← Back to GEO Academy
GEO basics

Reading 'GEO: Generative Engine Optimization' (I): Define Visibility Before Optimization

The first reading of the KDD 2024 paper GEO: Generative Engine Optimization explains why AI search is a retrieval, generation, and citation problem - not a conventional results-page ranking problem.

Published 07/25/2026 8 min read
GEOGenerative Engine OptimizationAI search visibilityGEO research

Reading 'GEO: Generative Engine Optimization' (I): Define Visibility Before Optimization

It is easy to reduce GEO to "SEO for AI search." The KDD 2024 paper *GEO: Generative Engine Optimization*, by Pranjal Aggarwal, Vishvak Murahari, and colleagues, starts from a more precise question: when AI search compresses multiple sources into one answer, how should a content creator define being seen?

This first reading does not begin with rewriting tactics. Its value is the research frame: define the system, define visibility, and only then discuss optimization.

Why the paper defines a generative engine first

The paper abstracts products such as BingChat, Google SGE, and Perplexity.ai as a generative engine. Rather than returning only blue links, it retrieves sources, uses a generative model to combine them, and delivers an answer with citations or source grounding.

Traditional SEO commonly asks, "What position is my page in?" A generative engine shows a synthesized answer instead. The same source may appear early, in the middle, or at the end; it may support a core conclusion or merely provide background. Rank alone cannot describe those differences.

Visibility is more than a brand mention

The paper treats a website's visibility, also called impression, in a generated answer as GEO's central objective. A source's role has at least three dimensions:

  1. How much of the answer comes from that source.
  2. Whether that contribution appears early enough to be noticed.
  3. Whether the source materially supports the answer's judgment.

This is why a screenshot proving that a brand was mentioned is not enough. A late citation for a minor fact has a different commercial meaning from an official source used to support the opening recommendation.

Two families of visibility metrics

The paper proposes objective measures. Word Count estimates how much text an attributed source contributes; Position-Adjusted Word Count adds the effect of placement, so the same 30 words generally carry more attention near the start than at the end.

It also proposes subjective measures inspired by G-Eval, using a model to assess relevance, influence, uniqueness, position awareness, quantity awareness, click likelihood, and information diversity. The method has limits, but its lesson is valuable: a GEO report should not count words mechanically without asking whether the source changed the user's understanding.

What this changes for a business

GEO is not an instruction to force an AI to recommend a brand. The paper studies content visibility under black-box conditions. A business can make official pages, documentation, cases, data, and third-party evidence easier to retrieve, cite, and integrate; it cannot control a model's output.

Nor is GEO a replacement for SEO. The experiments use sources drawn from search results. SEO helps a page enter the source pool; GEO also considers how the page is represented once it enters a generated answer.

The paper's GEO-bench further reinforces the need for a fixed question set. If prompts change every time, no team can distinguish content effects from platform variation, source changes, or different wording.

From research definition to an operating baseline

Before producing more content, measure the basic facts: on which AI platforms does the brand appear, how much explanatory space does it receive, where does it appear, which sources are used, and whether competitors receive more of the answer?

GEO Radar at https://www.georadar.top can help establish that measurement baseline through fixed question sets, multi-platform review, brand mentions, recommendation position, and competitor co-occurrence. The first step in GEO is not publishing more pages. It is making "Can AI see us?" a repeatable set of measurements.

Sources for this article