Generative Engine Optimization (GEO) refers to measures that help content appear in the answers of generative AI systems such as ChatGPT, Gemini or Perplexity. The term comes from a 2023 study by researchers at Princeton University and other institutes. That study showed that paragraphs with figures, quotes and source references were adopted more often in test environments. Much of what is sold as GEO today goes beyond this and is not backed by evidence.
The term GEO has appeared in almost every AI visibility offer since 2024, often alongside AEO and LLMO, and rarely with an explanation of what it actually means. Yet GEO has a clear origin and a manageable, evidence-based core. This page separates what comes from the research, what is practical experience, and what is a sales pitch.
Where the term comes from
Generative Engine Optimization was coined in November 2023 by a research group led by Pranjal Aggarwal and Vishvak Murahari, with Princeton University, Georgia Institute of Technology, the Allen Institute for AI and IIT Delhi involved. The study "GEO: Generative Engine Optimization" (arXiv 2311.09735) examined which changes to a text increase the likelihood that a generative search system uses the text in its answer and cites it as a source.
To do this, the researchers built their own test environment with questions and sources, then systematically altered the sources: adding figures, quotes, source references, improving readability, or adding technical terms or keywords. They then measured how often and how prominently each variant appeared in the generated answers.
What is backed by evidence
Three findings can be drawn from the study, each specific to the test environment examined:
- Figures, quotes and source references helped. Paragraphs containing concrete statistics, direct quotes from experts or references to sources were adopted significantly more often than the original versions.
- Keyword stuffing didn't help. Adding extra search terms brought no benefit, and in some tests was even a disadvantage.
- The effect depended on the topic. What helped with health-related questions had a weaker effect on other topics. There's no single measure that works the same everywhere.
The exact percentage figures from the study are deliberately left out here, as they relate to the test environment and not to real websites in real AI systems. What remains is the direction: concrete, well-evidenced and quotable content is favoured. This matches what Google states in its documentation on AI features about AI Overviews, namely that the same quality standards apply there as in normal search.
What is assumption
Much of what is offered under the GEO label goes beyond the study and is not publicly evidenced:
- That an llms.txt file increases visibility. No major provider has confirmed that it evaluates one.
- That certain Schema.org types are favoured by language models. Structured data helps Google understand content, but a direct effect on ChatGPT or Perplexity isn't documented.
- That a model's training data can be deliberately influenced. What content goes into training isn't public, and models are retrained at intervals of months.
- That visibility in AI answers can be increased by a specific percentage. Such figures usually come from providers' own measurements and can't be verified.
What can be observed instead: the systems companies recommend mostly rely on live search. OpenAI describes the OAI-SearchBot in its documentation, which reads websites for ChatGPT's search feature. Perplexity searches with every request. Google feeds AI Overviews from its search index. This means what's on your website today and found in search can appear in answers within a few weeks. You can check in two minutes whether crawlers even have access, using the test at Can AI crawlers read your website?.
Distinction from AEO
| AEO | GEO | |
|---|---|---|
| Question | Is my paragraph used as the answer to a specific question? | Does the model know my company and mention it, even without a direct question? |
| Unit | The paragraph on a page | The company as a concept, i.e. as an entity |
| Lever | Structure of the text: question, answer first, concrete statements | Consistent data across all sources, mentions on other websites, clear positioning |
| Example | "How long does a bathroom renovation take?" is answered by your paragraph | "Which plumbing company in Frauenfeld would you recommend?" names your company |
The two work together: a website with quotable paragraphs strengthens the entity, and a well-known entity is more likely to be chosen as a source. What AEO specifically requires is covered under What is AEO?.
Practical rules for SMEs
From the evidence-based core and how the systems work, six rules follow that can be implemented without an agency:
- Concrete figures instead of adjectives. "Three to four weeks" instead of "fast", "from CHF 18,000" instead of "fair". Figures with a date and conditions.
- Your own experience as evidence. "In around half of our renovations over the past three years" is a source reference only you can provide.
- Quotes from experts, with names. A sentence from the site manager, with name and role, is what the study measured as a quote.
- Consistent company details everywhere. Name, address, services on your website, business profile, directories and social media all matching exactly.
- Mentions on other websites. Associations, suppliers, local authorities, media. Every mention is evidence that the company exists and does what it says.
- Measure monthly. Ten fixed questions in ChatGPT, Perplexity and Google, results in a table. Tools for this are under Measure AI visibility.
None of these rules are new. They reflect what good websites have been doing for years. GEO doesn't change the work, it explains why it now also has an effect in a second place. The post Appearing in AI answers: three strategies for SMEs describes what's already visible in AI answers for SMEs and what isn't.
Frequently asked questions
Is GEO the same as AEO?
No, even though both terms are often used interchangeably. AEO aims at directly answering a question and being named as the source. GEO is broader and asks whether a language model knows a company or piece of content at all and uses it in generated text, even without anyone asking specifically.
Can GEO get ChatGPT to recommend my company?
You can influence it, not force it. Whether a model names a business depends on many sources: search results, directories, reviews, mentions on other websites. A business that appears consistently there and makes concrete statements on its own website has better chances. There's no guarantee.
Does my website need to be in the AI's training data?
Not necessarily. Most systems that recommend companies access search results live when answering, rather than answering only from training. ChatGPT with search, Perplexity and Google's AI overview all work this way. That's why changes to a website take effect within weeks, not only at the next model training.
Which figures on GEO are reliable?
Few. The Princeton study cites improvements within its own test environment, which don't transfer directly to an SME website. Figures from providers promising to boost visibility by a certain percentage are usually not publicly verifiable. Only your own measurement with fixed questions over several months is reliable.