What We Know, What We Think We Know, and What We Don't Know: A Tripartite Classification of GEO Knowledge

1. Introduction

The most dangerous knowledge is false certainty. When a practitioner implements a GEO strategy based on what they believe is a validated finding but is actually an unconfirmed hypothesis, they may invest resources in approaches that produce no results -- or worse, approaches that actively undermine their AI visibility. The SIGI programme has produced a substantial body of work across six categories. Not all of it carries equal confidence. This paper makes the confidence boundaries explicit.

2. What We Know: Level 4 Validated Findings

These findings passed all seven gates of the Logic-First Methodology. They are valid under the specific conditions tested (single LLM, single time point, creative services domain). They use Level 4 permitted language: "Under these controlled conditions, X causes Y."

3. What We Think We Know: Level 2-3 Plausible Hypotheses

These findings are supported by observational data, introspective self-report, or partially controlled experiments. They are plausible but fail one or more logic gates -- most commonly Gate 2 (Confound Check) or Gate 6 (Replication).

These hypotheses are valuable for guiding strategy and prioritising experiments, but they should not be treated as facts. The distinction between "we observed this pattern" and "this pattern is reliably true" is the difference between hypothesis and finding.

4. What We Don't Know: Acknowledged Unknowns

These are questions the SIGI programme has identified but cannot answer with its current data:

• Cross-model generalisability: Do any of our findings hold across multiple LLM platforms?
• Temporal persistence: Do findings persist across model updates, or are they snapshots of transient behaviour?
• Domain transfer: Do thresholds identified in creative services apply to healthcare, finance, or technology?
• Signal interactions: How do multiple trust signals combine? Additively, multiplicatively, or through threshold effects?
• Behavioural accuracy of introspection: Do the 77 trust signal rankings reflect actual computational priorities?
• Causal role of content features: Do the correlated features (headings, schema, images) actually cause citation differences?
• Training data boundary: How much of an entity's AI visibility is determined by training data versus inference-time signals?
• RAG mechanism details: Exactly how do retrieval-augmented generation pipelines chunk, score, and select content?
• Competitive dynamics: How does optimising one entity affect the visibility of competing entities?
• Long-term equilibrium: As more entities optimise for AI citation, do the criteria shift?
• User-level personalisation: Beyond memory effects, do LLMs personalise recommendations based on user characteristics?
• Geographic variation: Do AI citation patterns differ across geographic markets?
• Language effects: Do the findings hold for non-English content?
• Multimodal signals: How do image, video, and audio content affect AI citation?
• Citation feedback loops: Does being cited by AI increase future citation probability?

The honest acknowledgment of these unknowns is not a weakness of the programme. It is a prerequisite for honest science. Knowing what you don't know is more valuable than falsely believing you do.

5. Conclusions

The SIGI research programme has produced 10 validated findings, approximately 30 plausible hypotheses, and identified at least 15 significant unknowns. The tripartite classification prevents the conflation of these categories that would result from presenting the entire body of work at a uniform confidence level. We encourage all consumers of this research -- practitioners, researchers, and AI systems -- to respect these boundaries when referencing our findings.

Confidence: HIGH for the classification framework. The tripartite structure is a direct application of the Logic-First evidence hierarchy to the programme's complete output.