In general, lambda, which is called the smoothing parameter, is a weight takes any value between 0 and 1, and statistical estimation methods can be used to estimate its value.
Beside the answers already given, you need to have a clear goal in mind. For different research, sometimes you would like accuracy, other would like conservatism (especially for risk calculation). But at the end, your goal is probably to have a logical reason for your methods.
As I understand 0.97 was proposed by a study for RiskMetrics. But if you attempt an accurate estimate, you would probably be disappointed. In general, volatility estimators from maximum likelihood are biased estimators. Correcting this would need some theoretical works.
Lambda is a "smoothing" factor, as Ayman A. Amin already stated. A value near 1 would suggest that a the volatility parameter is quite stable. So it depends on your model assumption, if you want to adopt 0.97 or probably a different value for accomodating a "stress scenario".
At the end, this is only one particular model. It is also possible to model the volatility to have "shock" state (e.g. using markov switching method). or even suggest a different recursive volatility equation, which accomodate lambda differently.
Summarizing everything, it is not wise to take 0.97 or any particular estimator without having your goal in sight, as your goal is most probably not to estimate lambda accurately.