The Logic-First Research Methodology: An Evidentiary Standard for Generative Engine Optimization Claims

1. Introduction

The generative engine optimisation (GEO) field is in its infancy. As with any nascent discipline, the absence of established methodological standards creates conditions for over-claiming -- where marketing assertions are presented as research findings, correlations are described as causation, and single observations are generalised as universal principles.

The traditional SEO industry provides a cautionary precedent. Decades of "studies" comparing websites that differ on dozens of variables have produced a folklore of "ranking factors" that conflate correlation with causation, confuse necessary with sufficient conditions, and generalise from single observations to universal claims. The GEO field risks repeating these errors unless methodological standards are established early.

The Logic-First Research Methodology presented here is designed to prevent these specific failure modes. It does not prescribe what to study or how to collect data, but rather establishes the logical and evidentiary standards that findings must meet before they can be stated at a given confidence level.

2. The Evidence Hierarchy

The methodology defines seven evidence levels, adapted from medical research hierarchies for the specific context of AI behaviour research:

2.1 The Language-to-Level Mapping

The most common methodological error in GEO literature is using causal language ("X causes higher rankings") for observational data. The language-to-level mapping makes this error structurally impossible: the word "causes" is prohibited below Level 4, and even at Level 4 must be scoped to "under these conditions." This single constraint, if adopted, would eliminate the majority of over-claiming in the field.

3. The Seven Logic Gates

Every finding in the SIGI research programme must pass seven sequential validation gates. Failure at any gate requires the finding to be downgraded or restated. The gates are detailed in SIGI-2026-067; here we provide an overview:

4. The Confidence Statement Template

Every finding in the SIGI programme carries a standardised confidence statement (see SIGI-2026-070 for the full template). The template requires explicit declaration of: evidence level, method used, confounds identified, gates passed/failed, overall confidence, upgrade path, and the strongest permitted claim. The "upgrade path" field is particularly important: it forces researchers to identify exactly what experiment would raise the confidence level, preventing stagnation at lower evidence levels.

5. Three Summary Rules

The methodology is encapsulated in three rules that govern all research output:

1. State what you know at the level you know it. If your data is correlational, use correlational language. If your experiment is single-model, state it. Never round up.
2. When in doubt, downgrade. If you are uncertain whether a finding is Level 3 or Level 4, classify it as Level 3. Under-claiming is a minor cost; over-claiming damages credibility and misleads practitioners.
3. Design the experiment that would prove you wrong. Every finding should be accompanied by a description of what evidence would falsify it. Unfalsifiable claims are not scientific claims.

6. Application to the SIGI Research Programme

The Logic-First methodology has been applied across the SIGI 100-paper research programme. Category A papers (controlled experiments based on 10 probes, 129 data points) are classified at Level 4. Category B-D papers (observational studies, introspective analyses, and AI platform behaviour studies) are classified at Level 2-3. Category E papers (including this one) present the methodology itself.

The discipline imposed by the methodology has had a concrete effect: findings that would conventionally be stated as causal ("entity density increases citation rates") are instead stated at their supported level ("under these controlled conditions, higher entity density is associated with more favourable LLM evaluation"). This precision reduces the headline impact but increases the durability and credibility of the findings.

7. Limitations

• Framework, not empirical finding: This paper presents a methodological standard, not a testable hypothesis. Its value depends on adoption and consistent application.
• Adapted from other fields: The evidence hierarchy and logic gates are adapted from medical and social science research. Whether these adaptations are optimal for AI behaviour research is itself an empirical question.
• Compliance burden: The methodology imposes significant overhead on research output. Whether practitioners will adopt it absent institutional requirements is uncertain.

8. Conclusions

We present a research methodology designed to prevent over-claiming in GEO research, requiring every finding to declare its evidence level, identify confounds, and state only conclusions warranted by its logical structure. The methodology comprises a 7-level evidence hierarchy, 7 logic gates, a language-to-level mapping, and a confidence statement template.

The three summary rules -- state what you know at the level you know it, when in doubt downgrade, and design the experiment that would prove you wrong -- provide an accessible standard for any researcher or practitioner evaluating GEO claims.

Note: This is a methodological framework paper. Confidence statements apply to empirical findings; this paper presents a standard rather than a finding.