Why is Peano arithmetic undecidable?

Essentially, yes, but I believe the undecidability of Peano arithmetic (henceforth, PA) follows from the way the proof of Gödel's incompleteness theorem goes, rather than being a consequence of the fact that PA is incomplete. The proof (outlined below) starts by showing that PA can talk about computable relations, and goes on to show from this how you can construct an unprovable sentence. However, we can take a different approach to show that PA is undecidable: if PA can talk about computable relations, then you can formulate a sentence in the language of PA that is true if and only if a given algorithm halts / does not halt. (Briefly: An algorithm is the same thing as a computable partial function, and an algorithm halts on some input if and only if the corresponding partial function is defined on that input.) So if an algorithm can decide whether arbitrary sentences in PA are provable or not, we have an algorithm which solves the halting problem... but there are nonesuch. So the key point is that PA is rich enough to talk about computation. An essential ingredient for talking about computation is the ability to encode a pair of numbers as a number and the ability to recover the original pair from the encoding. It's not immediately clear to me that $+$ is insufficient to do this, but it's certainly plausible that you need at least two binary operations.

Here is a sketch of the proof of Gödel's first incompleteness theorem: First let's select a sufficiently powerful theory $T$, e.g. PA, and assume that it is a consistent theory, i.e. does not prove a contradiction.

1. We show that we can encode formulae and proofs in the models of $T$.
2. We show that $T$ is powerful enough to talk about computable relations.
3. We show that there is a computable relation $\mathrm{Prf}(m, n)$ which holds if and only if $m$ encodes a valid proof of the sentence encoded by $n$.
4. The above shows that there is a computable relation $Q(m, n)$ which holds if and only if $n$ encodes a formula $F(-)$ with one free variable and $m$ does not encode a valid proof of $F(n)$.
5. So we can define a formula $P(x)$ by $\forall m. Q(m, x)$. This means $P(x)$ holds if and only if there is no valid proof of $F(x)$, assuming $x$ encodes a formula $F(-)$ with one free variable.
6. But $P(-)$ is a formula with one free variable, and it can be encoded by some number $n$. So is the sentence $P(n)$ provable or not? Suppose it were. Then, that means $P(n)$ is true, so the theory asserts that there is no valid proof of $P(n)$ — a contradiction.
7. So we are forced to conclude that the sentence $P(n)$ is not provable. This is the Gödel sentence which we wished to prove the existence of, so we are done.

Note I haven't said anything about whether $P(n)$ is actually true. It turns out there is some subtlety here. Gödel's completeness theorem tells us that everything that can be proven in a first-order theory is true in every model of that theory and that every sentence which is true in every model of a first-order theory can be proven from the axioms of that theory. With some stronger consistency assumptions, we can also show that $\lnot P(n)$ is also not provable in PA, and this means that there are models of PA where $P(n)$ is true and models where it is false.

The key point is that the phrases "$n$ encodes a formula ..." and "$m$ encodes a valid proof of ..." are strictly outside the theory. The interpretation of a particular number as a formula or proof is defined externally, and we only define it for "standard" numbers. The upshot of this is that in a model $\mathcal{M}$ of PA where $P(n)$ is false, there is some non-standard number $m \in \mathcal{M}$ which $\mathcal{M}$ "believes" is a valid proof of $P(n)$, but because it's non-standard, we cannot translate it into a real proof.

While this may not be the intuitive idea you're looking for, I'll note that the same ideas behind Godel's proof that Peano Arithmetic is incomplete (that is, that there are true sentences of PA that can't be proven) can also be applied to show that Presburger Arithmetic is 'hard' in a tangible sense; essentially, one can encode multiplication of small enough numbers (specifically, the predicate '$x\times y = z$' with $x$, $y$ and $z$ 'only' doubly-exponential in the length of the formula) into a formula involving only addition. This means that if sentences in Presburger Arithmetic of length $n$ could be decided without multiplying numbers of size at least (roughly) $2^{2^n}$, then Presburger Arithmetic would suffer from the same Godelian paradox that Peano Arithmetic does and be incomplete. Paul Young wrote up a much better exposition of this for a conference a couple of decades ago than I can possibly give; have a look at http://books.google.com/books?id=2vvg3mRzDtwC&pg=PA503&lpg=PA503 for the details.