---
id: "concept-position-effects"
type: "concept"
source_timestamps: ["§ The Emerging Field of Bot Psychology"]
tags: ["e-commerce", "algorithmic-bias", "ui-ux"]
related: ["concept-bot-psychology", "question-optimizing-conflicting-biases", "contrarian-bot-rationality"]
definition: "The phenomenon where AI agents favor products based on their spatial location in a display, with preferences varying significantly across different AI models."
sources: ["geo"]
sourceVaultSlug: "hbr-seg-geo"
originDay: 3
articleStem: "hbr-ext-13-ai-upending-marketing"
sourceUrl: "https://hbr.org/2026/02/ai-is-upending-marketing-on-two-fronts"
sourceTitle: "AI Is Upending Marketing on Two Fronts"
---
# AI Model Position Effects

Similar to humans, AI agents exhibit **position effects** — a bias toward products based on their spatial display location in an e-commerce sandbox. But these preferences are highly *irrational* and vary drastically across foundation models.

Research from **Columbia and Yale** demonstrated that while GPT, Claude, and Gemini **all prefer products in the top row** of a display, their specific preferences within that row diverge completely:

- **GPT** strongly favors the **first (leftmost)** position
- **Claude** prefers the **middle** position
- **Gemini** favors the **right** side

This inconsistency creates a massive optimization challenge for retailers designing page layouts for machine customers — see the unresolved [[question-optimizing-conflicting-biases]]. Position effects are one of the three pillars of [[concept-bot-psychology-d13]] and the "irrational" half of [[contrarian-bot-rationality]].

**Enrichment caveat:** Position effects are well documented in *human* choice behavior (primacy/recency, middle-choice bias), and LLMs may inherit or exhibit similar heuristics. But as an experimental observation in *one sandbox*, this should not be treated as a robust cross-domain rule — it may vary by prompt, UI, model version, and training. Domain experts favor **non-spatial, standardized data feeds** and continuous testing over betting on fixed position biases that may disappear as models evolve.


## Related across articles
- [[concept-ai-model-segmentation]]
- [[claim-llm-processing-styles-vary]]
- [[claim-model-idiosyncrasy]]