I define a quaternion to be a sum of a scalar and a 3D vector (from $\mathbb{R}^3$). You can also call the scalar and vector parts the real and imaginary parts, respectively. The real inner product is $\langle x,y\rangle=\mathrm{Re}(x\overline{y})$, which is the usual dot product $x_1y_1+x_2y_2+x_3y_3+x_4y_4$. (There are different ways to build up to this, each with its own pedagogical merits, but I'll just leave it as a given you can check.) As a result, multiplying on the left or right by a unit quaternion is an isometry. In particular, if $p$ is a unit quaternion then $p^{-1}=\overline{p}$, and then
$$ \langle pxp^{-1},1\rangle=\langle px\overline{p},1\rangle=\langle px,p\rangle=\langle x,1\rangle $$
so $pxp^{-1}$ is a vector (aka pure imaginary) if and only if $x$ is.
The representation of quaternions by $2\times2$ complex matrices is a separate thing. You can transport the conjugation action over hear, if you want, though. For this, the unit quaternions become $\mathrm{SU}(2)$, the pure imaginary quaternions (real 3D vectors) become the lie algebra $\mathfrak{su}(2)$, which is the set of complex matrices satisfying $X^\dagger+X=0$. Then $SU(2)$ acts on $\mathfrak{su}(2)$ by conjugation, and it is a very quick exercise to check that if $A^\dagger A=I$ and $X^\dagger+X=0$ then $(AXA^{-1})^\dagger+(AXA^{-1})=0$ (keep in mind $(UV)^\dagger=V^\dagger U^\dagger$ and $A^\dagger A=I$ implies $A^{-1}=A^\dagger$). This is called the adjoint representation in Lie theory.