In text chat, LLM-powered conversational agents (CAs) often include supplementary cues about uncertainty and provenance alongside the main response. These cues help users judge reliability and notice potential AI hallucinations. In immersive environments, these agents often take embodied form as embodied conversational agents (ECAs). Shifting from text chat to a speech-led 3D avatar removes persistent inline text, when spoken they are easy to miss and can disrupt comprehension. In this study, we explored different ways to surface such cues in ECAs. We conducted a within-subjects study (N = 24) with AI hallucination identification tasks in VR, comparing a No-Cue baseline with three designs that each encode uncertainty and source provenance: Embodied-Cues (avatar gestures and posture), Icon-Cues (visual indicators on probability of hallucinations and source), and Text-Cues (color-coded text spans with inline citations). We evaluate how the three designs affect hallucination-detection efficiency, user trust on ECA, immersion, cognitive load, and self-reported ability to identify and interpret AI hallucinations.