Multimodal Entity Indexing: A Conceptual Framework for Autonomous Visual Entities in Generative Engine Optimization

Researcher: Tumengkol, S.S.
Type of work: Conceptual paper (draft). It is not a dissertation and has not been peer reviewed.
Last updated: October 26, 2026
Download the conceptual paper (PDF)

Status of the work

This page previously reported empirical results, including a citation-increase figure and a benchmark table. Those claims have been removed because the supporting data have not been published, and the conceptual paper (PDF) itself states that no experimental data were generated. At present this work is a conceptual framework with a research agenda. There are no results that can be cited as evidence.

Abstract

The rapid development of Large Language Models (LLMs) and Large Multimodal Models (LMMs) is shifting information retrieval from document-centric indexing toward multimodal semantic representation. This paper proposes Multimodal Entity Indexing (MEI), a framework for treating visual assets as knowledge-bearing entities within Generative Engine Optimization (GEO). The framework has three components: Autonomous Visual Entity (AVE), Information-to-Pixel Ratio (IPR), and Tri-Anchoring Context System (TACS). The paper presents no empirical evidence; it sets out a theoretical foundation and a research agenda for later studies.

Framework components

• Autonomous Visual Entity (AVE): a visual asset that can be represented on its own in a multimodal vector space.
• Information-to-Pixel Ratio (IPR): a proposed measure of factual information density relative to pixel cost. Its formula is being reformulated before it is tested.
• Tri-Anchoring Context System (TACS): contextual alignment through three text anchors: file name, deep alt text, and caption.

Next steps

Empirical validation is being designed as a controlled experiment, with hypotheses and an analysis plan registered before any data are collected. The protocol is published on a separate research domain, multimodalentity.org. Results, including null or negative ones, will be reported there.

Disclosure

The author runs PahamSEO, a business in the SEO field, so a potential conflict of interest exists. "PahamSEO Research Institute" is an internal research initiative within the PahamSEO ecosystem and is not a separate legal entity.

References

1. Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., ... & Sutskever, I. (2021). Learning transferable visual models from natural language supervision. Proceedings of ICML, PMLR 139, 8748-8763.
2. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2023). GEO: Generative Engine Optimization. arXiv:2311.09735. (Check the author list and title directly on arXiv before publishing.)