Designing for the Machine : How Vertical Placement and HTML Structure Drive Visibility in Gemini

Sammanfattning

As generative artificial intelligence becomes an increasingly common tool for information retrieval, the way these systems interact with web content represents a significant paradigm shift in digital information access. Modern large language models (LLMs) such as Gemini utilize Retrieval-Augmented Generation (RAG) to incorporate external web data into their responses. However, research indicates that these models often exhibit a positional bias, typically prioritizing information at the beginning or end of a text se- quence, a phenomenon known as the “U-shaped” attention pattern, while often overlooking content in the middle. This thesis investigates the extent to which vertical placement and HTML-structural formatting influence the visibility and prioritization of information within Gemini-generated summaries. While previous studies have focused on plain text, this research utilizes a controlled web environment to test how semantic tags, such as headings and lists, and the presence of irrelevant information affect which factual claims are reproduced in Gemini-generated summaries. Using an experimental framework focused on information salience, the study measures “answerability” by analyzing the presence of specific factual claims across varying response lengths generated via the Gemini API. Answerability is defined as the proportion of predefined factual claims reproduced in the generated responses. The results show that semantic HTML structure generally improves information retrieval, while vertical placement effects appear to be content-dependent rather than universally consistent across webpages. In contrast, irrelevant textual noise produced a limited measurable impact under the tested conditions. The findings suggest that positional bias observed in plain-text studies does not transfer uniformly to web-based retrieval environments and that the structure of the webpage influences information visibility in more context-dependent ways than previous research has suggested. This study contributes to the emerging field of Generative Engine Optimization (GEO) by exploring how webpage design choices may influence information visibility in Gemini-generated summaries.