The tool loads a small face-landmark model (Google's MediaPipe Face Landmarker) into your browser the first time you open it — about three megabytes, cached after. When you point the camera or upload a photo, the model maps 468 reference points on your face. The page reads just four of them: face length (top of forehead to chin), forehead width, cheekbone width, and jaw width.
Two ratios do the rest. A face longer than 1.4× its widest point reads as long. Widest at the cheekbones with a narrower forehead and jaw reads as diamond. Three zones roughly equal reads as square or round depending on jaw angle. Wider forehead with a pointed chin reads as heart. Everything else lands in oval — the versatile default that works with most styles.
Nothing is sent anywhere. The image is processed in your browser, the four measurements are derived in your browser, the classifier runs in your browser. Close the tab and there's no trace. We can't see the photo even if we wanted to.