The following solution is a special case of a general method.
The question asks
Let $a_ 1=\sqrt{6}$, $a_{n+1}=\sqrt{6+a_n}$. Find
$\lim_{n \to \infty} (a_n-3)6^n$.
With $\,q\,$ as a parameter, define the function
$$ F(x) := \sum_{n=0}^\infty c_n \frac{x^n}{f_n}\;\;
\text{ where }\;\; f_n:= \prod_{k=1}^n (1-q^k) \tag{1} $$
and where $\,|q|\ne1.\,$ Define the constants
$$ L := q/2, \quad \text{ and } \quad K := L^2-L. \tag{2} $$
Also define
$$ a_n := A\left(\frac{x_0}{q^n}\right) \;\;
\text{ where } \;\; A(x) := L - q\, x\, F(x) \tag{3} $$
and where $\,x_0 = A^{-1}(a_0)\,$ depends only on
$\,q\,$ and $\,a_0.\,$
Tthe value of $\,x_0\,$ computed this way does not
suffer from floating point rounding problems.
The equation
$$ a_{n+1} = \sqrt{K+a_n} \quad \text{ or } \quad a_{n+1}^2 = K+a_n \tag{4} $$
implies that the coefficients of $\,F(x)\,$ as polynomials in $\,q\,$ satisfy
$$ c_0 = 1, \quad c_{n+1} = \sum_{k=0}^n c_{n-k}\,c_k\,
\frac{f_n}{f_k f_{n-k}}. \tag{5} $$
Note that
$$ F(0) = 1,\; A(0) = L \; \text{ and }\;
a_n\to L. \tag{6} $$
The approach of $\,a_n\,$ to the limit $\,L\,$ is given by
$$ L - a_n = q \, \frac{x_0}{q^n}
F\left(\frac{x_0}{q^n} \right)
\approx q\frac{x_0}{q^n}. \tag{7} $$
This implies
$$ \lim_{n\to\infty}(a_n-L)\,q^n = -q\,x_0. \tag{8} $$
In the case in the question,
$$ q = 6,\; L = 3,\; a_0 = 0,\;
x_0 \approx 0.56094292329064. \tag{9} $$
In the case from Art of Problem Solving Online,
$$ q = 4,\; L = 2,\; a_0 = 0,\;
x_0 = \frac{\pi^2}{16},\; F(x) = \frac{\sin(\sqrt{x})^2}x. \tag{10} $$
NOTE: About my method.
Some of it is based on Koenigs function but my version is more constructive.
If we have a function $\,T(x)\,$ and define a sequence by
$\,a_{n+1} = T(a_n)\,$ where $\,a_n\to 0\,$
such that $\, a_n \approx c/q^n,\,$ then use the Ansatz
$\, a_n = F(x/q^n)\,$ for some function $\,F\,$ with a power
series expansion in $\,x\,$ with coefficients that depend on $\,q.\,$
The coefficients are uniquely determined by the function $\,T.$
The mode of convergence determines the proper Ansatz. For
example, $\,T(x) := x-x^2\,$ requires a different Ansatz.
NOTE: If $\,q=1\,$ then the convergence $\,a_n\to \frac14\,$ is
much slower and the above analysis does not hold. Instead the one
for $\,T(x) := x-x^2\,$ is needed. Something similar if $\,q=-1\,$
and other roots of unity.
NOTE: I was going to mention chapter 8.3 of Asymptotic Methods in Analysis
by de Bruijn which I am familiar with but didn't have the exact page number.