The ideal AR x-ray vision should enable users to clearly observe and grasp not only occludees, but also occluders. We propose a novel selective visualization method of both occludee and oc-cluder layers with dynamic opacity depending on the user's gaze depth. Using the gaze depth as a trigger to select the layers has a essential advantage over using other gestures or spoken commands in the sense of avoiding collision between user's intentional commands and unintentional actions. Our experiment by a visual paired-comparison task shows that our method has achieved a 20% higher success rate, and significantly reduced 30% of the average task completion time than a non-selective method using a constant and half transparency.