Debs,
I am in the position of doing similar "social skills" groups with two groups of teenagers - kids with Specific Language Impairment and kids with Autism.
Typically, the students with SLI typically don't have difficulties with generalisation but, having had years of difficulty with communicating, do have social communication difficulties. Some have particularly compromised verbal ability which they are acutely aware of and some have extreme difficulties getting a message across relative to my students with autism.
I have had some of these students who also have subtle and not-so-subtle pragmatic difficulties, difficulties with understanding facial expression/intonation etc, picking up on cues, processing "real time" conversations etc. I would say their abilities are on a continuum.
The group of teens I work with who have ASD are mainly verbal and most (though not all) have age appropriate concrete language skills. Some of them are actually pretty good at conversations too. I can think of at least 3 of 15(ish) who would be unlikely to be "spotted" as having social communication difficulties by someone who knew nothing about the spectrum. Some have very limited social interest and much more impaired verbal ability. Some have impaired verbal ability and social interest. A very mixed group.
Working with these groups, there are stark differences in practice with reference to generalisation. The group with SLI don't need as much structure or explanation of aims as the group with ASD - typically, their groups are much more fluid and might involve each student having a specific target or targets related to improving their group communication e.g. make eye contact when trying to get someone's attention, write down a word/use a VOCA if the communication breaks down etc. They know them, they get talking, they track it. When a new target is set, they would receive a reminder if not using their target initially 'in the moment' but very quickly after 1 minute, after 3 mins, after 5 etc.. it's easy to reduce the cuing. In this group, the students are also quite good at tracking eachother's targets "in the moment" - so at the end of a 10 minute discussion about, say, a recent football game, each student is pretty good at giving feedback to their peers about what they could have done to help them e.g. "you needed to slow down, I couldn't understand you when you were talking about x" etc (albeit in less clear language than this!).
The group with ASD can also provide this type of feedback with the right structures but it is a good deal more complicated e.g. we are using visual cues as above and video feedback, too. It seems that it is just "too much" for them to try and work out what they are doing in the moment (as Amberlight so eloquently describes) and also think about what others are doing. And the peculiarities of each interactive, dynamic conversation changes the learning "load" so to speak so having a general awareness of a target just doesn't work as well for them as it does for the group with SLI: there is a lot more prompting/explaining/direction in this group.
In addition, even when it does work, peer-to-peer feedback needs to be carefully managed as it can be more abrupt and create tension where two minds don't meet so systematic rating systems tend to be more helpful than relying on student reflection alone e.g. giving eachother marks out of 5.
We use Comic Strip Conversations with both groups where there is significant "trouble" that we need to discuss - here, again, the differences between the groups are stark, as for the students with SLI, simply "working through" the conversation on paper is often enough to prevent a recurrence of the difficulty. Any minor change in variable (which, of course, is to be expected in something as dynamic as interaction) and the student with ASD can find it very, very difficult to apply.
Ideally, I would like to see the supports/structures for the group with ASD being gradually removed.. e.g. I am working towards them being able to, first of all, track their own contribution to a conversation with reduced visual cues and then track those of others, but I'll be honest with you, I don't know what is achievable in terms of generalising real changes to "online" conversation. If it were an ABA programme, I guess I would say that we are stuck at the teaching stage a lot of the time because we just can't control the variables of a real conversation well enough to do the systematised generalisation that works well with other skills. The vast majority of my students understand the mechanics of conversation quite well and can tell you all about the importance of taking turns etc... but that doesn't make it much easier in a real conversation. So, the lessons are really an ongoing lesson in the "exceptions to the rules" which are, well, pretty constant.
Even where students are showing good learning of the target e.g. realising that they were about to interrupt and telling me that they nearly did it, this reflection in itself detracts from their ability to listen to and contribute to the ongoing conversation which in itself reduces generalisability.
It makes sense: if you had to surgically analyse the minutiae of how/why a conversation was working, how would you manage it and communicating your message at the same time? This is one of the reasons that I prefer to take a "talk, analyse, talk" approach - we start talking, we keep talking, the videoing/visual cues go down but don't need to be talked about or reflected on until the conversation topic ends.
Will it make a long-term difference? I really, really don't know. I work hard at trying to make it do so, but I don't think anyone has the answer when it comes to conversation. I don't think we know enough and, as Amberlight suggests, there is an automatic integration of such an array of skills in typical adult conversation that is difficult to replicate in intervention.
Aside from all that, with the tools we have now, as moondog suggests, the key for both groups, in terms of generalisation, is follow through by regular school staff pretty much constantly: everyone needs to sing off the same hymn sheet, as it were.