Caption Injection for Optimization in Generative Search Engine

Xiaolu Chen, Yong Liao, Wu, Haojie, Chen, Zhen, Liao, YongarXiv 2025 · DOI · arXiv

TL;DR

As search engines evolve from ranked lists to AI-generated summaries, a new field called Generative Search Engine Optimization (G-SEO) has emerged. This paper introduces Caption Injection, the first multimodal G-SEO technique that boosts a content source's visibility in AI responses by converting image data into strategically placed text descriptions, significantly outperforming traditional text-only optimization methods.

What Problem Does This Paper Address?

In the era of Generative Search Engines (GSEs)—think Perplexity, Gemini, or SearchGPT—the goal of SEO has shifted. It’s no longer just about appearing at the top of a list; it’s about ensuring your content is actually selected and synthesized into the final answer generated by a Large Language Model (LLM). This is known as subjective visibility.

While researchers have started exploring G-SEO for text, most methods ignore the rich information contained in images. As search engines move toward Multimodal Retrieval-Augmented Generation (MRAG), there is a massive gap: how do we optimize multimodal content so that the AI "notices" and includes it in its summary?

Methods

The authors propose a three-stage pipeline called Caption Injection to bridge the gap between visual data and textual generation:

  1. Structural Generation: First, a Vision-Language Model (VLM) analyzes an image to create a structured "Object-Action-Scene" (O-A-S) triad. This ensures the core semantic meaning of the image is captured objectively.
  2. Alignment Refinement: An LLM then takes this raw description and the surrounding article text to "rewrite" the caption. The goal is to make the visual description semantically consistent with the article's tone and context, preventing "semantic drift."
  3. Semantic Injection: Finally, the LLM identifies the most natural insertion points within the original text to weave in these new descriptions.

By turning images into high-quality, context-aware text, the method makes the source material more "attractive" to the generative engine's retrieval and synthesis process.

Datasets

The researchers evaluated their approach using MRAMG, a comprehensive benchmark for Multimodal Retrieval-Augmented Generation.

  • Scope: It includes 4,800 query-content pairs across six distinct domains (Arxiv, Manual, Recipe, Web, Wiki, and Wit).
  • Significance: This dataset is specifically chosen because it provides a balanced mix of text and image sources, allowing for a rigorous test of how modality affects AI perception.

Key Insights and Ideas

  • Uniqueness is the Secret Sauce: The most striking finding was that Caption Injection primarily boosts the uniqueness dimension of content. By adding visual semantics that other text-only sources lack, the content becomes more "irreplaceable" to the LLM during the generation phase.
  • The "Subjective Visibility" Metric: The paper moves beyond simple click-through rates, using an adapted G-EVAL 2.0 metric. This evaluates content across seven dimensions, including relevance, influence, and "positional salience," providing a much more nuanced view of how AI "values" a source.
  • Text vs. Image Preference: Interestingly, the study found that in purely textual scenarios, "multi-shot" (repeated) injections help reinforce concepts. However, in multimodal settings, over-injecting captions can actually backfire by distracting the LLM from core textual details.

Key Results

  • Superior Performance: In multimodal settings, Caption Injection achieved a 1.12% relative improvement in subjective visibility, outperforming the best text-only baseline (Fluency Expression Optimization) by 0.41%.
  • Cross-Modal Adaptability: Unlike traditional SEO tricks that often fail when the medium changes, Caption Injection maintained stable performance across both unimodal and multimodal search environments.
  • Domain Sensitivity: The method was particularly effective on the "Wit" and "Recipe" datasets, where visual information is naturally more critical to the user's query.

Limitations and Future Work

The authors identify a few "hard nuts to crack":

  • The Long-Text Problem: On datasets with extremely long articles (like the "Manual" category), the added captions often got "diluted," leading to lower optimization gains.
  • Shallow Fusion: Currently, the method is "shallow" because it converts images to text before injection. The authors suggest that future work should explore "deep" multimodal interactions where the LLM processes raw visual features alongside text.
  • Black-Box Dynamics: Because GSEs are often proprietary black boxes, the exact "preference" of the underlying model remains a moving target that requires constant adaptation.

Summary updated

Abstract

Generative Search Engine (GSE) leverages the Retrieval-Augmented Generation (RAG) technique and the Large Language Model (LLM) to integrate multi-source information and provide users with accurate and comprehensive responses. Unlike traditional search engines that present results in ranked lists, GSE shifts users' attention from sequential browsing to content-driven subjective perception, not only driving a paradigm shift in information retrieval but also highlighting the importance of enhancing the subjective visibility of content in generative search. In this context, Generative Search Engine Optimization (G-SEO) methods have emerged as a new research focus. With the rapid advancement of Multimodal Retrieval-Augmented Generation (MRAG) techniques, GSE can now efficiently integrate text, images, audio, and video, producing richer responses that better satisfy complex information needs. Existing G-SEO methods, however, remain limited to text-based optimization and fail to fully exploit multimodal data. To address this gap, we propose Caption Injection, the first multimodal G-SEO approach, which extracts captions from images and injects them into textual content, integrating visual semantics to enhance the subjective visibility in generative search. We systematically evaluate Caption Injection on MRAMG, a benchmark for MRAG, under both unimodal and multimodal settings. Experimental results show that Caption Injection significantly outperforms text-only G-SEO baselines under the G-EVAL metric, effectively improving the subjective visibility of content perceived by users, and demonstrating the practical benefits of multimodal information in G-SEO.
BibTeX
@misc{chen2025caption,
  author = {Xiaolu Chen and Yong Liao and Wu, Haojie and Chen, Zhen and Liao, Yong},
  title = {{Caption Injection for Optimization in Generative Search Engine}},
  year = {2025},
  eprint = {2511.04080},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.04080},
  howpublished = {arXiv},
}
Figure