Measuring GEO Visibility: Prompt Corpora Define the Answer Market

Olivier Martinez

cs.IRSep 6, 2026

Measuring GEO Visibility: Prompt Corpora Define the Answer Market

Authors: Olivier Martinez

Organizations: Sciences Po

Abstract

GEO (generative engine optimization) visibility scores aggregate source appearances, citations, or brand mentions in generated answers. The prompt corpus selects the situations evaluated, while weights determine their relative importance. Together they define an "answer market" that need not represent actual user demand. Prompt wording can alter retrieval, competing sources, and generated answers. Scoring then requires identifying the appearances, citations, or mentions of interest. If a language model performs this task, its instruction can change the score assigned to an unchanged answer. Our critical survey examines how these choices help define what a GEO score measures. It draws on research into whether indicators measure the intended phenomenon, total survey error, and information retrieval evaluation. The framework specifies situation annotation, prompt formulations, execution conditions, weights, and scoring rules. When weights are unknown or remain to be chosen, the framework reports sets of admissible scores. It distinguishes values compatible with data and assumptions about a target population (partial identification) from variation across weighting conventions (normative sensitivity). A citation alone does not establish a source's contribution. The article defines a comparison of answers generated with and without a source in a controlled documentary context, distinct from an intervention on the full engine with competing sources. The framework is supported by reproducible calculations. No new experiments are reported; its general empirical validity remains to be assessed.

Figures & tables

Function	Embedded decision	Consequence for interpretation
Specify the task	Definition, choice, justification, or list	Elicited behavior and success criterion
Select situations	Included needs, languages, people, and domains	Support over which the score is informative
Realize the need	Words, format, examples, and precision	Sensitivity to formulation
Condition the interaction	History, candidates, and assumed knowledge	Available information and trajectory
Make the output evaluable	Required sources, length, or format	Produced answer and ease of scoring

Table 1: Functions of prompts within the measurement protocol.

Stratum	System A	System B
Open-ended discovery	0.30	0.50
Named verification	0.90	0.60

Table 2: Fictional analytical example: success probabilities by stratum.

Study, status, and period	Scale examined	Contribution and inferential boundary
GEO; KDD 2024; version of 28 June 2024	10,000 queries, including 1,000 test queries	Supplied sources: conditional effect, not organic discovery
SAGEO; KDD 2026; version of 7 August 2026	2,700 test queries from nine datasets; 171,003 documents	Pipeline rerun; controlled distribution and initially retrieved target
Grossman; SIGIR 2026; 7–8 Dec. 2025	11,500 queries in the main collection	Mobile, Newark, two days; activation and sources on a benchmark, not a representative market
Sharma; 2026 preprint; Dec. 2025	112 products; 336 + 784 trials per API	Exact mention, nonequivalent conditions; illustration
Xu; 2026 preprint; Mar.–Apr. 2026	55,393 trending queries	Activation by observed categories, not a causal wording effect
Żatuchin; 2026 preprint; 29–30 Aug. 2026	234 usable runs, 12 rejected; few needs	Exploratory comparison, partial design, and nonconvergent control

Table 3: Scale, status, and inferential boundaries of visibility studies.

Published set	Queries	Activation $p_{\mathrm{act}}$	Bound on $p_{\mathrm{req}}$ , without mention data	$p_{\mathrm{req}}$ if $p_{\mathrm{cond}}=40\%$ (assumption)
Entire benchmark	11,500	65.6%	[0; 65.6%]	26.24%
ORCAS	5,000	51.5%	[0; 51.5%]	20.60%
Amazon Retail	500	17.4%	[0; 17.4%]	6.96%
ELI5	1,000	94.6%	[0; 94.6%]	37.84%

Table 4: Published activation and derived calculations; the final column is hypothetical. The total contains the subsets and is not an additional stratum.

Step	Required action and record	Decision(s)
1	Define the target and intervention. Specify the construct, unit, population or normative priority, period, metric, and denominator. Define the boundary of $S$ and the inputs of $H$ before choosing controls.	D1, D3
2	Document annotation and overlap. Retain accessible events or vignettes, the rules $A_{\eta}$ , disagreements, exclusions, and links between families. Identify missing cells; do not assign them an observed performance.	D1, D2
3	Fix or bound the weights. Declare the origins of $w$ and $v$ , constraints on $\Pi$ , and reasons for each scenario. Distinguish estimated empirical weights, conventions, and uncertainty classes; reserve control variants for final checks.	D2
4	Trace execution. Record the complete prompt, accessible system instructions, history, parameters, date and time, interface, model/version identifier when exposed, and request and response fingerprints. A stable identifier does not guarantee the absence of silent updates: interleave conditions, use sentinel controls, and segment collection if a break is suspected.	D1, D3
5	Code outputs and validate the judge. Apply the state scheme in Table 6 and the attribution coding in Section 8.2, drawing on Section 7.1. Retain the raw text and coding rule; audit outputs with independent judgment when necessary. Also version the judge’s instruction and check scoring differences on constant outputs.	D1
6	Calculate sets and their uncertainty. Produce cell-level scores, the score under the main convention, and extrema under $\Pi$ . Compare systems under the same weights. Separately add appropriate sampling or repetition uncertainty; publish the constraints active at the extrema. Multiverse analysis provides a transparency precedent for defensible analytical choices [ Steegen et al., 2016 ] .	D2, D4

Table 5: Seven implementation steps linked to the four decisions.

Observed state	Minimal coding	Consequence for the denominator
Activated module, usable answer	Activation = 1; separate mention and citation fields; support if assessable	Included in per-query rates and among activations
Valid non-activation	Activation = 0; presence in module = 0	Included per query; absent from the conditional rate
Explicit refusal	Refusal status; activation coded separately	Retain, then apply the target convention
Empty output received	Empty status; known or undetermined cause	No automatic recoding as non-activation
Missing collection or technical error	Missing status; reason and attempt	Neither observed brand absence nor silent exclusion

Table 6: Coding scheme for observed states.

Existing documentation field	Addition for a prompt corpus	Expected output
Motivation and uses	Estimand, unit, denominator, intervention	Target specification and defined contrast
Composition and collection	Families, variants, $A_{\eta}$ , overlap	Annotation scheme and missing cells
Processing and scoring	Origins of $w,v$ , class $\Pi$ , judge instruction	Weights, constraints, and cell-level scores
Maintenance and distribution	Engine state, fingerprints, competition, and changes	Version and differences between collections

Table 7: Extending Datasheets to document estimands and interventions.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Dimension	Statement for the present article
Objective and unit	Conceptual synthesis of measurement decisions; one work, versions grouped
Realized corpus	44 references; no exhaustive population of relevant works enumerated
Discovery	Initial dossier, titles, keywords, citation tracing, and extension during revision
Selection and annotation	Connection to construct, selection, formulation, execution, scoring, or transport; narrative coding without independent duplicate selection
Target and weights	Argumentative coverage; neither known inclusion probabilities nor an average of effects
Missing coverage	Non-English literatures and unidentified studies; unknown extent

Appendix

Table 8: Corpus card for this survey’s documentary selection.

Reference	Level actually used	Permitted use in this survey
Messick (1994)	ETS record and abstract	Intellectual lineage of validity; not the main methodological anchor
Frick et al. (2025)	PMLR record and abstract	Stated objective of prompt-conditional ranking
Zobel (1998)	Introduction, p. 307, indexed primary extraction	Pooling problem and stated general conclusion; no quantified effect
Manski (2003)	Introduction, pp. 1–2, publisher preview	Positioning of partial identification; no theorem from the book attributed without reading

Appendix

Table 9: Restricted readings and corresponding uses.

Generative Engine Optimization

Generative Engine Optimization studies how to make web content visible, cited, and accurately represented in AI-generated search answers.

See all 40 papers

Don't Measure Once: Measuring Visibility in AI Search (GEO)
Apr 8, 2026Julius Schulte, Malte Bleeker, Philipp KaufmannGenerative Engine Optimization Information Retrieval
Global Center for Entrepreneurship and Innovation University of St. Gallen 9000 St.Gallen, Switzerland · Marketing und Customer Insight University of St. Gallen 9000 St.Gallen, Switzerland
What Gets Cited: Competitive GEO in AI Answer Engines
May 25, 2026Rahul Vishwakarma, Shushant Kumar, Ratnesh JamidarGenerative Engine Optimization
Sprinklr · Gurugram, India · Dubai, UAE
GEO-Flag: Detecting and Measuring GEO-Optimized Web Content
Aug 17, 2026Junjie Chu, Ye Leng, Mingjie Li +3Generative Engine Optimization IR Evaluation
CISPA Helmholtz Center for Information Security · HPE · University of Waterloo

Apr 8, 2026J/K move · Enter open · S save

Don't Measure Once: Measuring Visibility in AI Search (GEO)

Julius Schulte, Malte Bleeker, Philipp Kaufmann

Global Center for Entrepreneurship and Innovation University of St. Gallen 9000 St.Gallen, Switzerland · Marketing und Customer Insight University of St. Gallen 9000 St.Gallen, Switzerland

Generative Engine Optimization Information Retrieval Generative AI Evaluation

Open paper AskPDF arXiv

As large language model-based chat systems become increasingly widely used, generative engine optimization (GEO) has emerged as an important problem for information access and retrieval. In classical search engines, results are comparatively transparent and stable: a single query often provides a representative snapshot of where a page or brand appears relative to competitors. The inherent probabilistic nature of AI search changes this paradigm. Answers can vary across runs, prompts, and time, making one-off observations unreliable. Drawing on empirical studies, our findings underscore the need for repeated measurements to assess a brand's GEO performance and to characterize visibility as a distribution rather than a single-point outcome.

Abstract

Generative Engine Optimization

Explore similar work

Don't Measure Once: Measuring Visibility in AI Search (GEO)

What Gets Cited: Competitive GEO in AI Answer Engines

GEO-Flag: Detecting and Measuring GEO-Optimized Web Content