MindFuse: Towards GenAI Explainability in Marketing Strategy Co-Creation

Aleksandr Farseev, Marlo Ongpin, Qi Yang, Ilia Gossoudarev, Yu-Yi Chu-Farseeva, Sergey I. Nikolenko

Abstract

Generative AI (GenAI) has primarily been utilized in digital marketing as a high-speed content factory. However, this approach often lacks strategic depth and interpretability. This paper introduces MindFuse, a novel framework designed to transition GenAI from a simple generator to a strategic co-creator. By integrating Large Language Models (LLMs) with unsupervised clustering and Click-Through Rate (CTR) prediction, MindFuse extracts actionable personas and narratives from real-world advertising data. The framework provides explainable visual heatmaps and autonomous performance analysis, bridging the gap between raw content production and data-driven strategic planning.

Introduction

The digital marketing landscape is currently overwhelmed by a high volume of AI-generated content. Despite this abundance, marketers face a persistent "black box" problem: while performance metrics like CTR are readily available, the underlying strategic reasons for an ad's success or failure often remain opaque. Traditional manual audits to identify customer personas and thematic pillars are too slow for modern "always-on" social media environments. MindFuse addresses these challenges by providing an explainable, data-driven pipeline for marketing strategy co-creation, allowing agencies to move beyond content generation toward strategic understanding.

The MindFuse Framework

The MindFuse architecture operates through several integrated modules that transform raw ad data into structured strategic insights:

  1. Content Pillar Identification: The system utilizes LLMs to parse ad corpora, extracting structured semantic units such as targeted human needs, product categories, and emotional archetypes.
  2. Dual-Layer Clustering: To move beyond surface-level descriptions, the framework employs X-Means clustering on ADA embeddings.
  3. Layer 1 (Personas): Clusters the "Audience" pillar to identify distinct psychological and behavioral profiles.
  4. Layer 2 (Thematic Challenges): Clusters the "Insights" pillar to find recurring pain points or aspirations.
  5. Narrative Synthesis: The system pairs these personas and challenges into a "hero + conflict" structure, which an LLM then uses to generate high-quality creative briefs and campaign stories.
The MindFuse pipeline: Extracting structured strategic insights from raw ad content using LLMs and prompt engineering.

Explainability and Performance Analysis

Beyond generation, MindFuse focuses on interpretability through Explainable Content Scoring. Built on the SODA architecture, this module predicts CTR and generates attention-based heatmaps. These heatmaps visualize which specific regions of an ad—such as the Call to Action (CTA) button, background, or product—are driving user engagement.

Furthermore, the system features an Autonomous Performance Analysis module. Using a custom prompt structure (Knowledge + Role + Task + Guiding Questions + Data Input), it acts as a zero-shot analyst, interpreting live campaign telemetry to suggest budget reallocations or creative iterations based on emergent strategic reasoning.

Experimental Validation and Results

The researchers validated MindFuse using a corpus of 1,120 B2B-focused ads from ride-hailing platforms and live campaign telemetry involving over 65,000 impressions. Real-world testing across four organizations, including performance and creative agencies, yielded several key results:

  • Efficiency Gains: MindFuse delivered 2x to 12x efficiency gains in agency workflows. Tasks like audience research and competitor monitoring that previously required 5 hours were reduced to just 30 minutes.
  • Predictive Accuracy: The attention heatmaps were highly validated; removing elements identified as "high-attention" led to a significant CTR drop of up to 64%.
  • Strategic Depth: The system successfully identified three distinct B2B personas (Efficiency Enthusiasts, Financial Empowerment Champions, and Efficiency Innovators) from raw data, providing granularity typically reserved for expensive manual research.
  • Visual Sensitivity: Ablation studies revealed that visual clarity is paramount; for instance, replacing a transparent background led to a more severe CTR drop (64%) than removing the primary visual element (58%).

Conclusion

MindFuse represents a significant shift toward the concept of "Hybrid Agencies," where the competitive edge lies in the synergy between human imagination and AI-driven technical fluency. While the authors note that AI suggestions can occasionally remain generic, the integration of empirical grounding with third-party audience databases and full-funnel attribution remains a promising path for future development.

Conceptualizing the Hybrid Agency: AI as a strategic partner in the boardroom, augmenting human creativity with data-driven insights.

Summary updated

BibTeX
@inproceedings{farseev2025mindfuse,
  author = {Aleksandr Farseev and Marlo Ongpin and Qi Yang and Ilia Gossoudarev and Yu-Yi Chu-Farseeva and Sergey I. Nikolenko},
  title = {{MindFuse: Towards GenAI Explainability in Marketing Strategy Co-Creation}},
  booktitle = {ACM Multimedia},
  year = {2025},
  doi = {10.1145/3746027.3758167},
}
Figure