Authors: Ansh Yadav
Abstract: Consumers increasingly meet brands not in a list of ranked links but inside a single synthesised answer written by a large language model (LLM). This paper develops a conceptual framework for Generative Engine Optimisation (GEO), the practice of shaping the probability that a brand is retrieved, cited, named, and recommended by generative search engines, and synthesises the empirical evidence available as of September 2026. We formalise brand presence as a four-stage retrieval–citation–recommendation funnel with a parallel parametric-memory route, define a family of measurable outcomes (citation share, share of model voice, ghost-citation rate, recommendation rank) and combine them into a Brand Visibility Index (BVI). We then synthesise eight primary and secondary studies, graded by evidential tier, spanning the original GEO benchmark of Aggarwal et al. (2024), the earned-media audit of Chen et al. (2025), the verifiability audit of Liu et al. (2023), and industry and institutional studies of click behaviour and brand mentions. Four regularities emerge. First, the click economy around AI summaries is contracting: two independent designs report roughly 40–47% fewer organic clicks when a summary is shown. Second, content-level interventions that raise evidence density (quotations, statistics, source citations) improve inclusion in generated answers by up to about 40% in controlled tests, while keyword stuffing does not. Third, engines exhibit a systematic preference for third-party, earned-media sources over brand-owned content, and branded web mentions correlate with AI Overview brand mentions at 0.664 against 0.218 for backlinks. Fourth, being cited and being named are distinct events: in one industry audit, 61.7% of domain appearances were citation-only. We derive eight testable propositions, specify a pre-registrable audit and field-experiment protocol, and discuss ethical limits, including the boundary between legitimate optimisation and adversarial manipulation. No new primary data are reported; the contribution is conceptual and integrative.
DOI: http://doi.org/10.5281/zenodo.23077972
