Two people have written in with the same confusion: they traced a careful outline and the gallons still looked wrong. In both cases the outline was not the problem. The photo was. Here is what the tracer actually does with your picture, and what to get right before you take it.
Everything the tracer draws on screen, walls, shelves, the floor, is scaled by inches-per-pixel factors set once, from the two tape measurements you type in, and then frozen. Editing the outline afterward moves area and volume the way it should, but it never touches the scale itself, and the scale is never re-derived from the outline. Area comes from that scale on both axes, and depth is typed in separately from the tape, so volume goes as the square of the scale, not the cube. Ten percent off on the scale is about twenty-one percent off on the final gallons. Getting those two tape numbers right, and getting the lines that represent them onto the correct edges of the photo, matters more than how precisely you trace the rim.
Shoot from as high and as close to straight down as you can manage: a ladder, a balcony, an upstairs window, or a drone if you have one. The tracer corrects a tilted photo using the two measured spans, solving a separate scale for each axis so both taped numbers come back exactly right instead of being averaged into something that fits neither. That fixes foreshortening along those two directions. It does not fix the gradient that runs across the rest of the frame in a steeply angled shot, where things near the camera are larger than things far from it in a way a two-axis correction cannot fully untangle. The straighter down the photo, the smaller that leftover error is, which is the real reason height and angle matter more than any other choice you make before you press the shutter.
Pull a tape across the rim twice, roughly at right angles to each other, and type both measured lengths in. One span alone would give the tracer a single uniform scale and no way to check it against anything. Two roughly perpendicular spans let it solve for a foreshortened photo and honor both real-world numbers at once, which is also how you can tell it worked: the status line flips from provisional to set from measured spans once both are in. Put each measurement line exactly where you actually held the tape. A span that drifts off the true edge feeds the tracer a wrong real-world length for a photographed distance that was correct, which is the same kind of error as a bad camera angle, just introduced by hand instead of by the lens.
A purple cross marks the spot on the ground directly under the camera. Anything below the rim, a shelf, a floor, a deep pocket, sits further from the lens than the rim does, so it photographs closer to that point than it really is, pulled toward it. This is relief displacement, the same effect you get looking straight down a stairwell: the bottom step appears to sit almost under your feet, nearer the centerline than it truly is, and the effect grows with how far below the rim a feature sits. Camera height above the rim is what the tracer uses to undo it: enter the real height, in feet, that the camera or drone was above the rim when the photo was taken, and it pushes each below-rim point back out, away from that spot, by a factor of (height + depth) / height. A drone shot from well overhead needs almost no correction, because the whole pond is nearly the same distance from the lens. A phone held at chest height over a three-foot-deep pond needs a lot, because the floor is meaningfully closer to the camera in relative terms than the rim is. Get the height wrong and that factor moves the wrong way: for a pond really shot from 73 inches up with a 36-inch-deep floor, the true push-out factor is (73+36)/73, about 1.49. Enter 36.5 inches by mistake and the factor becomes (36.5+36)/36.5, about 1.99, pushing the floor out too far and inflating the gallons. Enter 200 inches instead and the factor drops to (200+36)/200, about 1.18, pushing it out too little and deflating the gallons.
If what you have is a straight-down drawing or a plan rather than a photo with any real perspective in it, plan mode is the right setting. It still gets its scale from the same two measured spans, and depths are still whatever you type in, but it applies no camera-height correction at all, because a plan view has no camera to correct for. Using it on an actual photo throws away a real correction the tracer could otherwise make; using the normal mode on a plan applies a correction to distortion that was never there.
Some of this cannot be fixed by working the tool harder. A steeply oblique photo, taken low and to one side because there was no way to get above the pond, leaves that near-far gradient uncorrected no matter how carefully the two spans are measured. A dug shelf that grades softly into the slope below it, with no crisp line to follow the way the rim's ground-level edge has one, is genuinely the hardest of the three outlines to trace exactly, and small disagreements about where it runs move the answer by real, if modest, amounts. If there is truly no way to get above a pond, and no plan or drawing to fall back on, the honest answer is that the estimate carries more uncertainty than usual, and it is worth saying so rather than presenting a falsely precise number.