Large Language Models (LLMs) have demonstrated strong performance in data-to-text generation (D2T) but often suffer from factual inconsistencies. A key contributing factor to these inconsistencies is source-reference divergence—a mismatch between the input data and the reference text. In this work, we focus on modeling this divergence explicitly. We introduce three complementary metrics—Unigram-Level Divergence, Named Entity-Based Divergence, and Field-Aware Divergence—to quantify the degree of divergence, and we leverage these signals to guide generation. Specifically, we propose a divergence-aware selective generation framework that enables LLMs to abstain from generating outputs when divergence is high. Experiments across three diverse D2T datasets (WikiTableText, ViGGO, WebNLG) and three LLMs (Qwen2.5, FLAN-T5, OPT) show that our method significantly improves factual consistency, particularly at moderate coverage thresholds (70–80%). We observe consistent gains across four popular factuality metrics—AlignScore, QAFactEval, SummaC-Conv, and UniEval-Fact—without sacrificing fluency or diversity. These findings underscore the value of modeling divergence and establish selective generation as an effective strategy for improving factual consistency in LLM-based D2T.