DEJAN continuously harvests and parses low-level grounding payloads across major model architectures, ingesting structures like Google Gemini's grounding chunks and support segments, OpenAI's web search call annotations, and Anthropic's server tool results. By resolving Vertex AI redirect URIs to canonical destinations and comparing retrieved sources against inline citation spans, the platform establishes clear conversion metrics—such as the domain-level trust ratio—to expose which pages models browse but ultimately reject during synthesis.

This telemetry directly drives automated optimization loops that maximize citation acquisition and factual accuracy. Through a Bayesian content optimization engine, extractive snippet and page variants are iteratively generated and evaluated using multi-sample re-ranking to determine which content structures reliably gain placement. Concurrently, a veracity evaluation engine scores post-citation fidelity by tracking whether source claims survived intact, were omitted, or suffered model-induced distortions, guaranteeing that cited brand representations remain both prominent and verifiably truthful.