Thanks to @user26857, for his great hint.
Assume
\begin{equation}
f(x) = a_n x^n + a_{n-1} x^{n-1} + \cdots + a_0 , \quad
g(x) = b_m x^m + b_{n-1} x^{n-1} + \cdots + b_0 ,
\end{equation}
and $n \ge m$, $a_n \ne 0 \ne b_m$.
We have $\text{LM}(f)=x^n$ and $\text{LM}(g)=x^m$.
We call $L= \text{LCM}[ \, \text{LM}(f), \text{LM}(g)] = x^{n}$.
Then by definition
\begin{eqnarray}
S(f,g) &=& \frac{L}{\text{LT}(f)} f - \frac{L}{\text{LT}(g)} g \\
&=& \frac{x^n}{a_n x^n} f - \frac{x^n}{b_m x^m} g \\
&=& \frac{1}{a_n} \left (
f - \frac{ a_n x_n}{b_m x^m}g \right) \\
&=& \frac{1}{a_n}
\left (f - \frac{\text{LT}(f)}{\text{LT}(g)} g
\right )
\end{eqnarray}
On the other hand the first step on the Euclidean algorithm is the
division of $f$ by $g$. This step is achieved by finding the
remainder
\begin{equation}
r_1 = f - \frac{\text{LT}(f)}{{\text{LT}(g)}} g .
\end{equation}
Setting the $1/a_n$ aside (up to a multiplication by scalar any non-zero
multiple of the GCD$(f,g)$ is also Gr\"{o}bner basis member)
we have that $S(f,g)=r_1$. If $\text{deg}(r_1) < \text{deg}(f)$ we are done with this remainder, otherwise we keep reducing it until $\text{deg}(r_1) < \text{deg}(f)$.
The Buchberger algorithm starts with $G=\{f, g \}$, and then,
if $r_1 \ne 0$, add (append) $r$ to $G$.
That is $G = \{f, g , r_1 \}$. The next step is to reduce $r_1$
with respect
to $G$. Since $\text{deg}(r_1) < \text{deg}(f)$, we only need to reduce
$r$ with respect to $g$. That is, we divide $r$ by $g$, and keep doing
this until $r_n=0$. (We do not prove here that the algorithm finishes. It does.
That is not the purpose of this exercise). Then
$G=\{f, r_1, \dots, r_{n-1} \}$. We find that
$\text{GCD}(f,g)=r_{n-1}$, and it is found as if we were doing Euclid's algorithm instead of Buchberger's algorithm.
That $G$ can be reduced to $G=\{r_{n-1} \}= \{ \text{GCD}(f,g)\}$ is another story.