…according to an article at MedPageToday. Briefly, carb counting involves estimating the digestible carbohydrate grams in a meal (often called net carbs), then dosing rapid-acting insulin based on those grams and the individual’s prior responses to insulin. It turns out there’s not a lot of hard clinical evidence to back up the practice. A quote from the article:
Carbohydrate counting is the best known method for matching insulin dosing to meals, and is the recommended dietary strategy for achieving glycemic control in type 1 diabetes, though that recommendation has been largely based on expert consensus, Bell said.
One commentator said it doesn’t work very well because most folks aren’t very good at it, they’re not vigilant enough. Why do we so often want to blame the patient?
A review panel “compared carbohydrate counting with usual care, which consisted of either general nutrition advice or low dietary glycemic index (GI) advice.” They found no significant differences in hemoglobin a1c between the approaches.
Paleo Diabetic 
I would suggest that when carb counting doesn’t work it is because the carb count is too high.