Cloudflare and ETH Zurich have outlined the operational challenges posed by AI-driven crawler traffic and proposed approaches to improve cache efficiency in content delivery networks. Cloudflare engineers report that AI bot traffic now exceeds 10 billion requests per week, creating patterns that differ significantly from human browsing behavior. These patterns affect latency, cache hit rates, and resource utilization across CDN and backend infrastructure.
About a third of Cloudflare’s traffic comes from automated sources, including search engine crawlers, uptime checkers, and AI assistants. AI crawlers are the most active, responsible for roughly 80 percent of self-identified bot requests. These bots issue high-volume parallel requests, often accessing rarely visited pages or scanning sites sequentially to support AI services like retrieval-augmented generation.
AI crawler behavior diverges from human browsing in several ways. Many AI crawlers maintain a high unique URL ratio, access diverse content types, and issue requests that do not effectively reuse cached content. Unlike human users, AI crawlers do not leverage browser caching or session continuity, and multiple independent instances can generate repeated requests for the same content. Cloudflare modeling shows that iterative loops by AI agents create consistently high levels of unique content access, displacing frequently requested human content in edge caches
Erika S, a systems engineer, noted in her X post:
The 70-100 percent unique access ratio in RAG loops explains the cache churn I experienced during recent fine-tuning. LRU failing under AI load makes German hosting unpredictable.
AI traffic increases cache miss rates for CDNs, reducing the effectiveness of strategies such as least recently used cache eviction, cache speculation, and prefetching. Simulated AI crawler traffic caused a measurable drop in cache hit rates for a single CDN node, increasing origin server load and slowing response times.
BeePopCommunity, a technology observer, emphasized the widespread operational impact in X post:
AI traffic breaks assumptions built for humans.
Amy Lee, CFO at Aerospike, described the broader effect on databases in a LinkedIn post:
YES! AI traffic is breaking traditional cache architectures, not just at the CDN layer but all the way to the database. When 70-100 percent of requests are unique, access patterns stop being predictable enough to cache. Most databases perform well when conditions cooperate. AI traffic is systematically eliminating optimized conditions. The ones that hold up were never dependent on those conditions in the first place. We see this in production at Aerospike: 1-2 million mixed read/write operations per second with predictable tail latency is the requirement, not the exception.
To address these challenges, Cloudflare and ETH Zurich propose AI-aware caching strategies. These include separating human and AI traffic into distinct cache tiers, testing replacement algorithms such as least frequently used or first in, first out, and exploring machine learning-driven policies that adapt dynamically. Complementary measures like structured feeds or pay-per-crawl models can help control AI access while preserving cache efficiency. Cloudflare emphasizes that AI-driven services require different caching approaches than human traffic. The proposals highlight potential operational and technical adjustments for websites to serve both human users and AI agents efficiently, underscoring the need for updated cache architectures.