The "random" sample is obviously very, very skewed: If you would be asked to name a random conjecture, it probably will be a "famous" conjecture, and the longer a conjecture stands, the more famous it tends to be.
But that is not my question. Has anybody tried yet a reliable statistic on truely "random" conjectures and the time they need for solving? One would need to pick random math manuscripts (already a challenge - how to avoid inherent skewing?), check whether they contain some open conjectures and check when they were solved later on. This means work, work, work (you would need some expertise in all part of math - or "outsource" the job).
My conjecture :-) is that conjecture longevity follows a Poisson statistic, and the open parameter is rather of the order years than centuries.
Addendum: The entirely non-random example :-) by Kuperberg about the G2 knot polynomial (hey, I just like the paper :-) lists six questions "open to the author" (incidentally showing that the longevity might also be zero, in the case that another mathematician already proved your question - happens all the time - I think such cases should better be excluded from the statistic). Judging from the many papers I DLed "0" is a much more probable number for open conjectures in a paper, though. (Needs a statistic, too.)