Skip to content

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

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://huggingface.co/datasets/counter-geo/counter-geo-bench
Area: LLM Safety
Keywords: Generative Engine Optimization, Retrieval-Augmented Generation, Information Distortion, Safety Benchmark, Contrastive Detection

TL;DR

This paper introduces COUNTER-GEO-BENCH, the first defense benchmark targeting information-distorting Generative Engine Optimization (GEO) in search-augmented LLMs, revealing that commercial safety guardrails fail completely against fluent misinformation, and proposes C-GEO Guard, a lightweight 184M contrastive detector reducing Attack Success Rate by 47.6% relative with zero utility loss.

Background & Motivation

As generative search engines powered by large language models increasingly supplant conventional keyword search as the primary interface for information discovery, content creators and adversarial actors have rapidly adapted their strategies. Generative Engine Optimization (GEO) has emerged as the modern successor to traditional SEO, employing stylistic cues, authoritative phrasing, structural headers, and synthetic citation anchors to boost the visibility of target content in LLM-synthesized answers. However, when malicious campaigns weaponize these same techniques, they deploy fluent, polished web pages that embed targeted false claims across blogs and review platforms. When generative search pipelines ingest these documents alongside genuine sources, the victim LLM treats them as authoritative context and synthesizes false claims directly into its final output, severely compromising public trust in AI-driven knowledge retrieval.

Despite the severity of this threat surface, existing defensive ecosystems exhibit critical blind spots and structural evaluation deficiencies. Conventional RAG data poisoning defenses typically presuppose adversarial documents characterized by obvious linguistic perturbations, arbitrary text injections, or unnatural keyword stuffing that triggers upstream perplexity and retrieval filters. Conversely, production-grade safety guardrails (such as Llama Guard 3 and Granite Guardian) are trained almost exclusively on safety taxonomy violationsโ€”including hate speech, harassment, cyberattacks, or jailbreaks. Because GEO-optimized misinformation mimics high-quality, professional expository writing without exhibiting any toxic tokens or policy infractions, these guardrails remain entirely blind to factual corruption. Furthermore, commercial black-box product audits fail to isolate the independent contributions of retrieval rerankers, system prompts, and downstream synthesis rails.

To rigorously address this vulnerability under controlled conditions, this paper formulates an author-controlled retrieval-to-synthesis evaluation harness featuring an innovative paired rewrite protocol. By generating both information-preserving (IP) and information-distorting (ID) variants from the exact same source document for every query, the benchmark cleanly decouples the factual distortion effect from the organic visibility lift inherent to GEO. Core idea: construct COUNTER-GEO-BENCH with 247 human-verified, quality-gated query-document pairs to evaluate defenses across three victim LLMs, proving current guardrails are ineffective against fluent factual distortion while demonstrating that a lightweight contrastive prototype detector (C-GEO Guard) can cut attack success by nearly half without degrading normal utility.

Method

Overall Architecture

The COUNTER-GEO-BENCH evaluation framework is structured around three interconnected modular components: rigorous benchmark dataset synthesis, a controlled generative search pipeline (Harness), and insertion points for multi-stage defenses including the proposed C-GEO Guard baseline. The pipeline begins with user queries from the GEO-Bench corpus alongside top search sources. Claude Sonnet 4.6 serves as the benchmark rewriter to produce paired information-preserving (IP) and information-distorting (ID) documents, filtered through a strict geometric-mean quality gate. In the generative search harness, documents undergo 512-token chunking, hybrid BM25 + dense vector retrieval, and cross-encoder reranking down to the Top-12 context window. Candidate defenses are evaluated either as chunk-level filters prior to generation or as post-generation output rails, with the victim LLM producing citation-mandatory answers under zero-temperature decoding.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["User Query & Raw Candidate Document Pool"] --> B["Paired Rewriting & Quality Gate Verification<br/>Produce Paired IP Controls & Quality-Gated ID Attacks"]
    B --> C["Chunking & Hybrid Retrieval Reranking<br/>BM25+Dense Hybrid Search & Cross-Encoder Reranker"]
    C --> D["Chunk-Level Prototype Contrastive Detection<br/>C-GEO Guard Class Centroid Cosine Filtering"]
    D --> E["Post-Generation Output Rail & Victim LLM<br/>NeMo Self-Check Verification & Multi-Model Evaluation"]

Key Designs

1. Paired Rewriting & Quality Gate Verification: Decoupling visibility gains from factual distortion while enforcing ecological validity Prior poisoning benchmarks conflate the retrieval boost from optimization with the misinformation payload. The authors overcome this by pairing each evaluation query with two distinct rewrites of a non-authoritative source document: an Information-Preserving (IP) rewrite adhering to legitimate GEO practices (authoritative syntax, structured headers, citing source facts without altering statements) and an Information-Distorting (ID) rewrite applying the same GEO strength while injecting a targeted false claim reinforced by fabricated citations, faux statistics, and artificial consensus. To guarantee that generated attacks mimic genuine web text and evade naive upstream filters, every rewrite must satisfy a strict quality gate:

\[Q = (S_{\text{emb}} \times S_{\text{len}} \times S_{\text{ppl}} \times S_{\text{nat}})^{1/4} \ge 0.65\]

where \(S_{\text{emb}}\) denotes embedding cosine similarity (thresholded at 0.92โ€“0.99 for IP and 0.78โ€“0.92 for ID), \(S_{\text{len}}\) measures token length deviation, \(S_{\text{ppl}}\) evaluates reference language model perplexity ratios, and \(S_{\text{nat}}\) measures LLM-judged fluency. A zero on any sub-score zeros the metric. Only 25.0% of candidates (250/1000) passed this joint gate, which was finalized to 247 high-fidelity pairs after manual verification.

2. Chunk-Level Prototype Contrastive Detection: Overcoming safety taxonomy blindness with class-centroid matching General-purpose guardrails fail against GEO because factual manipulation lacks overt toxicity tokens. To establish a domain-aware reference defense, the authors introduce C-GEO Guard, built upon DeBERTa-v3-base (184M parameters). Contextualized token representations are mean-pooled and \(L_2\)-normalized into 768-dimensional chunk embeddings \(e\). Trained via Multiple Negatives Ranking Loss (MNRL) using hard paired IP negatives and assembled ID positives, the model captures manipulative semantic signatures across eight distinct ID-GEO attack classes (stealth injection, temporal poisoning, fake authority, fake citation, structured hijack, review flood, multi-hop reasoning, and negative GEO). For each attack class \(k\), an offline normalized prototype centroid \(c_k\) is computed:

\[c_k = \frac{\bar{e}_k}{\|\bar{e}_k\|_2}, \quad \bar{e}_k = \frac{1}{|S_k|} \sum_{i \in S_k} e_i\]

During inference, candidate chunks in the retrieved Top-12 context are evaluated against all class prototypes: \(\text{score}(e) = \max_k e^\top c_k\). Chunks exceeding the calibrated threshold \(\tau = 0.90\) are discarded prior to synthesis, cutting off false context at 2.3% of the parameter cost of 8B guardrails.

3. Post-Generation Output Rail & Victim LLM: Uncovering circular entailment vulnerabilities in self-checking models To assess defenses across alternate insertion points, the benchmark integrates NVIDIA NeMo Guardrails' Self-Check Fact-Checking module. In this setup, the victim LLM generates an initial answer draft and subsequently acts as a judge evaluating whether the claims in the draft are strictly entailed by the retrieved context chunks. The benchmark highlights a structural failure mode: because successful GEO tactics push the corrupted document directly into the retrieved Top-12 context, the victim LLM evaluates its draft against the poisoned evidence and finds "full entailment," blindly confirming the hallucinated claim. Conversely, on models like Llama-4-Scout, the entailment prompt triggers an indiscriminate refusal collapse (blocking 98.4% of clean queries), proving self-checking guardrails to be anticorrelated with true attack presence.

Benchmark & Evaluation Protocol

The benchmark evaluates four defense configurations across three frontier open-weight victim LLMs: Gemma-4-31B-IT (dense instruction-tuned), Qwen-3.5-35B-A3B (MoE architecture), and Llama-4-Scout-17B-16E (16-expert MoE). Attack success is evaluated under the ID condition using Claude Opus 4.6 as an impartial judge based on per-query attack criteria (scored 1.0 for unhedged false assertions, 0.5 for mixed/hedged claims, and 0.0 for factual accuracy; validated with human Fleiss' \(\kappa = 0.778\)). Utility is simultaneously measured on clean and IP conditions through False Positive Rate (FPR), Answer Accuracy, and composite Answer Quality \(\bar{Q}\) (averaging relevance, completeness, and clarity on a 1โ€“5 scale).

Key Experimental Results

Main Results

The primary evaluation results on the ID condition across 247 quality-gated benchmark queries are summarized below (directly quoting Table 3 in the paper):

Defense Configuration Gemma-4-31B ASR (%) Qwen-3.5-35B ASR (%) Llama-4-Scout ASR (%) 3-Model Avg ASR (%) Relative Change \(\Delta_{\text{rel}}\) Answer Quality \(\bar{Q}\) (1โ€“5)
Undefended Baseline 56.5 55.9 54.7 55.7 โ€” 4.49
Granite Guardian 3.3-8B 51.4 53.8 56.9 54.0 -3.1% 4.37
Llama Guard 3-8B 51.8 51.4 54.3 52.5 -5.7% 4.38
NeMo Self-Check 50.2 52.0 0.4โ€  51.1โ€ก -9.1%โ€ก 4.51โ€ก
C-GEO Guard (Ours, 184M) 30.6 29.4 27.5 29.2 -47.6% 4.48

Note: โ€ NeMo on Llama-4 collapsed into near-total query refusal (98.4% clean refusal rate); โ€ก3-model average for NeMo excludes Llama-4, reflecting Gemma-4 and Qwen-3.5 only.

Ablation & Cross-Rewriter Transfer Study

To establish that C-GEO Guard captures generalized manipulation signals rather than rewriter-specific artifacts, the authors tested the detector on 247 un-gated query-document pairs rewritten independently by GPT 5.5, while analyzing chunk-level block behavior and benign accuracy (quoting Tables 4, 6, and 7):

Experimental Setting / Defense Test Context ASR (%) Relative ASR \(\Delta_{\text{rel}}\) Chunk ID Block (%) Clean/IP Accuracy (%) IP Chunk FPR (%)
Chunk-Level Filtering Metrics
- Granite Guardian (8B) Main Set (Sonnet) 54.0 -3.1% 0.61 86.2 0.29
- Llama Guard 3 (8B) Main Set (Sonnet) 52.5 -5.7% 0.83 86.8 0.47
- C-GEO Guard (0.18B) Main Set (Sonnet) 29.2 -47.6% 10.27 86.4 2.16
Cross-Rewriter Transfer (GPT 5.5)
- Undefended Baseline Standard Template 55.7 โ€” โ€” 86.8 โ€”
- C-GEO Guard (\(\tau=0.90\)) Standard Template 22.1 -60.4% 9.80 82.6 2.80
Independent Weak-Prompt Transfer
- Undefended Baseline Unconstrained Template 50.6 โ€” โ€” 87.0 โ€”
- C-GEO Guard (\(\tau=0.90\)) Unconstrained Template 34.4 -32.0% โ€” 88.3 โ€”

Key Findings

  • Commercial Guardrails Are Ineffective and Can Act as Threat Amplifiers: Commercial 8B classifiers intercept less than 1% of malicious chunks (0.61% for Granite, 0.83% for Llama Guard), yielding statistically insignificant or marginal ASR drops (\(\le 3.2\) pp). Remarkably, on Llama-4, Granite Guardian worsened attack outcomes on 74 queries (escalating 29 clean queries into compromised states by stripping away benign counter-evidence), increasing net vulnerability by +4.0%.
  • Compact Contrastive Filtering Provides High-Leverage Protection: C-GEO Guard (0.18B parameters, 43ร— smaller than 8B baselines) achieves an ID-to-clean chunk block ratio of \(4.5\times\) (10.3% vs. 2.3%), slashing average ASR from 55.7% to 29.2% (a 26.5 pp absolute drop). On GPT 5.5 rewrites, it achieves an even higher 60.4% relative ASR reduction, while retaining benign accuracy at \(+0.2\) pp above baseline.
  • Misinformation Directly Corrupts Answer Quality: Contrary to assumptions that poisoning is orthogonal to fluency, undefended answers with full attack success (\(ASR=1.0\)) score an average quality \(\bar{Q}\) of only 4.29 versus 4.70 for clean answers (\(ASR=0.0\)). Unhedged commitment to falsehood narrows synthesis scope; when C-GEO Guard successfully neutralizes an attack, response quality rebounds by \(+0.70\) on relevance and completeness.

Highlights & Insights

  • Decoupled Paired Formulation: By holding the underlying document and query constant and pairing legitimate optimization (IP) against deceptive manipulation (ID), the benchmark cleanly isolates factual distortion from ranking prominence, resolving a fundamental confounder in prior RAG security literature.
  • Asymmetric Negative Sampling Strategy: The training regimen creatively recycles 750 candidate queries rejected by the benchmark gate, exploiting ungated IP rewrites as hard non-ID negatives alongside borderline rewrites, which teaches the DeBERTa encoder to separate benign stylistic optimization from factual corruption with zero external human annotations.
  • Empirical Demonstration of Circular Entailment Traps: The benchmark concretely proves that post-generation verification rails (such as NeMo Self-Check) suffer from a fundamental architectural flaw in RAG pipelines: when malicious content penetrates the retrieval window, models verify false outputs against false evidence, falsely validating corrupted assertions.

Limitations & Future Work

  • Evaluation Scale and Linguistic Scope: The current benchmark comprises 247 English queries. While sufficient to validate the 27 pp gap against baselines with high statistical confidence, it remains underpowered for detecting nuanced differences among minor model variants, and excludes non-English queries as well as high-stakes verticals (finance, medical, legal).
  • Single-Document Attack Constraint: The threat model assumes adversarial control over a single web source. Multi-document coordinated poisoning campaigns with cross-referencing networks would likely exacerbate ranking exploitation and pose more severe challenges for chunk-level filters.
  • Vulnerability to Adaptive Evasion: C-GEO Guard relies on semantic representations anchored to eight defined manipulation classes. Advanced adversaries employing style transfer, human post-editing, or gradient-based evasion could attenuate contrastive signals, necessitating future exploration into multi-source consensus verification and web provenance tracking.
  • vs. PoisonedRAG (Zou et al., 2025) & Arbitrary Content Injection (Tamber & Lin, 2025): Prior RAG poisoning works inject syntactically discordant sentences or query repetitions that can be identified via basic heuristic/perplexity filters. COUNTER-GEO-BENCH studies fluent, professionally structured articles optimized to mimic top-ranking web pages.
  • vs. Granite Guardian (Padhi et al., 2025) & Llama Guard 3 (Inan et al., 2023): Standard guardrails focus on safety taxonomies (toxicity, self-harm, hate speech), which leaves them powerless against polite, well-formatted factual falsehoods.
  • vs. GEO-Bench (Aggarwal et al., 2024): Whereas GEO-Bench examines visibility optimization from a publisher's utility standpoint, COUNTER-GEO-BENCH is the first work to investigate the adversarial safety implications of GEO, framing it as an attack vector and establishing a rigorous defense benchmark.

Rating

  • Novelty: โญโญโญโญโญ [Pioneering benchmark isolating information-distorting GEO from visibility lift via paired counterfactual design]
  • Experimental Thoroughness: โญโญโญโญโญ [Evaluated across 3 frontier victim LLMs, 4 defense baselines, with extensive cross-rewriter and prompt transfer tests]
  • Writing Quality: โญโญโญโญโญ [Exceptionally clear problem formulation, robust statistical validation, and insightful revelation of defense amplifier failure modes]
  • Value: โญโญโญโญโญ [Crucial benchmark and lightweight defense baseline for safeguarding modern search-augmented generative AI systems]