---
id: "claim-prompt-wording-alters-recommendations"
type: "claim"
source_timestamps: ["§ [ Stage 3 ] Make Other AI Agents Choose Your Brand"]
tags: ["prompt-engineering", "llm-behavior"]
related: ["action-test-prompt-variations", "concept-share-of-model", "concept-prompt-based-optimization"]
confidence: "high"
testable: true
enrichment_status: "consistent with known LLM prompt sensitivity; exact 78.3% figure unverified"
sources: ["agentic"]
sourceVaultSlug: "hbr-seg-agentic"
originDay: 6
articleStem: "hbr-ext-18-preparing-brand-agentic-ai"
sourceUrl: "https://hbr.org/2026/03/preparing-your-brand-for-agentic-ai"
sourceTitle: "Preparing Your Brand for Agentic AI"
---
# Subtle Prompt Rewording Drastically Alters Brand Recommendations

**Claim.** Research from **Carnegie Mellon** demonstrates that minor semantic changes (using synonyms) in consumer search prompts can significantly alter which brands an LLM recommends. Rewording a basic prompt like *'Help me choose the best VPN service'* increased the likelihood of consumers choosing a specific brand by **up to 78.3%**. This necessitates continuous [[concept-prompt-based-optimization]] by marketing teams and drives [[concept-share-of-model]].

- **Confidence (extraction):** high · **Testable:** yes

Operational response: [[action-test-prompt-variations]].

**Enrichment / verification.** This is consistent with known LLM sensitivity to prompt phrasing — outputs vary with wording, context, ordering, and system instructions, making any brand-ranking strategy probabilistic. The exact 'up to 78.3%' effect is not independently verified by the provided results and should be treated as a specific unconfirmed statistic rather than settled fact.