The challenge
A regional airline group with operations across three continents was losing competitive ground in metasearch placement — primarily because their search response times were too slow. Internal data showed look-to-book ratio dropping below industry benchmark in three key markets.
The technical root cause was a series of inefficient paths between metasearch partners, the airline’s booking API, and downstream GDS calls. Each hop added latency, and the existing CDN deployment wasn’t optimized for the request patterns.
The approach
We focused on four areas:
Edge path optimization
We rearchitected the metasearch-facing edge to reduce hops, add intelligent routing to the GDS endpoints, and implement short-TTL caching for repeat searches within a session.
API security overlay
The previous deployment didn’t differentiate between metasearch partner traffic and scraping bots. We added partner authentication and rate-limiting at the edge — with explicit allowlists for verified metasearch partners — and applied bot challenges only to unknown traffic.
Booking flow protection
Payment and account creation flows got dedicated security policies. Bot Manager rules tuned for fare scrapers were paired with account protection rules targeting credential stuffing.
Observability
Real-user monitoring deployed across the search and booking flows surfaced latency hot spots that synthetic monitoring had been missing. We tuned aggressively against P75 of real traffic, not synthetic averages.
The results
Search response time at P75 dropped from 460ms to 220ms — a 52% improvement. The look-to-book ratio recovered above benchmark within 90 days of launch, with the strongest gains in markets where the airline had previously been losing placement.
Booking uptime through the period hit 99.99%, including through two weather-driven demand spikes that previously would have triggered manual scaling.
What made it work
The airline’s data team owned the success metrics from the start — look-to-book ratio, search latency by market, booking conversion. Every engineering decision was framed against those metrics, which kept the work focused on outcomes rather than vanity numbers.
The willingness to break the legacy assumption that “we cache everything for an hour” was also critical. Modern fare data has a TTL of seconds, not minutes. Getting the cache strategy right meant being honest about freshness requirements per content type.