I think it's the coalescence with its sister species and selecting several cultivars or varieties in every species tested , and you'd better find a age-known species in the same genus or family as the reference to estimate your species's age.

I think your suggestion to use the two numbers as estimates of the upper and lower bounds is a good one. Extinctions will bias both of those bounds toward extremes: using the age of coalescence with a sister species will give an age that is too old if a closer sister species existed but is now extinct; using the deepest coalescence between populations within the species of interest will give an age that is too young if other populations (with older lineages of alleles) once existed but are now extinct. I agree with Luxia Yuan that you will need a good mutation rate calibration. *BEAST includes good methods for the kind of estimation you are proposing, but it assumes that there has been no gene flow among the populations or species. IMa2 is a more complex (and realistic) population model because it can account for gene flow effects in its estimates of divergence times, but requires a lot of loci to fit the model to more than 2 population samples.

I like the idea of using these as the upper and lower limits. I think, however, that this question really depends on what you mean by the age of a species. If you are thinking in terms of anagenetic character change (with the species defined by some character based criterion such as morphological traits, mate-preference attributes or the fixation of mutations that cause post-zygotic incompatibility) then it is a murky problem. If it is simply about genealogical splitting from the sister species then it is easier. Divergence among sister lineages over-estimates the time of divergence by not accounting for gene-diversity and coalescence within the ancestral population. The correct distance is approximates by Lynch & Creases' (1990) concept of net nucleotide divergence (nucleotide divergence among populations - average nucleotide diversity within each population) . Although as Michael Hart suggests it would be best to estimate this using coalescent methods implemented in IMa2 or BEAST.