Well testing is a very important part in the evaluation of gas discoveries. It is used to define the characteristics of a reservoir, to find boundaries and see a potential pressure depletion, which could verify the existence of a limited reservoir at an early stage. This thesis evaluated the analytical test interpretation methods. An important point is the non-uniqueness of a well test interpretation. The same pressure curve can be the result of very different conditions, leading to difficulties in the interpretation. In order to find out about the exactness of the estimation of producible volumes from an early well test, data of all performed pressure build-up tests from the RAG Rohoel-Aufsuchungs AG - concession in the Molasse in Upper Austria and Salzburg is digitized, and therefore around 600 tests can be analyzed. The used method of deriving the average drainage pressure is highly controversial. The analysis of the archived data shows that in both cases, for open-hole and cased-hole tests, gas volumes are often estimated inexactly. The possibility to build numerical models of the tests is presented and evaluated. The standard well testing software can be used to model the acquired insights and to converge it interdisciplinary with the extensive knowledge of the geologist. Dynamic 2D- and 3D- simulation of the tests with a commercial simulation software allows to analyze different geologic environments and show clearly that the determination of average pressures in reservoirs with boundaries is very risky, and can lead to a severe overestimation, but also a slight underestimation of the reserves. Without a rough estimation of the lateral extension beforehand, a prediction of the proper case is not possible. Therefore, the estimation of reserves in these environments with the material balance method is erroneous. The compiled well test interpretation data and the results from the simulations are used to feed a neural network. With real-world data it is only possible to find numerical relations under certain preconditions, like the exclusion of samples with a lower production than 5 MMscm or the differentiation between formations. These reservoirs can be defined as geologically similar formations, which perform likewise during the test and during production. First trials to feed a neural network with simulation results and use this method to improve the prediction accuracy show that the capability of predicting gas initially in place strongly depends on the predefined knowledge about geological conditions. This method is, therefore, only partly applicable to real-world problems.