Resident lurking Electrical Engineer here. In my experience, it massively depends on what you’re doing.
For logic/comms stuff, the USB ones can be quite good. They tend to be higher performance per $ spent but work best when you’re not having to frequently adjust settings. They’re also typically less hassle to integrate into other stuff and automate using python scripts or whatever, but that usage may be a little more ‘niche’ from a hobby perspective.
For general use, I would always strongly recommend starting with a cheap standalone unit. You’ll be constantly adjusting timebase, channel gain, voltage offset, trigger level and direction, etc. to hunt for the parts of signals that you’re looking for. It’s a tool used for insight so making it easy and quick to gain those insights is the name of the game, really.
There are a few key specs to watch out for, especially with the cheaper ones. In general, I say to buy more bandwidth than you need because ideally you want it to be 10x higher than the fundamental signal you’re looking at. This is because any signal that doesn’t look like a sine wave has much higher frequency components in it. A 10MHz SPI bus has plenty of 100MHz signal components if you want the edges to look ‘crisp’, which you do if you’re trying to find out why your SPI bus isn’t working and it may be because you’ve got glitches on those rising/falling edges.
Sample depth is hugely important and in genera more is always better. This lets you take a longer capture once triggered and then ‘zoom in’. It’s kinda like the equivalent of having a higher resolution camera and being able to zoom digitally. You don’t want to have to frame everything up perfectly every time, so deeper sample depth lets you ‘see’ more. I would aim for something in the mega-point range, but
Waveform capture rate/update rate is another huge one. This is basically how often the scope can trigger. It can be pretty important for making the tool easy to use, especially if you’re looking for faults. Have a low update rate means that you may only see every 10th time the scope triggers, so if something happens rarely, you’re not likely to see it.
There are also a bunch of other useful features that it might be worth considering, like digital channels, comms bus decode, extra channels, some advanced triggering options, on-screen FFT performance, waveform generation output, that kind of thing. Some of those will ever be as good as a standalone unit so we don’t tend to focus on those features as much on the more expensive end, but on the cheaper end it can be quite useful because you’re less likely to buy a standalone waveform generator, for instance. On the other hand, a cheap USB logic analyzer may be a better bet for less money than adding it to the scope. The best thing about a mixed signal scope vs an analog scope and separate digital logic analyzer is that you get all the information on the same screen and synchronized which can be useful.
It’s also worth being careful because at the low end, especially if you stray from the ‘known’ brands, it’s possible to save a small amount of money by losing a LOT of capability.
Another thing that’s worthwhile considering is that with scopes a lot of features are software locked, needing a license to unlock them. For certain models there are easy and safe ways to unlock these features without paying for them.
As for recommendations, Rigol is definitely the standard for gear on the low end. The 1000 series has some great units in it.
The DS1054Z is usually the de facto recommendation for a good entry level scope. 4 channels is nice, 50MHz is what I’d consider the bare minimum and the memory depth is pretty decent. Waveform capture rate is good. It’s hackable to get to 100MHz bandwidth.
The Siglent SDS1104X-E is also a 4 channel scope but with 100MHz bandwidth which is enough for the majority of situations. It also has good memory depth and an excellent waveform capture rate. It has better analog electronics so would be my pick over the Rigol, assuming the same price. I believe it’s also easily hackable to get to 200MHz bandwidth and I think some extra decode capabilities.
The GW Instek GDS-1054B is also a similar scope to the Rigol but at a cheaper price. 50Mhz, 4 channel, good memory depth, good waveform update rate, well regarded.
The EEVBlog community spends a LOT of time debating this kind of stuff, so that’s where I’d go if I were trying to decide between them. I try to avoid the place because I find the discourse a little bit too ‘aggressive’ and personal for my tastes. A few too many ‘engineering alpha wannabes’ crammed into the same space, I think 
. Guess i have a lot more to learn 