The best reason I know (other than for the fun!) for going beyond Google Analytics/Matomo-Piwik with one's web analytics methodology is ... for the money!
I took my first step about a decade ago, and worked my way slowly but steadily, well over my head, ever deeper. :)
However, as various studies and white papers have found in the past year:
* the greatest impact has been derived by the the travel vertical (third in value increase) and the greatest value by the retail industry (tied for second in impact).
* 'advanced' analytics, on average, across ~400 cases in 20 industries, added ~5% to the bottom line, machine learning based analytics another ~2%.
Note: the studies are all? of enterprise level sites.
Of course, 5-7% of $1,000 or even $10,000 probably isn't sufficient ROI. However, once one gets up in the $100,000+ (and there are a lot of quiet webdevs at and above that threshold)...
Note: I find those results to be low - unfortunately, of course, none of the studies detail wherefrom, what was changed, where to...
Plus, of course, there's the magic of:
It went "Zip" when it moved
And "Bop" when it stopped
And "Whirrr" when it stood still
I never knew just what it was and I guess I never will
---The Marvelous Toy by Tom Paxton
I've been playing for a few years now with LSTM-RNN (long short-term memory recurrent neural networks), which are good for modelling data and making predictions. Currently, this is my model creator methodology. Originally feeding heuristic algorithms; more recently, initiating GANs (generative adversarial networks) aka two neural nets in a zero-sum contest framework (wherefrom the adversarial) that take the base and predictive models plus live streaming data to determine whether a visitor is a bot, best (goal variable) contextual content match, which ad to show where, etc. All at sub-millisecond speed. Live, in real time, analytics in the driver's seat, not (just) the back seat.
While the above is fun (for me) and has delivered solid quantifiable profitable outcomes machine learning wasn't the huge leap it might have been as I'd already been using some of the more advanced web analytics methods. As noted above, if one has already invoked advanced analytics the machine learning may be a relative small increment, except that it can, as in my setup, (mostly) automagically simply continue on it's own. Note the mostly :) There is definitely a huge decrease in manual intervention (time!) required and that is a substantial value whether to life style or bottom line.
Regardless, it's really just a matter of classification (some descriptive model identifying many different relationships a la pattern recognition) handled via scripting and spreadsheets or one step further on to software algorithms or a second step to simple machine learning or a third...
Note: it helps to have a knowledge of mathematics and an understanding of statistics. Even with bog standard GoAn.
In play test at the moment is RL (reinforcement learning). You've probably heard of supervised and unsupervised learning; reinforced learning is the new third option. Neither inputs or outputs need be labeled nor less than optimal actions corrected. It learns via 'reward' or 'regret' (or both) feedback that updates state until an optimum is discovered.
Note: see 'multi-armed bandit problem.
Note: my current framework of choice is a UCB-ALP (Upper Confidence Bound-Adaptive Linear Programming) algorithm. It shows a lot of promise.
Note: I almost know what I'm doing. Now and then.
Come join me.
Out on the sharp edge.
Beware of nose bleeds.