A useful prediction is a prediction that matches experimental results.
And I would say that Newtonian gravity was a useful theory despite not matching experimental currently because when it was the best it mostly explained observations.
"Useful" in this context means useful for a future theory or experimentation, not utility. Newtonian theory certainly predicted many results, and all of them were useful in the above sense. A prediction doesn't have to exactly match observations to be useful, but that the theory produces a prediction which _could_ be tested against.
Even flat-earth theory makes useful predictions - albeit already proven false as time and time again, it predicts the wrong results.
What isn't useful is an unfalsifiable theory, which means the predictions it makes is not able to be tested, or the results of such a test could be construed to match if you squint a bit. A theory like creationism, for example.
> Even flat-earth theory makes useful predictions - albeit already proven false as time and time again, it predicts the wrong results.
I definitely agree usefulness is independent of fitting experimental data. gliop would seem to say flat earth theories are not useful.