, and it is not a specification. Anyone quoting the 40% figure at you as a promised traffic uplift is misreading a visibility metric as a business outcome.
How a generative engine actually builds an answer
You cannot optimise a process you cannot picture. Every current AI search surface runs a broadly similar four-stage pipeline, and each stage is a different intervention point.
The four stages of a generative answer, and what a site owner can actually influence at each one.
The first stage is where the biggest misunderstanding sits. Google’s own documentation states that AI Overviews and AI Mode “may use a ‘query fan-out’ technique — issuing multiple related searches across subtopics”. Your page is not competing for the query the person typed. It is competing across a set of generated sub-questions, each retrieving independently. A page can therefore be pulled into an answer for a query it has never ranked for, and can be absent from an answer for a query it ranks first on.
Why GEO is a distinct idea: citation came apart from ranking
If citations simply followed rankings, GEO would be a marketing word for SEO. The measurable position is that they no longer do.
Ahrefs analysed 863,000 keyword SERPs and 4 million AI Overview URLs in March 2026. Only 38% of the pages cited in AI Overviews also rank in the top 10. Another 31.2% sit at positions 11 to 100, and 31.0% come from outside the top 100 altogether. In the same study run in July 2025, roughly 76% of citations were drawn from the displayed search results; by March 2026 that had halved to about 38%, which is exactly what you would expect as fan-out retrieval takes over from SERP re-use.
Semrush found something compatible earlier, analysing 200,000 US keywords in September 2024: only 20% to 26% of AI Overview links matched the top 10 organic results, with around 11 links cited per overview. Two independent vendors, eighteen months apart, describing the same detachment.
The practical translation is that the unit of optimisation has changed. Classic SEO optimises a page against a query. GEO optimises a passage against a question that may never have been typed. That is a genuine difference in the work, and it is the strongest argument that GEO is more than rebranding.
What the research says actually works
The KDD paper tested nine content modifications against unedited baselines. The consistent winners were what the authors group as authority and evidence signals: adding quotations from credible sources, adding relevant statistics, citing sources explicitly, and improving fluency. Reported gains for the best-performing methods sit in the region of 30% to 40% on the paper’s visibility metrics, with effects varying substantially by domain — the authors are explicit that “the efficacy of these strategies varies across domains”.
The most useful result is the negative one. Keyword stuffing did not merely fail to help; it scored below the unmodified baseline. The oldest manipulation in search does not transfer to generative engines at all, because the engine is choosing text to quote, not matching a term frequency.
| Content change | Effect reported | What it means in practice |
|---|---|---|
| Adding quotations from credible sources | Among the largest gains | Quote a named authority rather than paraphrasing it anonymously. |
| Adding relevant statistics | Among the largest gains | A sentence with a figure in it is a sentence an engine can lift. |
| Citing sources explicitly | Substantial gain | Name the study, the sample and the date in the text, not just in a link. |
| Improving fluency | Substantial gain | Clear, self-contained sentences survive extraction; convoluted ones do not. |
| Keyword stuffing | Below baseline | Actively counterproductive. Do not carry this habit across. |
Source: Aggarwal et al., “GEO: Generative Engine Optimization”, KDD 2024. Directional evidence from a 2024 evaluation stack, not a current specification.
What independent data supports — and what it kills off
Two 2026 findings should shape what you spend on, and they point in opposite directions to most GEO sales material.
Structured data does not appear to buy citations. Ahrefs studied 1,885 pages that added JSON-LD between August 2025 and March 2026 against roughly 4,000 matched control pages. They found no citation lift across AI Overviews, AI Mode or ChatGPT, and a small but statistically significant negative effect of about 4.6% on AI Overviews. Google’s own guidance agrees in principle: “You don’t need to create new machine readable files, AI text files, or markup to appear in these features.” Keep schema markup for rich results, where it demonstrably works. Stop selling it as a GEO tactic.
Being talked about correlates far better than being linked to. Ahrefs measured 75,000 brands against AI Overview brand visibility using Spearman correlations. Branded web mentions came in at 0.664 and branded anchors at 0.527, against 0.218 for raw backlink counts, 0.295 for referring domains and 0.326 for Domain Rating. Ahrefs state clearly that correlation is not causation, and around 26% of brands studied had no AI Overview mentions at all. Read it as a strong steer toward earned coverage and third-party presence rather than a link-building target.
There is also a substantial set of GEO claims we deliberately do not repeat, because no published sample, date or method exists behind them: that FAQ schema is weighted some percentage higher, that structured pages get a multiple of the citations, that a fixed share of citations comes from the first third of a page, or that a named platform cites only a set percentage of pages it retrieves. If a statistic in a GEO deck has no sample size next to it, treat it as a slogan.
GEO against SEO: what is genuinely new work
This is the question a business owner actually needs answered, so here it is without hedging.
A working separation between the GEO tasks that are new, the ones that are SEO renamed, and the ones that simply matter more than they used to.
The left column is genuine. Designing a section so that a single paragraph answers a sub-question completely, without needing the two paragraphs above it, is a real craft change. So is maintaining a monitored prompt set, deciding which AI crawlers to allow, and reporting share of citations. But it sits inside an SEO engagement rather than beside it, which is why we are consistently against a separate GEO retainer. Google’s position, stated plainly in its documentation, is that there are “no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary”.
Doing GEO on a real site: a working sequence
For a Singapore business starting from a normal, reasonably healthy website, this is the order we would work in.
- Confirm eligibility first. Indexed, snippet-eligible, no accidental
nosnippetormax-snippetrestrictions, no retrieval crawlers blocked in robots.txt. Nothing else matters if this is broken — see technical SEO basics. - Map the sub-questions, not the keyword. Take your ten most commercially important topics and write out the fifteen to twenty questions a buyer would need answered around each. That set is a much better proxy for fan-out than a keyword list. Our keyword research guide covers how to build it from real demand data rather than guesswork.
- Rewrite for extraction. Each significant question gets a heading and an answer that stands alone in one short paragraph, with the number, date or condition inside the sentence rather than implied by the section around it.
- Add the evidence the research rewards. Named sources, sample sizes, dates, quotations from identifiable authorities. This is the single most transferable finding from the GEO paper, and it happens to be good writing anyway.
- Build off-site presence deliberately. Earned mentions, credible third-party lists, video, and accurate business listings. This is slow work with the strongest observed correlation to AI visibility.
- Instrument it before you judge it. A fixed prompt set checked monthly, Search Console’s generative AI impressions where available, and an isolated AI-referral segment in GA4.
Measuring GEO without fooling yourself
Measurement is where most GEO programmes quietly fail, because the tooling is genuinely immature and the gaps are not obvious.
Google announced Search generative AI performance reports for Search Console in June 2026, separating AI Overview and AI Mode impressions from standard organic data for the first time. The reports launched as a limited rollout and show impressions, pages, countries and devices — no clicks, no click-through rate and no query data. That is genuinely useful for direction and useless for attribution.
In GA4, the AI Assistants channel captures referrals from chatbots such as ChatGPT and Gemini. It explicitly excludes traffic from Google’s AI Overviews and AI Mode, which is classified as Organic Search. So “AI traffic” in your GA4 property means chatbot referrals only, and a large share of assistant referrals arrive with no referrer at all and land in Direct. Our guide to the GA4 reports worth checking and the UTM tagging guide cover how to segment this cleanly, and attribution models explains why the last-click view will understate it.
The measure that survives contact with reality is share of citations against a fixed prompt set: fifty to a hundred questions your buyers genuinely ask, checked on the same schedule across the same platforms, recording whether you appear and who appears instead. It is manual, it is noisy because these systems are probabilistic, and it is still the most honest signal available. Treat any single month’s movement as weather, not climate.
What GEO is not
Four red flags are worth naming, because all four are currently being sold in this market.
- Guaranteed citations. Nobody can guarantee them. The engines are probabilistic and, as the KDD authors put it, creators have “little to no control over when and how their content is displayed”.
- A separate technical file or markup you must buy. Google says the opposite in writing, and the schema evidence is negative.
- A rebuilt site “for AI”. If your site is crawlable and fast, it is already eligible. If it is not, that is a technical SEO fix with a known scope.
- A GEO retainer alongside an SEO retainer. One programme, one budget. If the two proposals overlap on 80% of the tasks, you are paying twice for the same work.
Market rates for SEO in Singapore vary widely by scope, and we break down what drives them in our SEO cost guide. The point here is narrower: GEO should change the composition of the work, not the number of invoices. If you want to see the outcomes the underlying work produces, our case studies are the honest version.
Frequently asked questions
What does generative engine optimization mean?
Generative Engine Optimization is the practice of improving how often and how prominently your content is selected, quoted and cited inside AI-generated answers such as Google’s AI Overviews and AI Mode, ChatGPT, Perplexity and Gemini. The term was introduced in a KDD 2024 research paper that defined it as a framework for improving content visibility in generative engine responses.
Is GEO different from SEO?
Partly. The eligibility layer is identical: you must be indexed, snippet-eligible and technically healthy, and Google states there are no additional requirements or special optimisations for AI Overviews and AI Mode. What differs is the unit of optimisation. Classic SEO optimises a page against a query; GEO optimises a passage against generated sub-questions. That is a real craft change, but it belongs inside an SEO programme rather than beside it.
Does the 40% visibility improvement from the GEO paper apply to my website?
Not directly. The paper reported gains of up to 40% on its own visibility metrics, principally a position-adjusted word count, measured on a 10,000-query benchmark using a 2024 evaluation stack. It did not measure clicks, sessions or revenue, and the authors note that effectiveness varies by domain. Treat it as directional evidence that evidence-rich writing gets quoted more, not as a forecast for your traffic.
Does schema markup improve GEO results?
The available evidence says no. Ahrefs compared 1,885 pages that added JSON-LD between August 2025 and March 2026 against roughly 4,000 matched controls and found no citation lift, with a small negative effect on AI Overviews. Google also states you do not need new markup or AI text files. Schema is still worthwhile for rich results in classic search.
What actually gets a page cited by an AI answer?
On the controlled evidence, the strongest levers are adding relevant statistics, quoting credible named sources, citing sources explicitly and writing clearly enough that a passage can be lifted without its surrounding context. Keyword stuffing performed below the unedited baseline in the same tests. Off-site brand mentions correlate more strongly with AI Overview visibility than backlink counts do.
How do I measure whether GEO is working?
Use three layers. Search Console’s generative AI performance reports for AI Overview and AI Mode impressions where you have access, a GA4 segment for AI assistant referrals while remembering that Google’s own AI surfaces are classified as Organic Search there, and a fixed prompt set of fifty to a hundred buyer questions checked monthly to track whether you are cited and who is cited instead.
Where this leaves you
GEO is a real research direction with one solid, well-evidenced core idea: generative engines quote text that carries evidence, and citation has detached from ranking. Around that core sits a large amount of commercial noise, some of it directly contradicted by controlled studies and by Google’s own documentation.
The practical version is unglamorous. Stay eligible, cover whole topics rather than single keywords, write claims that survive being lifted out of context, earn mentions off your own site, and measure with three imperfect instruments instead of one confident number. That is the whole of it, and it is close enough to good SEO that treating it as a separate purchase is the mistake to avoid.
The AI SEO guide for Singapore puts this in the wider context of both senses of AI SEO, showing up in AI Overviews goes deep on Google’s surface specifically, and the AI SEO service page sets out how we run this work for Singapore businesses.
Written by Adrian Tan, Singapore Digital Marketing. Last updated 5 August 2026.
For the practical version of that comparison — a job-by-job look at what actually changes in the work, rather than what the terms mean — see AI SEO vs traditional SEO.
