Clickbait posts have become rampant on social media platforms, causing a multitude of detrimental impacts, e.g., the propagation of disinformation. Most of the existing studies focus on text-centric clickbait. They often make judgments based on shallow single-modality features of the text content, ignoring the deep content and commonsense clues contained in the images. It is difficult to explain where the inconsistencies are, and these shallow features cannot indicate complex clickbait well. To address these problems, we propose a new approach to infer inconsistencies from a suspect-then-verify perspective. It can suspect each potential clue of the multi-modal content by asking diverse and incisive questions, and then verify each clue by commonsense reasoning. That can well find evolving bait tricks with good interpretability. Specifically, we first analyze and represent the multi-modal content and context of the post. We then question the authenticity and consistency of the content. We take into account knowledge such as external commonsense to generate reasonable and diverse questions, capturing various bait types better. To facilitate the generation process, we learn the questions’ reasoning structures and expressive patterns from open-source data. Next, we answer these questions to infer potential inconsistencies. The verification results are combined with typical clickbait features to derive the final prediction. We conduct a comprehensive set of experiments on three popular datasets to demonstrate the effectiveness of our approach.