Four Swing Analysis Apps, One Identical Swing: What They Each Said

I filmed one 7-iron. Face-on, 240fps, phone on a cheap tripod about waist height, four metres away, on the range at my home club on a still Tuesday morning. The strike was decent: 158 yards carry on the Rapsodo MLM2PRO sitting behind me, 7.2° down, 4,980rpm, 2.1° out-to-in path with a face 0.8° open to that. Slight pull-fade. The sort of shot I hit maybe six times in ten on a good day, and the reason I play off 11 rather than 8.

Then I fed that same MP4 into four of the most-recommended swing analysis apps and asked each one the same thing: what’s wrong, and what should I work on?

I got four answers. Not four phrasings of one answer. Four genuinely different diagnoses, two of which contradicted each other outright, and one of which described a swing fault that is physically incompatible with the ball flight the launch monitor recorded eight feet away.

That gap is the whole story. If you’re searching for the best golf swing analyser app, the useful question isn’t which one has the nicest interface. It’s which parts of what these things tell you are measurement, and which parts are a language model doing improv over a pose-estimation skeleton.

What each app actually said

Here’s the raw output, trimmed for length but not reworded.

Sportsbox 3D Golf (subscription, roughly £180/year on the annual plan) gave me numbers rather than prose. Its headline items on this swing:

Pelvis Sway (top)        4.1 in trail
Pelvis Thrust (impact)   1.8 in toward ball
Chest Bend (top)         38.4°
Lead Wrist Extension     19.2° (top) → 11.7° (impact)
Hand Depth (P4)          Off-plane, 2.3 in outside benchmark

The diagnosis it drove me toward: lead wrist extension (cupping) at the top, and hands too far from the body at P4.

Swing AI (the one that markets itself hardest to mid-handicappers, around £8/month) told me I had an “over-the-top downswing caused by early upper body rotation” and prescribed the classic towel-under-the-armpit drill.

HackMotion’s app, paired with the wrist sensor rather than reading video, reported lead wrist extension of 21° at the top flattening to 9° at impact. It said nothing at all about my pelvis, because it can’t see it. That restraint turns out to matter.

A general-purpose multimodal model (I used Claude with the video frames extracted, plus ChatGPT’s vision on the same frames as a cross-check) said my weight was hanging back through impact, that I was “falling away from the shot”, and that I should work on pressure shift into the lead side.

So: cupped wrist, over the top, cupped wrist, weight hanging back. One swing.

Which of those can survive contact with the launch monitor

The over-the-top diagnosis is the interesting failure, because it’s not merely unhelpful. It’s wrong in a way you can prove.

An over-the-top move produces a steep, out-to-in path with the club approaching from outside the target line, and unless the face is shut to that path you get a slice or a pull. My path was 2.1° out-to-in. That is not over the top. That’s inside the range most tour players live in on a mid-iron, and a lot of good ball-strikers sit around 2 to 3° left with a 7-iron by design. Attack angle was 7.2° down: normal, not steep. Swing AI looked at a face-on video, saw my trail shoulder move round early (it does, slightly), and reached for the most statistically common mid-handicap fault in its training distribution.

That’s the mechanism behind most bad AI swing advice, and it’s worth naming plainly. The app isn’t measuring your path. It’s pattern-matching your body shape to a label, then generating the standard prescription attached to that label. Fed a 14-handicapper’s face-on video, “over the top” is a decent bet. It just wasn’t true here, and the towel drill would have moved my path further left, into the pull-hook territory I spent last winter getting out of.

The weight-hanging-back read from the general-purpose models fails a different test. My pressure trace from the mat showed 78% into the lead foot by impact. What those models were seeing was secondary axis tilt, the normal backward lean of the upper body that every competent iron player has at impact, and interpreting it as a fault because from face-on, tilt and hanging back look similar in a still frame. This is a known failure mode: ask a vision model to judge weight distribution from a 2D video with no force data and it will confidently tell you something. It has no access to the quantity in question.

Sportsbox and HackMotion agreed, independently, on lead wrist extension at the top: 19.2° from video-based 3D, 21° from a sensor strapped to my wrist. Two different measurement methods, two degrees apart. That’s the signature of something real. And it’s consistent with the ball flight: a cupped lead wrist at the top tends to arrive with the face open, and my face was 0.8° open to path.

The rule this gives you

Trust the apps that measure. Discount the apps that describe.

MEASURED (defensible)          INFERRED (treat as a hypothesis)
-----------------------------  --------------------------------
Wrist angles (sensor)          "You're over the top"
Wrist angles (3D video, ±2°)   "Your weight is hanging back"
Pelvis/chest rotation & sway   "You're casting the club"
Hand path depth                "Your tempo is too quick"
Anything from the monitor      Any cause-and-effect story

The left column comes from a system with a physical quantity behind it. The right column comes from a system that has been asked to produce golf-coach-shaped text. Both feel equally authoritative in the app. Neither carries a confidence score, which is the single biggest design failure across all four.

For the fuller head-to-head, including the setups where each app broke down and how their outputs drifted across five repeat swings, the pillar piece AI swing analysis apps tested on the same swing has the full run.

What actually changed in my practice

Because two independent measurement systems agreed on the wrist, that’s what I worked on. Nine weeks, three sessions a week, mostly one drill: lead wrist flexion pumps at the top with a HackMotion audio cue set to beep above 12° extension. Boring. Effective.

Where the numbers went:

The dispersion number is the one that moved my scores. Not the extra four yards.

Had I followed the over-the-top prescription instead, I’d have spent nine weeks fixing a fault I didn’t have while the real one sat there. That’s not a hypothetical cost. It’s the ordinary experience of club golfers who take app advice at face value, and it’s why the free-and-chatty tier of this market is worse than useless for anyone who already has a ball-flight problem they’re trying to solve.

How to run this check on your own swing

You don’t need my kit to do the same filter. You need one video and one piece of ball-flight information, even rough.

Film face-on and down-the-line, waist height, phone perpendicular to the target line, 240fps if your phone offers it (iPhone 12 and up do; most mid-range Androids don’t, and 60fps will smear the impact frames badly enough to throw off any wrist measurement). Then note what the ball actually does. Starts left and curves right? Starts right and stays? Even without a monitor, start direction tells you roughly where the face was, and curve tells you the face-path relationship.

Now feed the video to your app of choice and check the diagnosis against the flight. If the app says you’re over the top and your misses are pushes and blocks to the right, the app is wrong. If it says you’re hanging back and you take a divot in front of the ball every time, the app is wrong. Contradiction between the diagnosis and the flight is the cheapest lie detector available, and it costs nothing.

Where an app gives you a number instead of a story, it earns more trust. Where two systems that work differently give you the same number, act on it.

One more thing about cost, since it’s the question I get asked most. Sportsbox at £180/year plus HackMotion’s sensor at around £250 is real money for a club golfer, and I wouldn’t tell a 24-handicapper to buy either. At that level your gains are in strike location and setup, both of which you can audit yourself with impact tape and a phone. The measurement kit starts paying for itself somewhere around 12, when your misses get specific enough that the fault you’re chasing is genuinely hard to see.