Recent advances in large language models (LLMs) have led to powerful AI-driven search and analysis platforms (e.g., GPT-4o, DeepSeek R1). While these systems can effectively locate and summarize data, they often rely on limited or single-source inputs, increasing the risk of low-quality or unverified outputs. This paper addresses the critical challenge of multi-source reliability by proposing a graph-based framework that (1) unifies heterogeneous data via LLM-driven schema extraction, (2) models multi-relational, time-evolving information in a Temporal Graph Attention Network (TGAT), and (3) provides final, user-facing analysis anchored on validated, cross-source evidence. We demonstrate our methodology on industry research data - where incomplete or inconsistent sources frequently lead to contradictory market size predictions - and show that our approach outperforms GPT-4o-only baselines in accuracy and interpretability. While we focus on industry forecasting as an example, the proposed framework is applicable to any domain requiring robust, multi-source AI search and analysis.