The tutorial that you link to does not explain what "translating" between Mealy and Moore machines actually means, which is however crucial. In every step of its computation, a Mealy machine first reads the next input character, and then produces the next output character while transitioning to the next state. A Moore machine, one the other hand, first produces the next output character and then reads the next input, after which it transitions to the next state.
These two modes of computation are different and hence there is no translation between them that does not alter the semantics of the machine. When you translate from Moore to Mealy, you can build a machine that ignores the next input, produces the next output, and only then takes input and state into account to transition to the next state. The resulting Mealy machine in a sense ignores the next input until it wrote the next output.
When you translate from Mealy to Moore, you have a problem: to compute the next output, you need the next input....which is only available to the machine after producing the next output. What you can do during the translation is however to build a Moore machine that produces its output delayed by one step. This requires you to store the last output in the state component, which blows up the automaton. Whether the resulting Moore machine is good for anything practical is a different question. Even with this semantics-altering translation, there is still the problem that the Moore machine needs to give some initial output, which needs to be defined. This could be an arbitrary one or a designated "no data" output character. This seems to be the choice of your tutorial - rather than having more than one initial state, you declare the one to be initial that is labeled by the output character representing ``nothing''.
Having to choose some initial output shows you that this is no one-to-one translation.