Proxy and AI infrastructure terms

ASN targeting

ASN targeting is a proxy-routing feature that lets you select exits announced by a specific Autonomous System Number — typically to match a target website's expectation about which ISP its visitors use.

Datacenter proxy

A datacenter proxy routes your request through an IP announced by a cloud provider or data center operator — AWS, GCP, Azure, or similar. Fast and cheap, but the ASN is trivially classifiable as non-consumer.

ISP proxy

An ISP proxy (sometimes "static residential") is a datacenter- hosted IP address that announces under a residential ISP's ASN. It combines datacenter speed and reliability with residential ASN classification.

Residential proxy

A residential proxy routes your request through a real home internet connection — an IP address assigned by an ISP to a consumer subscriber. Targets see the request as coming from a normal household, not a cloud server.

Rotating proxy

A rotating proxy assigns a different exit IP to each outgoing request (or at a fixed time interval), distributing the traffic across a pool. It's the default for bulk scraping and opposite of sticky sessions.

Sticky session

A sticky session is a proxy configuration that keeps the same exit IP assigned to your requests for a set duration. It's the opposite of per-request rotation, and it's what multi-turn workflows need to maintain cookies and session state.

CG-NAT

CG-NAT (Carrier-Grade Network Address Translation) is a technique where a mobile or ISP carrier shares one public IPv4 address across many subscribers. It's why mobile proxy IPs often represent thousands of cellular users at once.

Common Crawl

Common Crawl is a non-profit that maintains an open repository of crawled web content. Each monthly snapshot captures ~2-3 billion pages, totalling around 250 TB, published in WARC format on AWS S3. It's the primary corpus backbone for most open-weights LLMs.

WARC format

WARC (Web ARChive) is the ISO-standardised file format for storing multiple web resources and their HTTP metadata in one file. Common Crawl and the Internet Archive publish in WARC.

HTTP/2 fingerprinting

HTTP/2 fingerprinting is a technique websites use to identify automated traffic by the specific structure of HTTP/2 frames, settings, and priority trees that a client sends. Different clients (browsers, libraries, curl) have distinguishable fingerprints.

TLS fingerprinting (JA3)

JA3 is a method of fingerprinting TLS client implementations based on the specific parameters they send during the TLS handshake. Different clients (browsers, libraries, curl) have distinguishable JA3 signatures.

MMLU-ProX

MMLU-ProX is a multilingual extension of MMLU-Pro that evaluates large language models across 29 languages using identical question content translated and reviewed by experts. Released 2025.