In your case, the statistical model is not properly reflecting your conceptual model because the statistical model is actually proposing a three variable serial mediation. Your conceptual model, instead, only has two mediating paths. This would be the same as treating ITC/SMM as a covariate for the paths to TW and JS if PROCESS can handle it (I don't use Hayes PROCESS so it's not a definitive answer).

To build complex models, I'd recommend instead using STATA (using sem) or R (using lavaan). In both cases, you can either create latent variable models or use aggregate scores as manifest variables in a path model.

If you prefer point-and-click like SPSS and would rather not code, you can use the sem function in JASP (https://jasp-stats.org/2018/07/03/how-to-perform-structural-equation-modeling-in-jasp/) to run mediation path models.

Or you can also use jamovi (https://www.jamovi.org/) and get the jAMM package from their library to run GLM mediation models (https://jamovi-amm.github.io/). For both JASP and jamovi, you can easily create complex models by dragging variables into specific paths.