The Princeton Study on GEO: The Scientific Foundation

The scientific basis for Generative Engine Optimization comes from a study published in 2023 and presented at the KDD conference in Barcelona in 2024. Researchers from Princeton University, the Georgia Institute of Technology, and the Allen Institute for AI systematically measured for the first time how content levers can specifically increase seo-glossary/visibility/">visibility in generative AI responses – including robust effect sizes per measure.

Study Design and the GEO-Bench Dataset

The research team led by Pranjal Aggarwal, Vidhisha Balachandran, and Vishvak Murahari developed a benchmark called GEO-Bench. This includes around 10,000 search queries from nine different domains – ranging from medical topics to legal questions to product comparisons. For each query, responses from generative engines were analyzed, and the share of different sources in the citations was measured. The variety of questions was a deliberate study design element, as it allowed verification of whether the identified levers work across industries or are merely artifacts of specific fields.

Subsequently, the researchers deliberately varied specific properties of the source texts and measured how the citation frequency changed. This resulted in clean, causally interpretable effect sizes for each individual GEO measure. These numbers are now the gold standard against which reputable providers should measure themselves. The replicability of the results has since been confirmed multiple times by independent follow-up studies, underscoring the methodological rigor of the original work.

The Key Effect Sizes from the Study

The Princeton study identifies a handful of levers that work in nearly all domains. Measures that increase the credibility and verifiability of the statement are particularly strong. The following values come directly from the original publication and describe the average increase in visibility in the AI response.

• Cite Sources: +30 to 40% more citations through clear source references
• Statistics: +30% through embedded, concrete numbers and data
• Quotation: +30% through direct quotes from experts or original sources
• Fluency Optimization: +22% through clean, readable, well-structured language
• Technical Terms: +21% through precise terminology instead of generalities

These values are not marketing promises but averages over thousands of queries. From our consulting experience, we know: In highly competitive B2B domains, the effects often lie at the upper end of the specified ranges, as many competitors have not yet implemented GEO-specific optimizations. Another important point: The levers work cumulatively. Those who implement three or four of the measures simultaneously often achieve increases that significantly exceed the sum of the individual effects, as the language model disproportionately weights sources with multiple trust signals.

The Most Surprising Finding: 115 Percent for Underdogs

One of the most exciting results of the study concerns the question of who benefits most from GEO. While the top three domains in classic search results only see moderate gains through GEO, sites ranked fifth and below can increase their visibility in AI responses by up to 115.1 percent. Generative engines explicitly reward sources that are compelling in content but have previously been overlooked in classic ranking competition.

For medium-sized companies and niche providers, this is a strategic goldmine: Those who consistently optimize their content according to the Princeton levers and simultaneously build a healthy backlink profile can appear disproportionately in ChatGPT, Perplexity, and Google AI Mode even without a top-ranking position. Linkbuilding remains the foundation: Without external trust signals, a page does not even make it into the retrieval pool from which the models cite. This combination of Princeton-compliant optimization and targeted link building explains why smaller, high-quality brands suddenly appear in many niches ahead of established large publishers in AI responses.

In our client projects, we have reproduced this effect multiple times. A manufacturer from the industrial segment, who previously ranked between positions 8 and 12 on Google, was able to triple his citation share in Perplexity within four months through systematic GEO optimization and complementary link building. The classic ranking simultaneously improved to a stable position 4 – clear evidence that both disciplines reinforce rather than displace each other. Such results do not happen by chance but through a consistent integration of content optimization, technical maintenance, and targeted off-page work that binds all three levers to the same thematic axis.

The Princeton study is to GEO what Andrey Lipattsev's SEO statements from 2016 were to the backlink discussion: a scientifically hardened anchor that the entire industry can orient itself around.

What the Data Means for Your Practice

The study provides a clear prioritization. Those starting with GEO should first address the levers with the highest effect sizes: adding sources, incorporating statistics, embedding quotes. These measures can be quickly implemented in most content management systems and deliver measurable results within weeks. It is important that the optimizations are authentically embedded in the text and do not appear as retrofitted building blocks – language models reliably recognize unnatural structures and tend to downweight them.

At the same time, the backlink profile should be further developed, as the best content is of no use without trust signals from the open web. In our client projects, we therefore consistently combine the Princeton levers with thematically strong link building – this is the most reliable way to appear permanently in the responses of relevant AI systems. Those who implement both simultaneously achieve the first measurable citation increases within a quarter and lay the foundation for sustainable visibility.

Conclusion

The Princeton study has taken GEO out of the realm of speculation and transformed it into an evidence-based discipline. Those who take the effect sizes published there seriously have a clear roadmap: sources, statistics, quotes, fluency, technical language – supported by a strong backlink profile. This is the scientifically hardened basis on which successful GEO strategies will be built in 2026. The more brands understand and consistently apply these levers, the more important the additional off-page advantage will become – another reason to start building a robust link profile now.