Verified Misguidance: Measuring Structural Citation Failures in Search-Augmented LLMs

Researchers have identified systematic citation failures in search-augmented LLMs, where models cite real sources yet distort their meaning or select inappropriate sources. The CITETRACE dataset reveals that 30.6% of citations distort sources and up to 96% of users encounter misleading citations, with provider-level factors accounting for 88-96% of citation quality variance.

Search-augmented language models have become critical infrastructure for information delivery, processing millions of queries daily. Users inherently trust citations as verification that responses are grounded in legitimate sources, yet rarely fact-check cited materials themselves. This creates a systemic vulnerability where flawed citations can propagate misinformation at massive scale. The CITETRACE study exposes a structural failure pattern called Verified Misguidance—where models cite real, accessible sources but fail along critical dimensions: misalignment between query intent and source purpose, selection of domain-inappropriate sources, or distortion of source content. These failures create a troubling trade-off where models that faithfully represent sources often pick unsuitable ones, while models selecting appropriate sources may distort their meaning. The research framework evaluates citations across three dimensions using expert-validated matrices and a five-level fidelity rubric, providing the first comprehensive structural assessment of deployed systems. Provider-level differences explaining 88-96% of citation-quality variance suggest that source selection is governed by factors beyond individual model capability—likely retrieval ranking, training data composition, and retrieval system architecture. This finding indicates that improving LLM citations requires systemic changes at the provider level, not merely better base models. The implications extend beyond accuracy: users misled by verified sources face eroded trust in AI systems while believing they've engaged in due diligence. For AI providers, this represents both reputational risk and competitive differentiation opportunity. The CITETRACE framework establishes measurement standards that will likely become industry benchmarks, pressuring providers to improve citation reliability.

• →30.6% of citations distort their sources and 27.1% originate from domain-inappropriate sources across tested models
• →Up to 96% of users encounter at least one structurally misleading citation in responses, despite citations referencing real sources
• →Provider-level factors explain 88-96% of citation quality variance, indicating systemic architectural issues rather than model limitations
• →CITETRACE provides the first comprehensive evaluation framework for diagnosing structural citation failures in deployed search-augmented systems
• →The verified misguidance pattern reveals a fidelity-suitability trade-off where faithful models select inappropriate sources and vice versa