Effektives Prompt Engineering beim Einsatz von Generative Artificial Intelligence im Marketing

Abstract

Generative Artificial Intelligence contributes to increased productivity in marketing, prompting many marketers to adopt this technology. Controlled by natural language prompts, Large Language Models (LLMs) and Vision Language Models (VLMs) can generate original and realistic content in text and image form, supporting a wide range of use cases. Despite its immense potential, Generative Artificial Intelligence also presents several challenges. Since the generated content is probabilistic, errors cannot be ruled out. Additionally, due to the training data, models may incorporate biased opinions or outdated information. Furthermore, AI-generated content can lead to a reduction in brand authenticity. One solution to address these challenges is Prompt Engineering. Optimized prompts can significantly influence the generated content, thus leveraging the capabilities of Generative Artificial Intelligence. This work is based on the methodology of literature review and analysis, following the guidelines of Jan vom Brocke. Initially, the most important journals in the fields of Marketing and Artificial Intelligence were identified using the h index, and relevant publications were searched within them. The scientific discourse is expanded with papers from databases such as emerald.com, link.springer.com, researchgate.com, sciencedirect.com, and scholar.google.com. At the beginning of the work, key terms are explained. Next, the relevance of Generative Artificial Intelligence for marketers is demonstrated, and exemplary use cases are identified. The core of this work involves deriving insights from the literature on methods and guidelines that can be applied to Large Language Models and Vision Language Models. For the application of Prompt Engineering in Large Language Models, nine Prompt Engineering methods are elucidated. The exemplary applications using the PIA Camper example create a unified context for better reader understanding. Additionally, two dimensions of a matrix were derived from the literature to structure the methods by task type and desired interaction. In Vision Language Models, several guidelines need to be observed to achieve the desired outcome. During the iterative process, adjustments to the prompt modifiers or strategies to improve the results can be made.

Date of Award	2024
Original language	German (Austria)
Supervisor	Andreas Auinger (Supervisor)