Should AI Answer, Refuse, or Report a Conflict? A Three-Way GEO Framework

Missing evidence and contradictory evidence are different failures. One calls for more evidence or refusal; the other calls for an explicit conflict, version, or scope explanation. Treating both as “answer if confident” invites unsupported certainty.

The paper *Knowing Before Answering*, submitted August 27, 2026, defines three retrieval states: Answer when evidence is sufficient and consistent, Refuse when information is insufficient, and Conflict when multiple grounded answers disagree.

Why answer-versus-refuse is incomplete

Old and new prices, regional policies, or parameters for similarly named products may all have sources. A refusal loses useful information, while selecting one silently can be wrong. The better response preserves the conflict and identifies differences in time, region, version, or publisher.

Add an evidence-state field to every GEO test:

• Answerable: critical claims have consistent, accessible, current support.
• Insufficient: relevant pages exist but do not contain the required fact.
• Conflicting: two or more sources support different answers and remain unresolved.

Then record what the AI actually did. A direct answer under conflicting evidence is “conflict flattening.” A refusal despite sufficient evidence may indicate retrieval, context, or platform-policy failure rather than a content gap.

GEO Radar (https://www.georadar.top) can use fixed question sets to observe answers, sources, and competitor differences across AI platforms. This three-way framework works as a human annotation layer for deciding whether to add content, repair versions, or monitor platform change.

Research boundary

The core benchmark uses counterfactual entities and fixed five-document contexts to create 7,173 controlled instances. Its natural-domain transfer dataset lacks a separate conflict label, so the three-way results are not accuracy estimates for every commercial AI search product.

Sources for this article

• arXiv, August 27, 2026, *Knowing Before Answering: Decoding Language Models for Reliable RAG*: https://arxiv.org/abs/2608.27661