If you approach from the mathematics side of things, building on log₁₀ is the completely obvious approach to use. If it seems unintuitive, that’s just because you don’t understand logarithms.
> the autograding test cases did not include a test using a power of 10.
That’s a pretty glaring oversight. Boundary cases are probably the most important things to cover in such tests. I’d be wanting to test things like 0, 1, 9, 10, 11, 99, 100, 101, and so forth. (Also negative numbers, unless the problem explicitly constrained it to {0, 1, 2, …}.)
We often talk about edge cases in testing, but this reveals that the edges can be not just at the extremes, but also at intermediate value changes. Put another way (still fuzzy!), these are the interesting values. You test interesting values.
This would also be a good place to use property testing; instead of hard-coding a bunch of tests (or perhaps as well as), compare against a large number of generated values, comparing with a known-good implementation, probably just len(str(number)).
In a lot of cases, the most useful property is "for all inputs, the result is the right answer", but we don't know the right answer until we've written the algorithm, and there's not usually much point writing two algorithms unless, like here, one of those algorithms is trivial correct but inappropriate for some reason.
More generally, I feel like the more trivial a property is to check, the simpler the code to implement it is, and the less useful the test is in general. In the extreme case, a getter/setter pair is very easy to test with PBT, but you rarely need to be testing your getters and setters.