I think you may be confusing two things here. The ANOSIM statistic R is a scaled measure of the effect size (separation) between groups, 1 = perfectly/completely separated, 0 = no separation, so you already have a measure of separation and I don't see why you'd want to use NMDS-space for anything. Remember, NMDS is only a representation of the underlying resemblances, so it is better to work on those resemblances (or their ranks) directly.

On the other hand, you can put confidence regions around group averages in a NMDS ordination as a multivariate anallogue of confidence intervals in a univariate plot, put in to facilitate interpretation of ANOSIM tests (as in, test on replicates, interpret using means). This is available in Primer v7.

As far as I understood, ANOSIM is a global test - it only tells me if there a overall difference among groups or not but it does not tell me how intense the similarity or dissimilarity between specific groups are? How can I then say, group A is more similar to group B than to group C? I saw publications where authors computed pairwise ANOSIM and reported R-values but this becomes exhausting if you have 16 groups to compare. I hope I could express what I'm aiming for more clearly now.

Depends how you are doing your ANOSIM. In Primer the pairwise tests are output automatically, and there's an additional option to send the pairwise R values to a separate file. Theoretically you could then use the pairwise R values to produce a different sort of nMDS showing how similar means of groups are to each other but it's not something generally done.

Thank you very much Paul Somerflield for your response. I submitted my manuscript with a different Approach now. Let's see what the Reviewer will say About it. Kind regards,

Nadine