LLMs boost push notification CTR by 14.5% over static templates

A new paper from arXiv (2605.16264) by Nilesh Agrawal argues that push notification effectiveness has been held back by focusing on who to notify and when, while ignoring how to communicate. The author defines six dimensions of notification message quality: contextual relevance, clarity, actionability, novelty handling, linguistic freshness, and persuasive appropriateness. Across 28 reviewed deployments (from 142 screened), LLM-based composition improved click-through rates by +8% to +14.5% over static templates and +1% to +2.5% over mature slot-filling systems. However, the paper cautions that these gains span heterogeneous systems and are often misattributed to text generation alone, when adjacent components like targeting and ranking also contribute.

The paper also provides a three-criterion decision framework to help engineers decide when LLM generation is the binding constraint versus when optimizing other pipeline stages would yield more impact. It examines domain-specific applications in social media, food delivery, and e-commerce, and proposes a unified architecture that includes budget-aware routing, grounded generation, candidate ranking, diversity controls, and online learning. For product teams, the key takeaway is that message wording is an underinvested lever—and LLMs can unlock meaningful uplifts without changing targeting models or notification schedules.