GEO: Generative Engine Optimization (Aggarwal et al. 2024)

1. What this paper established

GEO: Generative Engine Optimization (Aggarwal et al., KDD 2024) introduced the academic term GEO.

The paper turned the previously informal goal of getting content into an AI-generated answer into a measurable optimization problem supported by a public benchmark.

The paper uses GEO in a narrow, benchmark-specific sense: content rewrites evaluated in a fixed test harness. Practitioners later expanded the term to describe a broader discipline. As a result, many claims made about GEO in 2026 concern practices that this paper did not test.

2. The problem the paper formalizes

The paper begins by describing a structural change in search. A conventional results page presents a ranked list of links, while a generative engine produces a synthesized answer. In that setting, visibility and ranking are no longer equivalent.

The paper reduces the problem to three elements:

• The generative engine is a black box. Content creators cannot change the model; they can change only the web content supplied to it.
• The goal is to maximize a source’s visibility within the synthesized answer, rather than its position in a list of links.
• Success depends on how the answer uses the source text. The paper therefore introduces a visibility metric rather than reusing search rank.

The answer loop and Gao et al.’s 2023 survey explain the retrieval and grounding processes that determine which sources an engine selects and uses. Aggarwal et al. assume those processes are already in place and focus on how changes to source content affect visibility.

3. Methodology: the GEO framework and GEO-bench

The researchers applied nine content-optimization methods to candidate web sources:

GEO-bench contains:

• 10,000 queries are divided into 8,000 training, 1,000 validation, and 1,000 test queries.
• The queries come from 9 datasets: MS MARCO, ORCAS-I, Natural Questions, AllSouls, LIMA, Davinci-Debate, Perplexity.ai Discover, ELI5, and a GPT-4-generated set.
• They span 25 domains, including Arts, Health, and Games, organized into seven categories.
• For each query, the top five Google results form the candidate source set.

The researchers tested two engines. Their internal generative engine prompted GPT-3.5-turbo with the top five Google results. They also used the live Perplexity.ai service to check whether the findings transferred to a deployed engine.

4. The paper’s impression metrics

The paper’s most widely reused contribution is its approach to measuring visibility. Rather than treating citation and mention as binary outcomes, the metrics account for how much of a source appears in an answer and where that material appears.

The paper proposes two measures:

Position-Adjusted Word Count (Imp_pwc):
  sum over cited sentences s of  |s| · e^(-pos(s)/|S|)
  divided by total response word count

= cited-sentence word count, exponentially discounted
  according to how late each sentence appears in the answer.

Subjective Impression:
  a composite score from GPT-3.5 across 7 subdimensions:
  relevance · influence · uniqueness · subjective position ·
  subjective count · click-likelihood · diversity

A yes-or-no citation count cannot distinguish a passing reference from a source that shapes much of an answer. These metrics capture that difference in prominence, an important aspect of citability. Many later vendor metrics and GEO measurement frameworks use the same position-weighted approach.

5. Detailed results

The main results and their limits are:

Taken together, the failed keyword-stuffing test and the smaller result on Perplexity.ai support two practical conclusions. Substantive additions performed better than keyword manipulation, and the reported percentages should be treated as upper bounds rather than expected gains.

6. Critical assessment

Three contributions remain important:

1. Naming and framing. The paper defined visibility in an AI-generated answer as a black-box optimization problem and gave the field a shared term for that problem.
2. The impression metrics. Position-adjusted, continuous measures of visibility provide more information than a binary citation count and have influenced later measurement systems.
3. A public benchmark. GEO-bench made the paper’s claims testable and opened them to independent criticism.

The conclusions also have four important limits:

1. External validity. The study tested an internal GPT-3.5 harness and the 2023–24 version of Perplexity.ai. ChatGPT Search, Gemini, and AI Overviews use different retrieval and synthesis systems, so the paper does not establish that the same effects apply to them.
2. Benchmark drift. Both GEO-bench’s corpus and the tested engines age. The 40% result belongs to a 2024 snapshot and should not be treated as a constant across time.
3. Later evidence is less optimistic. Puerto et al.’s C-SEO Bench, published in the NeurIPS 2025 Datasets and Benchmarks Track, found that many conversational SEO rewrites became ineffective or counterproductive when several parties optimized for the same engine. The single-actor gain in the GEO paper is an upper bound, not a competitive equilibrium.
4. The headline is easy to misread. Practitioner coverage often turns “up to 40%” into “about 40%.” The paper reports a per-method, per-domain maximum that fell to about 22% on a live engine and may decline further under competition and trust filtering, including E-E-A-T pressures.

Assessment. The paper supports adding source citations, statistics, and quotations instead of manipulating keywords. Its specific percentages are not reliable planning assumptions.

7. Reproducibility

The following resources were verified as of May 17, 2026:

The reproducibility rating is code-and-data. Both the method implementation and GEO-bench are public, so researchers can reproduce the main experiments independently. That level of access is uncommon among papers in this field.

8. Practical implications

For practitioners, the findings suggest four guidelines:

• Make substantive changes. Add credible citations, concrete statistics, and relevant quotations instead of relying on keyword manipulation.
• Measure prominence, not citation alone. A source can play a minor or substantial role in an answer even when both cases count as one citation. Include that distinction in AI citation tracking.
• Do not reuse the reported percentages across engines or domains. The same rewrite can perform differently on ChatGPT Search and Perplexity.ai.
• Expect competition to reduce gains for any one publisher. A method that helps one publisher may lose its advantage when competitors adopt it, as C-SEO Bench found.

9. Further reading

• Gao et al. 2023: RAG: A Survey explains the retrieval and grounding mechanisms assumed by the GEO paper.
• C-SEO Bench (Puerto et al. 2025) provides a less optimistic test of conversational SEO rewrites.
• Generative Engine Optimization explains how the broader practitioner discipline developed from the paper’s narrower definition.