Residential Proxy & Benchmarking Glossary
30 termsQuick-reference definitions for every metric and concept used in our proxy benchmarks. For detailed scoring methodology, see Methodology.
Metrics
P50 Latency
P50 (50th percentile) latency is the median time it takes for a request to travel through the proxy and return a response. Half of all requests complete faster than this value, half slower. We measure Time to First Byte (TTFB) through the proxy to the nearest Cloudflare edge, the same clean network measurement independent benchmarks like Proxyway use (response time to the nearest CDN). On that metric good residential is sub-second (the strongest providers land near 0.4–0.9s), 0.8–2s is mid-tier, and above 2s is genuinely slow. EU pools typically run faster than US for the same provider. (ISP/datacenter proxies are far faster, a different product.)
P95 Latency
P95 latency represents the time within which 95% of all requests complete. It captures tail latency — the slow outliers that P50 hides. A large gap between P50 and P95 indicates inconsistent performance. For time-sensitive scraping (e.g., price monitoring), P95 matters more than P50.
TTFB (Time to First Byte)
TTFB measures the delay between the client sending an HTTP request and receiving the first byte of the server's response. In proxy benchmarking, TTFB includes: DNS resolution, TCP connect to the proxy, TLS handshake, proxy-to-target connection, and the target server's processing time. It's the primary latency metric used in our Composite Score.
Success Rate
Success rate measures how often a request through the proxy completes with usable content — not blocked, not challenged, not timed out. We count CAPTCHAs and interstitial pages as failures (an HTTP 200 that serves a challenge page is not a success), and we measure it per context: transport (our controlled endpoints), neutral sites, and high-defense targets (Google, Amazon). One honest caveat from our own data (methodology v2.3): on high-defense targets, success rate is saturated — every credible residential pool lands in a similar band — so it confirms a provider works but can't rank one above another. For ranking, latency, session stability, and pool cleanliness separate providers far better.
IP Uniqueness Rate
IP uniqueness measures how diverse a provider's IP pool is in practice. A high uniqueness rate (e.g., 0.85) means 85% of requests exit through a different IP — indicating a large, well-rotated pool. Low uniqueness suggests a small pool where IPs are recycled frequently, increasing the chance of IP-level blocks.
Residential Rate
Residential rate indicates what fraction of a provider's exit IPs are actual residential addresses (ISP-assigned to homes) versus datacenter IPs. Residential IPs are less likely to be flagged by anti-bot systems. We classify IPs using ASN databases (DB-IP Lite) cross-referenced with curated datacenter ASN lists.
Geo Accuracy
Geo accuracy measures the reliability of a provider's geo-targeting. When you request a US proxy, does the exit IP actually geolocate to the US? We verify by capturing the exit IP via our /reflect endpoint, then checking it against GeoIP databases. Low geo accuracy means geo-targeted scraping will produce inconsistent results.
Download Speed (Mbps)
Download speed measures the sustained data transfer rate through the proxy, in megabits per second. Measured by downloading a test file through the proxy and calculating throughput. Important for bandwidth-intensive tasks like image scraping or large page downloads. Residential proxies are typically slower than datacenter proxies due to ISP last-mile constraints.
Deny Rate
Deny rate tracks how often a target responds with a hard block — HTTP 403 Forbidden, connection reset, or an empty response. A high deny rate means the provider's IPs are known to the target's blocklist. Distinguished from challenge rate, where the target presents a CAPTCHA or JavaScript challenge rather than a flat denial.
Challenge Rate
Challenge rate measures how often a target presents an intermediate verification step (CAPTCHA, Cloudflare JS challenge, reCAPTCHA) rather than serving content directly. Challenges are softer than denials — they can sometimes be solved — but they add latency and reduce effective throughput. A provider with low challenge rates has IPs with better reputation scores.
Latency Consistency
Latency consistency is the ratio between a provider's P95 and P50 latency. A ratio near 1× means requests take about the same time every time; a large ratio means the typical number hides slow outliers. We band it as steady (under 2.5×), variable (2.5–4×), and spiky (over 4×). Two providers with the same median can behave very differently under load: a P50 of 500ms with a P95 of 3s will miss timeouts a steady pool never hits. Shown on provider cards and compare pages; it does not affect the composite score.
Clean IP Rate
Clean IP Rate is the percentage of a provider's sampled exit IPs that have zero abuse reports in AbuseIPDB. Unlike anti-bot success rate (saturated across credible providers), pool cleanliness differentiates strongly — we observe a wide spread between providers. Dirty IPs carry prior abuse baggage: they hit rate limits, CAPTCHAs, and blocklists sooner, even when the request itself is polite. Measured continuously from real probe exits; displayed on the dashboard and use-case pages, weight 0% in the composite until trusted over a longer window (same policy as geo).
Scoring Dimensions
Composite Score
The Composite Score (score_v2_total) is a weighted combination of dimension scores (methodology v2.3): Core Network Performance (35%), Session Reliability (30%), Neutral Reachability (15%), and Target Reachability (20%). Geo Integrity and Clean IP Rate are measured but carry 0% weight until trusted over a longer window. Computed daily from rolling 30-day probe data. A score above 70 indicates good all-around quality; below 50 signals significant issues.
Core Network Score
Core Network Score (35% of Composite) evaluates fundamental proxy connectivity: TTFB, timeout rate, and controlled-plane success rate. Measured via Plane A controlled probes hitting our own endpoints every 15 minutes. High scores indicate fast, reliable connections; low scores suggest network congestion or poor infrastructure.
Session Reliability Score
Session Reliability Score (30% of Composite) tests whether sticky sessions actually work. We open a session, make sequential requests over the session TTL, and check if the exit IP changes unexpectedly. Unexpected rotation receives a heavy penalty (3x weight) because it breaks stateful scraping workflows like login flows or multi-page checkouts.
Geo Integrity Score
Geo Integrity Score answers: when you request a proxy in a specific country, do you actually get one there? It currently carries 0% weight in the Composite (under revision) — our single-source country check isn't rigorous enough to publish, and it returns to a nonzero weight only once multi-country, multi-database measurement ships. We'd rather score nothing than score it wrong.
Neutral Reachability Score
Neutral Reachability Score (15% of Composite) tests if proxies can reach neutral targets like httpbin.org and Cloudflare speed test. These services don't actively block proxies, so failures indicate fundamental connectivity problems (DNS issues, IP blacklisting at network level, broken TLS) rather than anti-bot detection.
Target Reachability Score
Target Reachability Score (20% of Composite) measures success against actively-defended targets like Google and Amazon. Tested via Plane C probes every 6 hours with strict safety rules (24h per-IP cooldowns, stop-after-deny). This score directly reflects how well a provider's IPs evade anti-bot detection systems.
Testing Infrastructure
Plane A (Controlled)
Plane A is the primary testing plane that measures fundamental proxy performance. Probes hit ProxyStats-owned endpoints (/healthz, /reflect, /tiny, /download) every 15 minutes from both EU and US vantage points. Because we control the targets, results are completely reproducible and unaffected by third-party server load or anti-bot systems.
Plane B (Neutral)
Plane B probes neutral external services (httpbin.org, Cloudflare speed test) that don't block proxies. Running every 15 minutes, it validates that a provider's proxies can reach the broader internet — not just our controlled endpoints. Failures here indicate network-level issues that would affect any scraping target.
Plane C (High-Defense)
Plane C tests proxy performance against targets with sophisticated anti-bot defenses. Runs every 6 hours (not every 15 minutes) with strict safety rules: 24-hour per-IP cooldowns, stop-after-deny logic, and rate limiting. This cautious approach prevents our benchmarking from degrading the provider's IP reputation while still measuring real-world bypass capability.
Probe Run
A probe run is one complete test cycle: connect to the proxy, send a request to the target, measure all timing metrics, record the exit IP, and store the result. Each probe run produces one row in the probe_runs table with timing, status, exit IP, and download metrics. Thousands of probe runs are aggregated into daily rollups for scoring.
Daily Rollup
Daily rollups aggregate all probe runs for a given provider + region + day into statistical summaries: P50/P95 latencies, success rates, session survival rates, and dimension scores. The rollup runs at 00:05 UTC and produces the data that powers the Composite Score, leaderboard rankings, and trend charts. Each rollup also receives a confidence level based on probe count.
Vantage Point
A vantage point is a server running probe workers in a specific region. ProxyStats currently operates two vantage points: EU-Central (Frankfurt, Germany) and US-East. Each vantage point independently tests all providers, giving separate latency readings per region. More vantage points mean more representative benchmarks; providers that perform well from both regions demonstrate globally consistent infrastructure.
Confidence Level
Confidence level categorizes how much data backs a provider's scores: Low (<10 probes/day), Medium (10-50), High (>50). Scores with low confidence should be interpreted cautiously — they may not represent steady-state performance. New providers start at low confidence and graduate to high within 2-3 days of active probing.
General Concepts
Residential Proxy
Residential proxies use IP addresses assigned by Internet Service Providers to real households. Because the traffic appears to originate from a regular home user, websites are much less likely to block it compared to datacenter proxies. The trade-off: residential proxies are slower (limited by home internet speeds) and more expensive per GB. They're the standard choice for scraping protected targets.
Datacenter Proxy
Datacenter proxies route traffic through servers in data centers (AWS, Hetzner, OVH, etc.). They're fast and cheap but easily detectable — anti-bot systems maintain lists of known datacenter IP ranges. Suitable for targets without anti-bot protection, but ineffective against Google, Amazon, or Cloudflare-protected sites.
Sticky Session
Sticky sessions (also called session proxies) keep the same exit IP for a configured duration (typically 1-30 minutes). Essential for stateful workflows: logging into a website, navigating multi-page forms, or maintaining shopping cart state. Without sticky sessions, each request would exit through a different IP, breaking any session-based interaction.
IP Rotation
IP rotation assigns a different exit IP for each new request (or after a set interval). This distributes requests across the provider's pool, reducing the chance any single IP gets flagged. Rotation types include: per-request (new IP every request), timed (new IP every N minutes), and geo-targeted (rotate within a specific country). Rotation quality depends on pool size and diversity.
Price per GB
The standard pricing unit for residential proxies. Providers charge based on total bandwidth consumed (upload + download). Typical range: $1-15/GB depending on provider and plan size. Lower price doesn't always mean better value — cheap providers may have lower success rates, requiring more retries and consuming more bandwidth to get the same results.