The Princeton GEO Study: Methodology, Results and Critique

The Princeton GEO study is the starting point for any serious work on Generative Engine Optimization. If you work in DACH B2B and want to understand why statistics, source citations, and quotations suddenly count as optimization levers in AI answers, you cannot skip this paper. Released as a preprint in 2023 (published at KDD 2024), it delivered the first measurable, reproducible evidence that content can be deliberately designed so that generative search systems cite it more often and more prominently. This article puts the study in context, explains its methodology, interprets the widely cited core result correctly, and names the limitations you need to keep in mind when applying it to today's systems and the German-speaking market.

What the Princeton GEO study is and why it still matters

The paper titled "GEO: Generative Engine Optimization" was written by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande. It was published at the 2024 ACM SIGKDD Conference (KDD 2024) and is freely available as arXiv preprint 2311.09735. The authors coined the term Generative Engine Optimization and defined it as a discipline of its own alongside classic search engine optimization.

Its relevance stems from the situation it addresses. Classic SEO optimizes for rankings, meaning positions in a list of results. Generative search systems such as Google AI Overviews, ChatGPT Search, or Perplexity instead deliver a synthesized answer and point only to selected sources. The Princeton study was the first paper to measure this new reality in a controlled experiment. It shifted the question from "How do I rank higher?" to "How do I get included in the generated answer?" This shift is exactly what makes the paper the founding document of a discipline whose tactics now show up in every GEO guide.

One point matters for context: the study is academic, not commercial. Its findings come from a peer-reviewed experiment, not from a tool vendor's marketing. That gives it a credibility advantage you cannot take for granted with many later industry studies.

The methodology: GEO-bench and the new visibility metrics

At the core of the study is a purpose-built benchmark framework called GEO-bench. It comprises 10,000 search queries, split into 8,000 training, 1,000 validation, and 1,000 test queries, drawn from nine different data sources across 25 domains such as art, health, or games. To make the dataset reflect real-world search, the authors kept the typical query distribution: 80 percent informational, 10 percent each transactional and navigational queries. This breadth is crucial because it prevents a result from holding true for only a single niche.

The second methodological contribution is two new metrics that make visibility in a generated answer quantifiable in the first place. Both measure something different from a classic ranking:

Position-Adjusted Word Count: This metric counts how many words of an AI answer trace back to a specific source and weights them by their position in the answer. A source that is used early and extensively gets a higher score than one that appears only on the margins. In other words, it measures the extent and prominence of the uptake.

Subjective Impression: This metric uses a language model as the evaluator and assesses the perceived relevance, influence, and usefulness of a cited source within the answer. It captures qualitative aspects that a plain word count cannot reflect.

The evaluation itself ran on a generative system built on a then-current GPT model that used search results as its source base. For each test query, the content of a target source was revised according to one GEO method, and the researchers then measured whether, and by how much, the visibility of that source in the generated answer changed.

The nine GEO methods tested and their effect

The study tested nine concrete content manipulations against an unoptimized baseline. The most important ones fall into three groups:

Authority signals (the winners): These include Statistics Addition (adding relevant numbers and statistics), Quotation Addition (adding relevant quotes), and Cite Sources (adding citations that point to credible evidence). The authors describe these three methods as the most effective in the study.

Linguistic optimization: This group includes Fluency Optimization (smoother, clearer wording), Authoritative (a more authoritative tone), and Technical Terms (adding specialist terminology). On Position-Adjusted Word Count, Fluency Optimization reached 24.7 points (19.3 without optimization), putting it roughly level with Cite Sources (24.6). Authoritative and Technical Terms had a moderate effect that depended heavily on the subject area.

Ineffective or harmful approaches: Keyword stuffing, the mechanical repetition of search terms, was among the weakest approaches and could even reduce visibility compared to the baseline. In the real-world test on Perplexity, keyword stuffing scored about 10 percent lower on Position-Adjusted Word Count than the unoptimized baseline. This is a key finding: classic SEO spam tactics do not carry over to generative systems.

The most valuable finding for practice lies in the combination. The authors showed that combining linguistic optimization with added statistics outperforms any individual method. For DACH B2B, this means: raise evidence density and readability together instead of picking a single tactic.

Interpreting the core result "up to 40 percent" correctly

The most frequently cited number from the study is the headline claim that visibility in generative answers can be increased by up to 40 percent. This number is often misunderstood in practice, so it is worth putting it in precise context.

First, it is a maximum, not an average. Specifically, according to the paper, the three methods the authors highlight (Cite Sources, Quotation Addition, and Statistics Addition) achieved a relative improvement of 30 to 40 percent on the Position-Adjusted Word Count metric (recalculated from Table 1 of the paper: roughly 27 percent for Cite Sources up to 41 percent for Quotation Addition). So it represents the top of what is possible under ideal conditions, not a guarantee for every piece of content.

Second, it is a relative improvement, not an absolute one. It describes the gain over the unoptimized baseline of the same source within the same answer context. If you are already cited prominently, you have less room to grow.

Third, and strategically this is the most important point, lower-ranked sources benefited the most. The study describes an equalizer effect: content that was barely visible without optimization gained disproportionately from GEO methods. For small and mid-sized providers in the DACH region, this is the real good news. GEO gives underdogs in particular a realistic chance of being mentioned instead of only making established market leaders even more visible.

Domain specificity: why the same method has different effects

An often overlooked finding of the study is its domain dependency. Across the 25 domains covered, no method worked the same way everywhere. In some subject areas, statistics lifted visibility the most; in others, it was quotations or a more authoritative writing style.

For B2B practitioners, this translates into a concrete instruction: there is no universal GEO formula. A technical industrial product, a legal advisory service, and a software-as-a-service offering need different priorities. Where trust comes from hard data, statistics density pays off. Where it comes from expertise and reputation, attributed quotations pay off. The study thus provides a methodology for testing, not a recipe. If you take GEO seriously, you measure the effect per topic area instead of laying a blanket strategy over all your content.

Criticism and limitations of the study

As important as the Princeton study is, its explanatory power has clear limits that you need to know for responsible use.

Synthetic and partly opaque evaluation: Visibility was measured by machine, partly with a language model as the evaluator. The generative system and its source selection remain a black box. It is not entirely clear why one source was preferred over another.

Model generation 2023/2024: The experiments ran on the models and search systems of that time. Today's engines such as Google AI Overviews, Google AI Mode, Perplexity, or Claude work with newer architectures, different retrieval, and in some cases completely different source logic. The specific percentages therefore do not carry over one to one to 2026.

Zero-sum setup: In the experiment, only a few sources competed for visibility within one answer. In such a closed setup, relative gains turn out larger than in the open real world, where a system chooses from millions of pages. Read the improvement figures as a directional indicator, not as an exact forecast for your own website.

Generalizability: The study tested generative synthesis, not every product in use today. That the mechanism works in principle is considered well established. That the exact numbers apply to a specific system today is an assumption, not a proven fact.

From study to practice: operationalizing the findings

The research results translate into four practical levers that you can put to work right away in DACH B2B:

Front-loading: The most important answer belongs at the beginning of a section. Generative systems favor early, clearly worded statements. This matches the finding that prominent placement is weighted more heavily in the Position-Adjusted Word Count logic.

Answer islands: Organize content into standalone, semantically complete passages that fully answer one question. Self-contained blocks like these are easier for a system to lift into an AI answer as a whole.

Fact and statistics density: Since Statistics Addition was among the strongest methods, concrete, sourced data density is a direct GEO lever. Your own proprietary numbers are especially valuable because they cannot simply be swapped for someone else's.

Schema and structure: Structured data and clean heading hierarchies help systems classify content correctly. They are the technical operationalization of the content findings.

Current state of research since Princeton

The Princeton study was the beginning, not the final word. Several follow-up studies confirm its general direction and put individual assumptions into perspective.

Extended frameworks have expanded the original catalog of methods and added structural features such as readability, formatting, and semantic clarity. The thrust remains the same: authority signals and good structure increase the likelihood of being cited.

Research into what drives AI visibility off the page is particularly revealing. An analysis of 1,000 domains found that domain authority is clearly related to AI mentions (Pearson 0.65, Spearman 0.57). Notably, nofollow links show almost the same relationship with AI visibility as follow links. In practice, this shifts the focus from raw link power toward reputation and mentions overall. Brand presence and backlinks are more closely related in the AI world than classic SEO suggests.

This follow-up research confirms the core of the Princeton study and adds an important dimension: on-page optimization alone is not enough; it has to be backed by external authority.

llms.txt in light of the evidence

An idea circulating in the GEO space is to provide language models with a curated content map in a file called llms.txt. For an evidence-based assessment, one point matters: this idea is not part of the Princeton study, and the study does not support it.

The state of play is sobering. Despite growing adoption, no major AI provider has officially confirmed that it uses llms.txt for answer generation. Industry analyses have so far found no reliable link between having an llms.txt file and higher citation rates. The gap between attention and actual effect is wide here. If you approach GEO seriously, treat llms.txt as experimental and low priority, and put your resources into proven levers such as evidence density, structure, and reputation first.

Transfer to DACH and Austria

The Princeton study is English-language and US-centric. Whether its findings hold in German-speaking B2B has a nuanced answer.

The mechanism is language-independent. Generative systems prefer sourced, well-structured, and authoritative content because of how answer synthesis works, not because of the English language. To that extent, the qualitative findings transfer. The exact percentages do not, because they are tied to the systems tested and to English-language datasets.

Market relevance in Austria is growing rapidly. According to Statistics Austria, 29.9 percent of Austrian companies with 10 or more employees were already using AI in 2025, up from 20.3 percent in 2024 and 10.8 percent in 2023. That is almost 10 percentage points more within one year. The years are not fully comparable, though, because since 2025 the survey has asked about eight AI technologies instead of seven. Without the new category, the 2025 figure would be 28.9 percent. Especially relevant for GEO: 51 percent of AI-using companies use AI for language generation. The technology behind generative search has already arrived in Austria's small and mid-sized businesses.

Two practical points are decisive for the DACH region. First, generative search measurably changes click behavior. According to an international Ahrefs analysis, the presence of an AI Overview now correlates with a 58 percent lower average click-through rate, compared with 34.5 percent in the earlier study from April 2025. Visibility in the answer is therefore becoming more important than the list position itself. Second, the remaining traffic is more valuable. An international Semrush study found that an AI search visitor is 4.4 times as valuable as a classic organic search visitor, based on conversion rate. For DACH B2B providers with long, consultation-heavy sales cycles, this combination clearly shifts priority toward GEO.

Further reading

The Princeton GEO study provides the scientific foundation, but not a ready-made strategy. For operational implementation in DACH B2B, connect the findings with the related topics in this knowledge base: the role of statistics and data as one of the strongest single levers, the relationship between brand mentions and backlinks, content structuring through front-loading and answer islands, and ongoing performance measurement with suitable GEO metrics. You get the most out of the study if you use it as a hypothesis generator and test its findings in your own topic area instead of reading the percentages as a promise.