Implementations vary in primitive: some read pixels and click coordinates from screenshots, others read a structured accessibility tree (the OS's built-in map of on-screen elements and their roles) and act on it, and some combine both with scripting layers like AppleScript for reliability. The trade-off is consistent across implementations — accessibility-tree and structured-API control is more robust than raw coordinate clicking, but any canvas-rendered interface (custom-drawn UI outside a browser) forces a fallback to screenshots and coordinates, which is inherently less reliable.