The integration of Generative AI into travel platforms is creating significant performance engineering challenges, particularly during peak seasons. When an AI-powered booking assistant goes live, it initially receives positive feedback, but under heavy load, response times climb, the AI layer queues, and users abandon the booking funnel. This reveals that the intelligence shipped becomes an untested bottleneck. The fundamental issue is that organizations are deploying GenAI capabilities without engineering them to perform, scale, or fail gracefully. Unlike traditional travel platform requests, where a query hit an API, a database returned results, and a response rendered in under a second, GenAI introduces variable token generation times dependent on prompt complexity and model load, context window management affecting cost and latency, non-deterministic outputs complicating regression testing and SLO definition, third-party model API dependencies within the critical booking path, and multi-step agent chains where a slow step poisons the entire journey. LLM latency is non-linear; under heavy load, response times spike, stall, or time out entirely, unlike standard microservices where latency increases predictably. The token generation rate drops, model API queues fill up, and downstream services become blocked. A traditional load test asking how many concurrent users can be handled is insufficient; the real question is how every layer of the AI pipeline behaves when 50,000 users ask it something unexpected simultaneously.