Abstract

Generative engine optimization requires a different measurement system from traditional search. A ranking position describes where a page appears. A generated answer describes how a brand is represented, whether it is recommended, what evidence supports the answer, and which alternatives appear beside it. This paper defines a measurement framework built around prompt portfolios and answer-level outcomes.

The unit of analysis

A single prompt is too unstable and too narrow to represent market visibility. The proposed unit is a prompt portfolio: a structured set of questions grouped by audience, intent, use case, category, and stage of decision. Each prompt receives a weight based on commercial relevance rather than estimated search volume alone.

Four dimensions of representation

The framework tracks presence, position, description, and evidence. Presence asks whether the brand appears. Position records whether it is mentioned first, included later, or omitted. Description evaluates which attributes and claims are attached to the brand. Evidence identifies the sources cited or implied in the answer.

Recommendation share

Recommendation share measures how often a brand appears inside the recommended set for a defined prompt portfolio. First-recommendation share is stricter. It measures how often the brand is presented as the leading option. Both metrics should be reported by audience and use case to avoid hiding category-specific weaknesses.

Citation coverage

Citation coverage compares the claims a brand wants to own with the sources available to support them. A high visibility score with weak citation coverage can be fragile. A strong citation network with low visibility may indicate a retrieval or content-structure problem.

Operating cadence

GEO measurement is most useful when repeated on a stable schedule with versioned prompts, recorded outputs, and clear change logs. Teams should separate real movement from normal answer variation by looking for patterns across the portfolio rather than reacting to one response.

Limitations

Prompt wording, location, personalization, and model updates can change results. Measurement should therefore emphasize direction, coverage, and repeated patterns. Exact scores should not be treated as permanent market facts.