Roughly speaking, GRH asserts that the M\"obius function $\mu$ is "orthogonal" to all Dirichlet characters $\chi$, in the sense that correlations such as $\sum_{n \leq x} \mu(n) \overline{\chi(n)}$ are very small.  This is the expected behaviour of the M\"obius function, and through various standard analytic number theory manipulations can also control correlations between the von Mangoldt function $\Lambda(n)$ (which basically encodes primes) and various other functions, e.g. linear phases $e(\alpha n)$.  On the other hand, GRH alone struggles to control self-correlations such as $\sum_{n \leq x} \mu(n) \mu(n+2)$ or $\sum_{n \leq x} \Lambda(n) \Lambda(n+2)$ (the latter being used to count twin primes).  For instance, in the function field case GRH is a known theorem, but we are still unable to obtain an asymptotic (or even a lower bound of the right order of magnitude) for twin primes, though it is possible to establish the infinitude of twin primes in this case by more algebraic means.

In contrast, the existence of a Siegel zero means that there is a quadratic Dirichlet character $\chi$ with which $\mu$ has very high correlation (this can be quantified precisely using the "pretentious" approach to analytic number theory developed by Granville and Soundararajan).  This would be very unusual behavior for $\mu$, and as such is actually a rather powerful piece of information.  As a crude first approximation, a Siegel zero allows one to replace $\mu$ with $\chi$ with acceptable error (in practice one has to be more careful at small primes, though, for instance by imposing a suitable preliminary sieve).  Thus for instance one could hope to approximate $\sum_{n \leq x} \mu(n) \mu(n+2)$ by something resembling $\sum_{n \leq x} \chi(n) \chi(n+2)$, which is relatively easy to bound nontrivially.  *Very* roughly speaking, Heath-Brown's arguments proceed by similarly replacing the von Mangoldt function $\Lambda(n) = \sum_{d|n} \mu(d) \log \frac{n}{d}$ with the variant $f(n) := \sum_{d|n} \chi(d) \log \frac{n}{d}$, which is roughly of the same order of complexity as the divisor function; in particular sums such as $\sum_{n \leq x} f(n) f(n+2)$ are tractable enough by known methods (e.g. Kloosterman sum bounds) that they are a useful approximant to the otherwise intractable $\sum_{n \leq x} \Lambda(n) \Lambda(n+2)$.