Skip to content
Glossary

What is a proxy success rate?

Definition

The share of requests that return a usable response. Without a sample size, a target, a date and a definition of success, the number means little.

Read the full guide407 Proxy Authentication Required: 407 vs 403 vs 429407 means your proxy refused you; 403 and 429 usually come from the site. Real responses from Squid, three anti-bot systems and…

A rate is successes divided by attempts, and both halves are choices. What counts as a success? Which attempts are counted: first tries only, or the final outcome after retries, and do timeouts count at all? Change either answer and the same run produces a different number. That is why two published rates usually cannot be compared, even before you ask how many requests sat behind them.

How much 98% can mean

A rate measured on a sample is an estimate with a margin around it. The Wilson score interval, the method the NIST/SEMATECH e-Handbook of Statistical Methods gives for a proportion, shows how wide that margin is. These are computed, not measured: the 95% interval around an observed 98% at four sample sizes.

Width of the 95% Wilson interval around an observed 98% success rate, by number of requests Drawn from the table below.
Width of the 95% Wilson interval around an observed 98% success rate, by number of requests. Bar chart of Interval width (points) by Requests in the sample.4 rows, drawn from the table that follows this chart.02.557.51012.5 points50: 10.110.150100: 6.56.51001,000: 1.81.81,00010,000: 0.50.510,000Width of the 95% Wilson interval around an observed 98% success rate, by number of requests. Bar chart of Interval width (points) by Requests in the sample.4 rows, drawn from the table that follows this chart.5050: 10.110.1100100: 6.56.51,0001,000: 1.81.810,00010,000: 0.50.5051015 points
Requests in the sample 95% interval for an observed 98% Interval width (points)
50 89.5% to 99.6% 10.1
100 93.0% to 99.4% 6.5
1,000 96.9% to 98.7% 1.8
10,000 97.7% to 98.3% 0.5

At 100 requests, an observed 98% is consistent with anything from 93.0% to 99.4%. Put a 96% from another 100 requests beside it and its interval runs from 90.2% to 98.4%; the two overlap almost entirely, so the gap between them is not evidence of anything. At 1,000 requests each, the same two rates give 96.9% to 98.7% and 94.6% to 97.0%, and the intervals barely touch. The numbers come from this script, standard library only, run on Python 3.15.0:

from math import sqrt

def wilson(successes: int, n: int, z: float = 1.96) -> tuple[float, float]:
    p = successes / n
    centre = p + z * z / (2 * n)
    margin = z * sqrt(p * (1 - p) / n + z * z / (4 * n * n))
    denom = 1 + z * z / n
    return (centre - margin) / denom, (centre + margin) / denom

for n in (50, 100, 1_000, 10_000):
    lo, hi = wilson(round(0.98 * n), n)
    print(f"98% of {n:>6,}: {lo:6.1%} to {hi:6.1%}  (width {(hi - lo) * 100:4.1f} points)")
98% of     50:  89.5% to  99.6%  (width 10.1 points)
98% of    100:  93.0% to  99.4%  (width  6.5 points)
98% of  1,000:  96.9% to  98.7%  (width  1.8 points)
98% of 10,000:  97.7% to  98.3%  (width  0.5 points)

The interval assumes independent attempts. Requests fired in one burst at one target share whatever that target was doing at that minute, so a burst of 10,000 is less informative than the table suggests. Spread the sample across hours.

What counts as success

Counting HTTP 200 as success is the common mistake, and it always flatters the result. A challenge page, a consent wall and an empty result set can all arrive with status 200, and so can a page served to a suspected bot with the data stripped out. Some anti-bot systems do this on purpose, because a page that looks like success wastes more of a scraper’s time than an honest 403.

Define success by content instead: assert that a field you actually need is present and plausible, and count everything else as a failure whatever the status code says. The 403 vs 407 vs 429 guide has a section on detecting the block that returns 200, and the Python proxy guide shows the same check in code. A rate measured this way is lower, and it is the one worth having.

What a published rate has to state

Report one rate per target, too. Averaging an easy target with a hard one produces a number that describes neither, and the mix decides the answer: send more traffic to the easy site next week and the blended rate improves while nothing about the proxy changed. A single headline figure across many sites hides exactly the comparison a buyer needs. The list below is the minimum.

  • Sample size, so the interval above can be worked out.
  • Target, because a rate against an echo service says nothing about a defended retailer.
  • When it was measured, and over how long, because pools and defences both change.
  • The definition of success, status-based or content-based.
  • The counting rule for retries and timeouts.
  • Concurrency and location, since rates fall as parallelism rises and vary by country.

A vendor figure missing these is a statement of confidence, not a measurement. This site prints no provider success rate unless we ran the test ourselves and can state every item on that list; how we test sets out the method. To compare candidates yourself, the proxy type guide includes a script that sends the same URL through each one at the same moment.

Frequently asked questions

Why do provider success rates differ so much from what I measure?
Usually the target and the definition. A rate measured against undefended sites, counting any 200 as success, comes out far higher than one measured against a hard target with content checks. Ask which target, which dates and which definition.
How large a sample do I need?
It depends on the precision you want. Around an observed 98%, 100 requests leave an interval about 6.5 points wide and 1,000 requests about 1.8 points. Spread the requests across hours rather than one burst.
Should retries count as failures?
Decide, then say which. A first-attempt rate measures how often the proxy works at once; a rate after retries measures whether the job finishes. Both are useful, they can differ widely, and a figure that does not say which cannot be compared with another.

Sources

  1. NIST/SEMATECH e-Handbook of Statistical Methods, 7.2.4.1: confidence intervals for a proportion (Wilson method)
  2. Wilson, E. B. (1927). Probable Inference, the Law of Succession, and Statistical Inference. JASA 22(158)
  3. RFC 9110: HTTP Semantics, what a 200 response asserts

All terms, A to Z