Short answer: A morphism of sites as you define it (a functor which preserves covers and the fibre products showing up in the gluing condition) gives rise to an adjunction $$f_s:\mathrm{Sh}(\mathbf{C}) \leftrightarrows \mathrm{Sh}(\mathbf{D}):f^s$$ of sheaf categories. The condition of covering-flatness ensures that $f_s$ is exact. (The resulting structure is known as a geometric morphism of topoi.)
EDIT to clarify subscript/superscript Notation: In the classic example of a map $f:X \to Y$ inducing a morphism of sites $f^{-1}:\mathrm{Op}(Y) \to \mathrm{Op}(X)$, the functors above are the direct image functor $${(f^{-1})}^s = f_\ast$$ and the inverse image functor $${(f^{-1})}_s = f^\ast.$$ ("The ${(-)}^{-1}$ exchanges subscripts and superscripts.") This notation is commonly used in the general case. The adjunction $(f_s \dashv f^s)$ is then written instead as $(f^\ast \dashv f_\ast)$.
Long answer: A functor of small categories $$f:\mathbf{C} \to \mathbf{D}$$ gives rise to an adjoint pair of functors between presheaf categories $$f_p:\mathrm{PSh}(\mathbf{C}) \leftrightarrows \mathrm{PSh}(\mathbf{D}):f^p.$$ The functor $f^p:\mathrm{PSh}(\mathbf{D}) \leftrightarrows \mathrm{PSh}(\mathbf{C})$ is given by composition with $f$. It preserves limits because limits in presheaf categories are computed pointwise. We may build its left adjoint $f_p$ explicitly: it is given by left Kan extension along $f$. The functor $f_p$ preserves finite limits if and only if $f$ is representably flat [nFF, Prop 2.6]. (The left Kan extension is pointwise computed by a colimit. Representable flatness ensures that these colimits are filtered. Finite limits commute with filtered colimits.)
Now assume $\mathbf{C}$ and $\mathbf{D}$ are sites. Sheafification and inclusion over $\mathbf{C}$ form an adjoint pair $$L_\mathbf{C}:\mathrm{PSh}(\mathbf{C}) \leftrightarrows \mathrm{Sh}(\mathbf{C}):I_\mathbf{C}$$ and similarly $(L_\mathbf{D} \dashv I_\mathbf{D})$.
If $f$ preserves covers and the fibre products present in the gluing condition, the functor $f_p$ preserves sheaves: the composite $$f^p \circ I_\mathbf{D}:\mathrm{Sh}(\mathbf{D}) \to \mathrm{PSh}(\mathbf{D}) \to \mathrm{PSh}(\mathbf{C})$$ lands in the subcategory $\mathrm{Sh}(\mathbf{C}) \subset \mathrm{PSh}(\mathbf{C})$. It has a left adjoint $$L_\mathbf{D} \circ f_p: \mathrm{PSh}(\mathbf{C}) \to \mathrm{PSh}(\mathbf{D}) \to \mathrm{Sh}(\mathbf{D}).$$ The adjunction restricts to an adjunction of presheaf categories $$f_s:\mathrm{Sh}(\mathbf{C}) \leftrightarrows \mathrm{Sh}(\mathbf{D}):f^s.$$ The functor $f_s$ is the composite $L_\mathbf{C} \circ f_p \circ I_\mathbf{D}$. We're again looking for a condition that will make this functor exact. As a left adjoint, it already preserves colimits.
Now comes the punchline: Covering-flatness of $f$ is equivalent to the requirement that the composite $L_\mathbf{C} \circ f_p$ preserve finite limits [nFF, Prop 2.15]. (The reason is the same as in the story for presheaves, except the fact we sheafify allows us to use a slightly weaker condition than representable flatness.) This means $f_s$ is exact iff $f$ is covering-flat.