There is growing interest in estimating the causal effect of unstructured media content on business and consumer outcomes (e.g., new product adoption, market sales, brand affinity, investments). A common challenge is that the driver (that is, the volume of media content) is likely endogenous to the outcome of interest, such that directly regressing the outcome on the driver may yield biased estimates of the causal effect. To address this issue, researchers frequently employ instrumental variables (IV) analyses. We propose a promising substitute to the classic IVs in the case of unstructured data (e.g. news text, social media images, videos), where the media corpus can be naturally represented as a combination of topical content predefined by domain expertise. Our proposed solution is inspired by Bartik IVs, also known as the shift-share instruments. We formalize the procedure to construct the Bartik instrument using outputs from topic models and discuss the technical considerations in its implementation and performance diagnostics. The proposed procedure is flexible and provides a principled way to search for valid IVs for text-based marketing research. The same discussion extends to audio, visual, and video-based research with minor modification (mostly differs by how to extract the topic components in each of the data modality).