INCITE: In-context Non-parametric Conditioning of Interpretive Transformer-state for citation Elicitation

Generative Engine Optimization (GEO) has concentrated on a problem that ends too early: the retrieval event. Once a document enters the active context window of a generative language model, retrieval is no longer the operative question. The question that actually governs citation is: what cognitive state does reading this document produce in the model, and does that state favor or disfavor attribution? This paper formalizes that question as a research program. This paper introduces the INCITE state (In-context Non-parametric Conditioning of Interpretive Transformer-state for citation Elicitation), a latent representational configuration arising in a language model's mid-to-late residual stream as it processes an in-context document, which causally mediates citation behavior. Drawing on three converging research programs, this study constructs a mechanistic account of how document text induces functionally distinct cognitive states in reading models and how those states determine citation outcomes. From the emotion-function interpretability literature, this work derives the principle that documents activate abstract functional representations analogous to epistemic affect states during reading, and that the valence and arousal geometry of those states directly predicts citation favorability. From the latent geometry program, this analysis derives the principle that documents function as soft constitutions, non-parametrically conditioning the model's residual stream into recoverable geometric families that either align with or diverge from citation-ready configurations. From feature-level optimization research, this paper derives empirical grounding for the layer-wise INCITE formation hypothesis: content features dominate citation because they engage mid-to-late layer integration processes; surface lexical features fail because they operate only at early layers where the INCITE state has not yet differentiated. This study proposes an INCITE-centric optimization framework, a taxonomy of citation-favorable and citation-suppressive cognitive induction properties, a full experimental evaluation protocol, and a set of falsifiable hypotheses. The central implication is practical and immediate: the next frontier of AEO and GEO is not content that ranks well or reads well for humans. It is content that produces epistemically favorable cognitive states in the language models that read it.