What is GEO (Generative Engine Optimization)? Definition, evidence and method

Pranjal Aggarwal and co-authors introduced the Generative Engine Optimization framework in a preprint submitted in November 2023 and later published at KDD 2024. Their work formalizes visibility in a generative engine, proposes the GEO-bench benchmark and tests several content transformations. The study reports gains of up to 40% in some experimental conditions and shows that effectiveness varies by domain.

The “40%” figure is often detached from its protocol. It does not mean that a page published on the web will automatically gain 40% more citations. The source is already present in the experimental context: the study mainly measures the influence or visibility of retrieved content, not its organic discovery in a changing commercial engine.

A critical review of 45 studies released as a preprint in July 2026 (opens in a new tab) separates activation, crawl, retrieval, context allocation, citation, factual uptake and user behavior. It concludes that none of the techniques reviewed has yet demonstrated a stable, longitudinal and cross-platform causal effect on organic discovery and downstream value.

Talking about a “ranking in ChatGPT” hides the mechanisms. Some answers use only model knowledge; others activate web search, consult an index or fetch a URL on demand. Candidate sources may then be filtered, reordered, truncated or replaced before synthesis. Finally, information that influenced the answer is not always displayed as a citation.

This chain explains why a technically sound site can remain absent, why a page may be cited for one wording but not a paraphrase, and why an editorial change may improve fidelity while reducing retrieval. Every stage requires a different signal and a different item of evidence.

Market vocabulary changes faster than practice. Creating a competing definition for every acronym mainly creates confusion. SEO remains the foundation for discovery, eligibility and authority. AEO emphasizes direct answers. GEO formalizes visibility in a generative answer and its measurement. LLMO is often used for language-model applications more broadly, including when no web search is activated. “AI SEO” is an umbrella term connecting conventional and generative search.

A GEO recommendation has value only when its evidence level remains visible. The register separates six statuses, requires a scope and connects every row to official documentation, a publication or an Edikka observation. A claim can be confirmed for Google and remain unknown for ChatGPT; its status must never exceed its scope.

Version 1.1.1 was reviewed on 2026-08-27. JSON and CSV distributions, the manifest, changelog and SHA-256 checksums are published under CC BY 4.0. A material correction triggers a new version, allowing the register to be cited, challenged and improved without erasing its history.

In May 2026, Google Search Central published a guide to optimizing for generative features. Its myth-busting section does not condemn every use of these techniques on the web; it says they are neither required nor advantageous for visibility in generative Google Search. The “Google Search” scope is essential: Google documentation does not describe how ChatGPT, Claude or Perplexity work.

A universal GEO policy quickly becomes false. Google Search relies on Googlebot and its index; OpenAI separates search, user actions and training; Anthropic publishes three crawlers; Perplexity separates its search crawler from its user fetcher; Microsoft now exposes citations in Bing Webmaster Tools. The first control is to name the exact surface and purpose.

The right order prevents a model, index, location or competitor change from being credited to a rewrite. It also prevents sacrificing a page that is already retrieved merely to optimize its form. Apply the following method to one pilot page before expanding it across the site.

On 22 July 2026, Google launched AI Overviews and AI Mode in France on mobile, desktop and in the Google app. This date marks a product-availability change for the French market; by itself it proves no traffic, citation or behavior change for a given website. Read the Google France announcement (opens in a new tab).

A young domain is not excluded from every citation. It can nevertheless face a greater absence risk when its content provides no distinctive evidence, recognized vocabulary or distribution. The two Edikka measurements answer different questions and must not be merged.

Edikka’s public robots.txt file allows search and grounding while reserving training for selected declared crawlers. The zero ChatGPT citations in MIA-FR J0 therefore cannot be attributed to an explicit OAI-SearchBot block in that file; server logs, CDN behavior and actual retrieval still require checking.

A serious GEO provider should show how it defines the corpus, separates platforms, archives answers, checks sources, handles variation and connects a change to a hypothesis. It should also explain what it cannot guarantee. An impressive dashboard does not compensate for an opaque method.

The register is not a certification and does not establish a universal algorithm. It exposes disagreements between documentation, research, observations and hypotheses. An official source describes its product; a research paper depends on its protocol; an Edikka observation depends on its corpus and date. These levels complement one another without becoming interchangeable.

The register’s first function is self-correction. An Edikka recommendation contradicted by new documentation must change status and leave a trace in the changelog. An obsolete register would be more misleading than no register at all: review is quarterly at minimum and occurs sooner after a material change by a cited operator.

Content version 1.1.0 applies that rule to itself: GEO-034 moves from “Unproven” to “Contradicted” after the addition of a longitudinal study showing that nearly 30% of domains cited in the observed AI Overviews were absent from the first-page results displayed in parallel. Version 1.1.1 adds DOI archiving and citation metadata only; none of the 39 claims changed.