Good question as to what Geni has access to and what Geni displays (right :) ) I will try to be more detailed here and put worst case hat on: Let's say Geni has access to only what it displays and let's run an example. (akin to running it on summarised data therefore)
I will digress first into proportions.
I currently have over 400 atDNA matches, ranging from my daughter who shares 40-60% but the longest (continuous) segment between us is only +- 5%. However, the next close cousin of mine only shares +-5% but here the longest continuous segment is 1%, a much larger proportion of total shared. (approx 10% vs 20%.)
The continuous proportion as a percentage of shared becomes much larger the more distant the relationship (over all the 100s of my matches, even getting close to 100% (i.e. the only common DNA therefore)) and therefore it is the measure which I believe Geni should target if employing this method. But it comes with a caveat...
So if my daughter and I share continuous 5%, then it means that one of my parents has/had more than 5% continuous DNA "exactly like that" with 100% certainty but 50% certainty until triangulation via children of my parents' brothers and sisters that could or could not be this exact 5% if the children matches to my daughter. As you can see, because the continuous proportion is so small for near relatives, this becomes distraught with difficulties.
However, for the 100s of distant matches where the continuous proportions are large, it becomes more and more certain that the continuous proportion is in fact the small proportion of the actual DNA of an ancestor. Now the problem becomes a bit different, in the sense that if there are more than one set of common ancestors within 8 generations, it is no longer just 50/50 but perhaps even 10%, but with the benefit it is the exact DNA of which ancestor. (With 8 generations the set is 2^8)
So that is why I proposed 4 generations... it gives higher reliability between what exact DNA it could be, and reduces the risk that the DNA is attributed to the wrong pair of ancestors. (Also to consider in future: Over the past 4 generations on earth, an individual having more than one set of 2nd grandparents has become remote, but that was not always the case - I have not factored this in below, effectively it should be 2x reliability :).)
For most of the matches I have in this range (common ancestors 4-5 generations back), where the genealogy part is sound, the longest continuous segment seems to be around 25%+ of the total match.
So getting back to how exactly... lets say X and Y performed atDNA tests and match within range of whatever measure of relatedness is used. If their Geni trees are complete and only part of the OWT, up to 8 generations, then I have some confidence Geni is able to accurately determine the path to their closest common set of ancestors. Geni then performs a check whether this is within 4 generations, and that there are no other path within 4-6 (say) generations. Geni then checks the % of the longest continuous match and adds an atDNA display on both sets of the nearest couple of ancestors each of 40% (the lowest, i.e. we "lose" 20%) times this proportion% times 4 (1/25%) on that common couple and ADDS with graphical display of the path (similar to the graphical display currently for Y-DNA and mtDNA) and for the two distinct unique paths to that set of ancestors propagates this specific indicator.
Now when Z performs an atDNA test, and matches to both X and Y and also only shares this same set of ancestors 4 generations ago, one should of course expect different proportions, however the smallest proportion (length) over X, Y and Z then becomes an even more reliable indicator that this is part of the exact DNA of one of those common set of ancestors and the proportions and reliability should be updated.
And now B performs an atDNA test and does match to X,Y, and Z, but does not share the same set of common ancestors, however, on the fifth generation she does. For me that would clearly show that the atDNA attributed to the ancestors of X, Y and Z's 4th generation parents can therefore be attributed to the 5th generation parents of the ancestor in the 4th generation which also shares B's, and we can therefore remove the 40% penalty applied initially to the calculation, AND get more certainty on the EXACT likely subset of DNA of the 5th generation ancestors... and with more and more testers and common ancestors, suddenly we have an algorithm that should be able to assist users tremendously...
Even though a pilot, I believe there would be numerous secondary benefits just from a genealogical usage (of Geni) perspective as well, a more complete OWT as one benefit, but which I am not discussing here. However through trial and error (with how many generations to go back) I am sure something great can be employed on Geni.