Full swing · Downswing

What AI golf swing analysis actually measures

Every golf app now claims AI. Almost none of them say what is being measured, which makes the claim impossible to judge and easy to oversell. This page is the specific version: what a vision model reads from an ordinary phone video, how confident it is, and where it stops being reliable.

There are two separate things happening under the label, and they have very different accuracy characteristics.

1Pose estimation: where your body is, frame by frame

A pose model locates 33 body landmarks — shoulders, hips, elbows, wrists, knees, ankles — in every frame of the video. This runs in your browser, so the video never has to leave your phone for this step. It is the most reliable part of the pipeline because it is a well-established computer vision task, and it is what makes a skeleton overlay and angle measurements possible at all.

2Phase detection: finding the six moments that matter

A vision model watches the video and identifies Address, Backswing Parallel, Top, Downswing Parallel, Impact and Finish, returning the frame where each occurs. This is genuinely useful for comparing two swings at the same point rather than at the same timestamp. Treat the frame numbers as close, not exact — at 30 frames per second the club travels a long way between frames, and impact in particular is often a frame either side of the truth.

3Classification: what kind of shot this even is

The same pass identifies the club category, the camera angle (down the line or face on), your handedness, and whether this is a full swing, a half swing, a chip or a putt. This sounds trivial and is not: it is what lets a library of hundreds of swings be filtered into something you can actually navigate, without you tagging every upload by hand.

4Coaching feedback: a language model reading the analysis

Finally a language model turns the above into a paragraph of feedback naming one priority and a drill. This is the least certain layer and should be read as a competent second opinion rather than a verdict — it sees one swing from one angle with no knowledge of your body, your injury history or what your coach told you last week.

5What it cannot measure from a phone video

No amount of vision modelling recovers clubhead speed, ball speed, spin rate, smash factor or launch angle from ordinary phone footage — those need a launch monitor or radar. Anything claiming those numbers from a single phone camera is estimating from assumptions. Depth is also weak from one camera angle, so lateral movement toward or away from the lens is the least trustworthy thing on screen.

Common questions

Is AI swing analysis accurate enough to be useful?

For body position, sequencing and comparing two swings at the same phase, yes. For precise frame-level timing, treat it as approximate. For anything involving ball or club speed, no — that is not recoverable from a phone video and should be measured with a launch monitor.

Do I need a special camera or a high frame rate?

A normal phone camera works. Higher frame rates genuinely help, because the difference between 30 and 120 frames per second is the difference between four frames through impact and sixteen. If your phone offers slow motion, use it — it improves every layer of the analysis at once.

Does my video get uploaded to a server?

For pose detection, no — that runs in your browser on your device. The vision analysis pass does require the video to be sent for processing, because the model runs server-side.

Can AI replace a golf coach?

No, and the honest reason is that a model sees one swing from one angle with no context. What it does well is the thing a coach cannot do between lessons: look at every swing you record, consistently, and remember all of them.

Last updated August 18, 2026