An image-shaped answer instead of a few numbers
A detection model's output is compact, a handful of numbers per object: four coordinates for the box plus a class and confidence score. A segmentation model's output is structurally different. Instead of a few numbers, it produces a full grid matching the input image's dimensions, where every single pixel position gets its own predicted class. If the input image is 640 by 480 pixels, the output mask is also organized as a 640 by 480 grid, and every one of those roughly 300,000 positions carries a class prediction.
This is why people describe a segmentation mask as an 'image-shaped answer'. You can literally visualize it as a picture, typically by painting each class a different color, and overlaying it on the original photo shows exactly which pixels the model assigned to which class. A detection output can't be visualized that way, boxes are just coordinates, but a mask is the same shape and resolution as the thing it's describing.