It will work as a rule of thumb, what you have to consider is the method used by arcmap to calculate flow accumulation. It is based off the D8 method, which assumes that 100% of a cell's input flows out to one of the 8 surrounding cells at the compass points: N, NE, E, SE etc.
This works as an approximation but you can't say that such a model would reflect the true catchment areas, as it is quite common for water flow to not follow 45 degree angles and to branch and recombine as it flows across the surface. This is one of the limitations of trying to model a landscape in a raster grid, but one which is being overcome to some degree with the use of LiDAR data and more complex analysis techniques which trace vectors of probabilistic water flow between cells based on work by Lea (1992).
General reading about these hydrology techniques should always include work by David Tarboton, especially his 1997 paper, which has been the groundwork for most DEM based hydrology for the last 15 years. If you are more inclined towards programming (particularly c++) you could also check out Jon Pelletier's book which has great algorithms of all the classic techniques in hydrology and DEM processing in easy to follow c++ samples.