---
type: "synthesis"
arc: "bot-behavior"
articles: ["a006", "a010", "a013", "a025", "a029"]
tags: ["model-heterogeneity", "segmentation", "volatility"]
id: "cross-day-model-heterogeneity"
sources: ["geo"]
sourceVaultSlug: "hbr-seg-geo"
originDay: 3
articleStem: "hbr-seg-geo"
sourceUrl: "(unified vault: 13 sources)"
sourceTitle: "HBR — Demand Ⅰ-A · GEO / AI-mediated discovery & agentic commerce"
---
A finding no single article owns but all reinforce: **'AI' is not a monolith — each LLM behaves like a distinct market segment, and cross-model variance is the norm, not the exception.**

- **Sabbah & Acar**: treat each model as a consumer segment — [[concept-ai-model-segmentation]], [[concept-reasoning-vs-non-reasoning-models]].
- **Dubois (SOM)**: [[claim-llm-processing-styles-vary]] — Chanteclair scores 19% on Perplexity, 0% on Llama ([[entity-chanteclair]]); Airbnb framed as 'uniqueness' vs 'local' vs 'flexibility' by model.
- **Gale (recall share)**: [[claim-ai-visibility-fragmented]] — of 716 brands, only 8.4% appear on all three engines.
- **Puntoni (position effects)**: [[concept-position-effects]] — GPT prefers left, Claude middle, Gemini right; creating [[question-optimizing-conflicting-biases]].
- **Dubois (luxury)**: [[claim-model-idiosyncrasy]] — the same Van Gogh cue yields indifferent/lower/higher WTP across Gemini/ChatGPT/Claude.

Compounding the spatial variance is *temporal* variance: model updates act as 'exogenous demand shocks' ([[claim-fixed-strategies-expire]], [[question-model-update-volatility]], [[question-som-volatility]]), which is precisely why the corpus repeatedly demands *continuous* testing ([[concept-continuous-ai-simulation-infrastructure]]) rather than one-off optimization. This heterogeneity is the reason the [[cross-day-persuasion-penalty-convergence]] must always be hedged 'by model and category.'