I think the answer is: neither, since this GGX distribution is NOT sampling the direction for a ray but a microfacet normal. According to the paper of GGX: Microfacet Models for Refraction through Rough Surfaces, this distribution is used to sample the normal direction for the microfacet, which you can think of as sampling a new normal direction with respect to the original normal direction (macroscopic). Therefore, the coordinate frame is spanned by the original macroscopic normal, for example, we use the macroscopic normal as local z-axis, then the $\theta$ here is the angle between the newly sampled microfacet normal and the macroscopic normal, as shown in the following figure.
So, how do you get the direction of the light? First, we get the normal we want just by using GGX, to sample a microfacet normal. Then, maybe we will start to sample surface event like whether we evaluate the diffuse part or the specular part. Take evaluating specular part as an example, now we have a incident direction, a microfacet normal, then we can directly calculate the outgoing direction.
Also, can I use this θ directly to calculate the BRDF?
Yes. Sampling an outgoing direction will also require you to calculate the path throughput. This throughput coefficient is calculated using the parameters you've sampled.
So, in all, sampling a new ray direction is usually pretty local around one specific normal (geometric / shading / sampled microfacet normal), so we will usually set the frame according to the normal. However, this is not the case when you are dealing with medium scattering (phase function). In this case, the sampling is usually done in the frame spanned by original incident ray direction.