Why GEO Security Reports Need a Benign Evidence Retention Metric

A system can achieve a low attack rate by citing almost nothing. That kind of safety creates thin answers, more refusals, and expensive human rework. A defense report must show how much normal evidence remains.

The GEO Defender paper uses Benign Evidence Retention (BER) to compare the number of benign sources used in a defended answer with a clean-reference answer. In its experiment, a static safety prompt averages 83.31% BER, while the two-stage method reaches 94.12% and reduces average attack success to 6.20%.

A practical measurement set

Choose questions with human-verified answers and save the normal source set before enabling a defense. Then record:

• proportion of benign sources retained;
• whether critical facts still have support;
• attack or erroneous sources used in answers;
• refusal and extra-retrieval counts;
• human repair time per answer.

BER above 100% only means more benign documents were cited in aggregate. It does not prove higher quality: sources may be redundant or relevant without supporting the conclusion.

GEO Radar (https://www.georadar.top) can help brands observe sources used across AI platforms and how competitor-source combinations change. When candidate pools are hidden, reports must label this as answer-level black-box observation rather than a full BER reproduction.

Cost boundary

Retaining more evidence can increase context and review costs. The goal is not maximum source count, but enough independent support for critical facts with minimal irrelevant material. The paper does not establish one threshold for every private repository, language, or production platform.

Sources for this article

• arXiv, September 2, 2026, *When Optimization Becomes Manipulation: Defending Generative Search against Malicious Generative Engine Optimization*: https://arxiv.org/abs/2609.02964