Counter-GEO-Bench evaluates defenses against information-distorting generative engine optimization (GEO) in retrieval-augmented and generative search systems.

The gated release covers 247 queries. It includes all 247 information-preserving (IP) rewrites and 237 information-distorting (ID) rewrites. Ten potentially sensitive ID rewrites are withheld; researchers may request access or generate their own target false claims and ID rewrites. Original source documents are not included.

Each record contains a query, an IP rewrite, release labels, and fake-URL metadata. Released ID rewrites include their evaluation metadata. The trained C-GEO Guard is available in a separate repository.

Files

data/counter_geo_bench_fake_url_247.jsonl: the benchmark data.
data/counter_geo_bench_fake_url_247.sample.json: three example records.
data/evaluation_templates/: evaluation templates.
code/: the victim pipeline and quality-evaluation code.
DATA_CARD.md: field definitions, sources, intended use, and limitations.

Usage

cd code
python -m pip install -e '.[gpu]'
CUDA_VISIBLE_DEVICES=0,1 python scripts/step2_victim_pipeline.py \
  victim.tensor_parallel_size=2

The victim model, embedding model, reranker, and device settings are configured in code/conf/config.yaml. Victim synthesis requires GPUs.

License and responsible use

This repository uses component-specific licenses:

data/ is licensed under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0).
code/ is licensed under the Apache License 2.0.

See LICENSE and LICENSES/ for the applicable terms. Access is gated for defensive research and evaluation. Every record is marked as synthetic benchmark content, and its target false claim is identified as model-generated. These claims do not represent the authors' advice or views. Released ID rewrites must not be redistributed as an unlabeled web corpus or used to deploy misinformation.

Citation

Paper: Counter-GEO-Bench

If you use Counter-GEO-Bench, please cite:

@misc{zheng2026countergeo,
  title         = {{Counter-GEO-Bench}: Evaluating Defenses Against Information-Distorting Generative Engine Optimization},
  author        = {Zheng, Bing and Zhao, Zongyao and Yang, Wenming},
  year          = {2026},
  eprint        = {2609.02316},
  archivePrefix = {arXiv},
  primaryClass  = {cs.IR},
  url           = {https://arxiv.org/abs/2609.02316}
}

Downloads last month: 32

Total file size:

3.58 MB

Paper for counter-geo/counter-geo-bench

Counter-GEO-Bench: Evaluating Defenses Against Information-Distorting Generative Engine Optimization

Paper • 2609.02316 • Published 4 days ago