The paper that defined GEO: visibility metrics, 9 methods

GEO (Generative Engine Optimization) is a framework for optimizing content to raise the chance that generative search engines such as Perplexity and Google AI Overviews cite it when they build an answer. The paper by Aggarwal et al., presented at KDD 2024, is the original GEO paper: it first defined the concept, its measurement metrics, and a benchmark ^[1].

Instead of listing search results, a generative engine reads several sources, merges them into one answer, and shows it directly. So visibility is now decided not by rank but by whether you are cited inside that answer. The problem is that content creators have little control over when and how their writing appears in the answer. GEO is an attempt to give creators a means of optimization in this new landscape ^[1]. (We unpack the basics of GEO further in What is GEO.)

What a generative engine is, and why SEO isn't enough

A generative engine works in roughly four steps. It rewrites the query into several queries, gathers sources through a search engine, condenses them with a summarization model, and produces a final answer with citations using a response model.

Figure 1. The structure of a generative engine: query rewriting, retrieval, summarization, and response generation with citations.

Traditional SEO was about climbing the rankings by fitting the search ranking algorithm. But a generative engine does not stop at keyword matching: a language model understands sources and queries more deeply and synthesizes an answer. On top of that, citations are embedded inside the answer text at different lengths and positions. So "what rank are we" alone cannot measure visibility.

Figure 2. A search engine keeps rank and original text intact, but a generative engine mixes citations into the answer, so measurement becomes multidimensional.

How to measure visibility

The paper's most fundamental contribution is defining impression (visibility) metrics for generative engines. The simplest, "word count," divides the word count of the sentences carrying a given citation $c_i$ by the word count of the whole response.

Here $S_{c_i}$ is the set of sentences citing $c_i$, $S_r$ is all sentences in the response, and $|s|$ is the word count of sentence $s$. The more of the answer a citation takes up, the larger the role that source played.

But word count alone counts a citation the same whether it appears early or late. Since citations placed earlier get read more, the paper also uses position-adjusted word count, which decays exponentially with position ^[4].

Beyond that, subjective aspects such as click probability, query relevance, a citation's influence, and uniqueness are not captured by word count. So the paper adds a subjective impression measured with G-Eval along 7 dimensions (relevance, influence, uniqueness, diversity, follow-up, subjective position, and citation count) ^[3]. A method's effect is compared as the relative improvement of the modified response $r^{\prime }$ over the original response $r$.

Nine GEO methods

The paper defines 9 engine-agnostic methods that can be applied to web content and evaluates them on GEO-Bench (10K queries, 9 sources, 25 domains) ^[1][2].

Table 1. The 9 GEO methods evaluated in the paper.

The key point is that, apart from numbers 3, 4, and 5, most of these change only how existing content is presented, without adding much new information. In other words, you can raise visibility without making up facts.

Figure 3. GEO before and after: a low-visibility source is cited earlier and at greater length in the generative response after optimization.

What worked

The results are clear. Cite Sources, Quotation Addition, and Statistics Addition were the strongest, and the best method improved on the baseline by 41% in position-adjusted word count and 28% in subjective impression. Traditional SEO keyword stuffing, by contrast, had almost no effect or even made things worse ^[1].

Table 2. Absolute impression of GEO methods (excerpt). Citation and statistics methods raise both metrics substantially ^[1].

Methods work better in combination. Using fluency improvement and statistics addition together beats any single method, and the best-performing combination differs by domain.

Figure 4. Heatmap of relative visibility improvement by method combination. Fluency + statistics is the strongest.

The lower the rank, the bigger the opportunity

More interesting still, the lower a page ranks in search, the larger the GEO effect. Adding source citations to the page ranked 5th in search raised its visibility in the generative response by a full 115%. Small creators who lose out on backlinks and domain authority can earn citations through content design alone, so GEO in effect levels the playing field ^[1].

Table 3. Relative visibility improvement (%) by search rank. Gains explode at lower ranks ^[1].

Different methods work for different domains

No single method works for every topic. The paper also analyzed which methods are strong in each query domain. An authoritative tone works especially well for debate-style questions, and source citations for factual questions ^[1].

Table 4. Domains where each method is especially effective ^[1]. Choose methods to fit your target topic.

It worked on a deployed engine too

Going beyond the synthetic benchmark, the paper also validated on the deployed Perplexity.ai. Source, statistics, and quote methods raised subjective impression by up to 37%, while keyword stuffing again came in below the baseline ^[1]. That signals lab results carry over to a real commercial engine.

Limits of this approach

The paper also states its own limits ^[1].

• Engines change: as generative engines evolve, the effect of these methods may shift, just as with SEO.
• Black box: the approach adjusts only text without knowing the engine's internals, so the causal reason why it works is not fully established.
• Effect on search ranking not measured: the effect of GEO edits on traditional search rankings was not measured separately.
• Tag noise: domain and intent tags were assigned automatically, so some errors may be mixed in.

The TRAIL view: citation is something you measure and design

The two pillars this original paper set up are the backbone of what we do today: redefining visibility so it can be measured, and optimizing it through content design. TRAIL Search tracks this visibility and share of voice (SoV) by engine, and our practical recommendations match the paper's conclusion: don't repeat keywords, build dense evidence with sources, statistics, and expert quotes, and polish sentences for fluency ^[1].

If this paper opened the door, later research is moving toward choosing methods automatically. FeatGEO, which optimizes at the feature level instead of editing the text directly, is one example ^[5]. Read alongside research on RAG source attribution, which traces which documents built an answer, measurement and design fit together as a pair. In the end, your place in AI answers grows from the moment you turn what you include and how you present it into a measurable problem.

You can continue with SAGEO Arena: edit only the body and you drop out of search, A reward ceiling for RL alignment, computed before training, and Do AI search sources change with the user's location?.