The purpose of this research is to investigate the current capabilities and performance of generative AI based on the use case of performing search queries on geodata as well as how well generative AI can produce maps. To achieve the purpose of this study, the capabilities of generative AI with aforementioned use cases has been tested using various AI models and benchmarked using a set of tests. The dataset to test the use case of search queries was collected by performing multiple API calls in order to get a pair based question-answer pair of how the current search query works. The other dataset containing choropleth maps of different regions related to keywords of that region was gathered similarly.For the first use case of performing search queries on geodata which is geodata text generation there were 2 methods tested, one using llama-7b-hf model trained on custom dataset as well as RAG model with Elasticsearch. The test metrics for this test generation key evaluation metrics, e.g., precision, recall faithfulness and relevancy. Results showed that the rag model performed better in terms of precision and recall, but that the RAG model outperformed in terms of faithfulness and relevancy. The key takeaway is that having a context to define the answer from massively helps with the hallucination problem of using generative AI models. The result for the RAG model showed promise as something that could be actually used with very limited drawbacks in terms of usual drawbacks for generative AI models like outdated datasets financial cost for training etc.For the second use case of how well generative AI can produce maps which is geodata image generation there were 2 methods tested, one using Dall-e-3 as well as Stable Diffusion model SDXL trained on custom dataset. The test metrics were simple visual evaluation of the output images of the models. Results showed that the Stable Diffusion model SDXL performed better in terms of actually producing city images looking like the original city. The key takeaway is that prompt engineering was an issue for both models, and that reduced image quality for the training set for Stable Diffusion model SDXL most likely reduced the overall quality of the product. The result for the Stable Diffusion model SDXL showed promise and if it had been properly designed with good describing prompts with higher quality dataset images the model could possibly see future usage.