Generative Engine Optimization: What Actually Gets Cited in AI Answers

Generative Engine Optimization, GEO for short, is the work of getting content to appear in the answers of generative systems: in ChatGPT, Perplexity, Gemini, Copilot, the AI Overviews and Google's AI Mode. The goal is not a position in a list of results but a mention in an answer — as a source, as a provider or as evidence.

The term comes out of a paper by Aggarwal and colleagues from Princeton, Georgia Tech, the Allen Institute for AI and IIT Delhi, published in November 2023 and presented at the ACM SIGKDD conference in 2024. They measured which interventions in a text raise the probability that a generative system draws on it.

Why GEO Came About

For about two years now there has been a new layer between question and answer. Where you used to get ten blue links and made the choice yourself, you now often get a finished answer with three to five source references underneath. Anyone absent from that answer does not exist for that user — regardless of what position their page holds.

What the Numbers Say

Ahrefs measured the effect twice with the same method: 300,000 keywords, half with and half without an AI Overview, click-through rates from aggregated Search Console data.

In eight months the effect nearly doubled. That is the most solid figure on the subject, and it is the reason a good position no longer carries on its own.

The second figure is the more revealing one. In March 2026 Ahrefs examined where the URLs cited in AI Overviews come from: 863,000 keyword SERPs, four million cited addresses. Around 38 per cent of the citations came from the first ten organic results of the same search. In July 2025 it was still about 76 per cent.

The reason lies in how the systems work: they break a question into sub-questions and search for each one separately. Whatever ranks for the original query is therefore only one source among several.

The Gartner Forecast Has Expired

Nearly every text on the subject carries a figure from Gartner: search volume down 25 per cent by 2026, organic traffic down 50 per cent by 2028. The press release is dated 19 February 2024.

The first deadline has now passed. The drop by a quarter did not happen. Google still holds over 90 per cent market share in search, as of April 2026. The reason was foreseeable: Google pulled the answering function into its own search with the AI Overviews rather than ceding volume to chatbots.

Passing the forecast on today as a statement about the future means quoting a prediction whose deadline has come and gone. The shift that did take place is a different one, and less comfortable in practice: the queries stayed. The clicks did not.

How a Generative System Arrives at Its Answer

Without this intermediate step every measure stays an assertion. An answer comes from two sources.

The first is training. What a model learned about a topic or a brand comes from text that was available at the time of training. That layer cannot be edited, it is old, and for smaller brands it is thin to empty.

The second is retrieval at runtime, technically retrieval augmented generation. On a question the system looks for matching documents, reads them and formulates the answer from them. This is where GEO applies: that layer can be influenced, because it consists of retrievable pages.

Query Fan-Out

The systems do not issue one search query but several. "Which agency fits for B2B link building?" turns into sub-questions about selection criteria, prices, providers and risks. Each sub-question gets its own search.

In practice that means a page can be cited without ranking for the original question at all — because it answers one of the sub-questions cleanly. And conversely a page at position 1 can go uncited if it offers no self-contained statement on any sub-question.

That explains the drop from 76 to 38 per cent. And it shifts the work: away from the one perfect page for a keyword, towards solid statements on many concrete questions.

GEO and SEO: What Stays, What Is Added

GEO does not replace SEO, it builds on it. A page that is not indexed does not get retrieved either.

The row that makes the difference is the fifth. For a ranking the page counts as a whole. For a citation what counts is the single sentence that can be lifted out and reproduced without context.

GEO and AEO: Why the Terms Belong Apart

Answer Engine Optimization, AEO for short, is often used synonymously. The terms overlap, but they do not mean the same thing.

• AEO is the older and narrower term. It comes from the era of featured snippets and voice assistants: optimizing for the one direct answer to a concrete question.
• GEO is broader. It concerns generative systems that assemble an answer from several sources, take context into account and follow up in dialogue.

The blurring is not a quarrel about words. Whoever does AEO optimizes individual question-answer pairs. Whoever does GEO additionally has to make the brand recognisable as an entity and get independent sources to say the same thing. That is considerably more work, and it falls in a different place.

The Levers, Ordered by Evidence

Plenty of recommendations circulate about GEO and few of them come with evidence. The following order separates the two.

Evidenced: What the Study Measured

The paper by Aggarwal and colleagues tested nine interventions on texts, against a benchmark of 10,000 search queries. The most effective ones raise visibility in generated answers by up to 40 per cent:

• Cite sources. Statements with a reference to where they come from get picked up more often than those without.
• Include figures. Concrete values instead of general descriptions raise the probability of being taken up considerably.
• Use quotations. Verbatim statements from nameable people or institutions work in a similar way.
• Write clearly. Readable, well-formed versions are preferred over convoluted ones.

What is remarkable about the list is what is missing: keyword density, text length, heading structure. The interventions that work concern how well a statement is evidenced, not how it is optimized for a search term.

One point of context: this is a laboratory finding on a benchmark, not a guarantee for your own market. The study itself notes that the effect varies by subject area.

Plausible: What Follows From How the Systems Work

These levers are not measured in a study, but they follow directly from how the systems operate:

• Spell entities out. Company names, product names, people with their role, places with the country — instead of "we" and "our offering". A model links named things, not pronouns.
• One statement per paragraph. What is meant to be quoted has to survive being lifted out.
• Answer sub-questions directly. The question as the heading, the answer in the first sentence below it, the reasoning after that.
• Tables for comparisons. Structured entries get reproduced correctly more reliably than running text with subordinate clauses.
• Make the date visible. Without a status, every entry is timeless, including the superseded one.
• Add structured data. Schema.org helps with assignment but does not replace the readable text.

The Lever Outside Your Own Website

Your own page is one source among many. Generative systems weight matching statements from independent sources higher than a provider's account of itself. When trade media, directories and guides say something different from your own page, your own page rarely prevails.

That is why part of GEO happens off your own domain: mentions in third-party publications, appearances in comparisons and overviews, evidence a system can lean on. How to build that is covered in AI Mentions; the fact page as its counterpart in Grounding Page.

This external evidence is exactly what you build predictably with GetMentioned: pick fitting publishers, compare SEO data and prices, book placements directly and document publications centrally.

What GEO Does Not Do

It Does Not Force a Citation

There is no mechanism obliging a system to draw on a particular page. Every lever raises probabilities. Anyone promising guarantees is selling something else.

It Does Not Replace Rankings

Even with only about four in ten citations coming from the top 10: those four are real, and discoverability through classic search remains the route by which a page enters the retrievable pool at all.

There Is No Control Over the Answer

A system formulates for itself. It can quote correctly and still place the quote in a context you would not have chosen. What is steerable is the source situation, not the result.

It Does Not Work Immediately

Weeks pass between publication and effect. A page has to be indexed, retrieved and weighted against other sources. Anyone seeing no effect after two weeks has not done too little; they have looked too early.

How to Measure Success

Getting started costs ten minutes: ask the systems that matter to you the questions a prospect would ask. Note whether your brand gets named, in what context, alongside which competitors, and which sources carry the answer.

That spot check becomes a measurement once it is repeatable: a fixed set of questions, a fixed interval, the same systems, documented results. How to set that up systematically, with metrics such as Share of Voice and Share of Model, is covered in AI Visibility.

Conclusion

Generative Engine Optimization is not a new discipline with tricks of its own but the extension of an old question under new conditions: why should a system draw on you in particular?

The evidenced answers to that are unspectacular. Statements that stand on their own. Figures with a source. Clear language. Names spelled out. And, off your own website, independent sources saying the same thing. That is less exciting than the forecasts the subject is usually sold with, and it lasts longer.