Methodology: 2026 ADI Lab Benchmark

Published: December 15, 2025 · Updated: September 30, 2026

This page documents how the 2026 benchmark was designed, what was measured, and where the results stop being useful. The short version: 16 browsers, 800 profiles each, 47 checks, 24 login flows, three proxy countries, and a 90-day account observation window.

1. Sample design

Browsers: 12 commercial anti-detect browsers, each on its current stable release at the start of the test window.
Profiles: 800 per browser, split evenly between Facebook (400) and Instagram (400).
Waves: 16 waves of 50 profiles per browser. Each wave used the same three sticky residential proxy countries and the same 24 login flows.
Repetitions: every performance measurement was repeated 10 times; the median is reported.
Observation: accounts were observed for 90 days from first login. Status was recorded at day 90.

The sample is balanced by browser and platform, not by use case. A team running 5 profiles and a team running 500 profiles should weight different metrics.

2. Hardware and network

AMD Ryzen 9 7950X, 64 GB DDR5-6000, Windows 11 24H2, 1 Gbps fiber.
Sticky residential proxies in the Netherlands, Germany, and the United States.
One browser instance per profile at a time for fingerprint checks; concurrency was measured separately on an isolated rig.

3. The 47 fingerprint checks

The checks cover browser identity, display, locale, graphics, fonts, hardware, input, network, permissions, media, browser surface, automation artifacts, transport, and isolation. The full list with pass criteria is in checks.json.

The checks are grouped into two reporting levels:

Check-level pass rate: the percentage of all 47 checks passed across all 800 profiles for a browser. Incogniton: 97.8%, or 46.0 of 47 checks per profile on average.
Full-profile pass rate: the percentage of profiles that passed all 47 checks at once. Incogniton: 91.4%.

The two metrics answer different questions. Check-level tells you how often an individual signal is clean. Full-profile tells you how often a whole identity is clean at the same time. Full-profile is stricter and more operationally relevant, because one leaked signal can be enough to link an account.

Why the numbers diverge. Most failing profiles failed a cluster of related checks, usually after a proxy timeout, a kernel mismatch, or a partial profile restore. Isolated single-check failures were uncommon. That is why a 97.8% check-level rate can coexist with a 91.4% full-profile rate without an arithmetic error.

4. Account outcomes

Facebook and Instagram: 400 profiles per platform per browser, observed for 90 days.
Ban: the account was disabled or permanently restricted by the platform during the window.
Other loss: a permanent checkpoint, identity-verification loop, or account closure that prevented normal use but was not a platform ban.
Survival: the share of profiles still active and unrestricted at day 90. Survival is not 100 minus the ban rate; it also counts other losses.
Combined ban rate: the unweighted average of the Facebook and Instagram rates, not a pooled count. This avoids letting the platform with more profiles dominate the summary.

Ban counts are small. At 400 profiles per platform, a 1.2% rate is about 5 bans and a 1.6% rate is about 6. The 95% Wilson intervals overlap substantially. Treat the top-three ordering as directional.

5. Performance

Cold-start launch: time from process launch to a usable first page with the proxy connected, measured 10 times per profile. Median and p95 are reported.
RAM per profile: resident memory after the first page load, measured per profile.
Stable concurrency: the highest number of simultaneous profiles that completed a 30-minute scripted workflow without a launch failure, crash, or proxy drop on the test rig.

Proxy handshake and target-site load time often dominate perceived speed. A 70 ms cold-start difference is small compared with a 1.4-second proxy handshake.

6. Statistical treatment

The benchmark is descriptive. We do not run a causal model, and we do not claim that a browser caused an account to survive. We report point estimates, sample sizes, and the practical uncertainty around the top-three comparisons. We do not rank differences smaller than the rounding precision of the metric.

7. Limitations

Single hardware platform and OS build.
Three residential proxy countries only.
Two platforms in the ban analysis, not all 24 login flows.
Small ban counts and wide confidence intervals.
Vendor-supplied review licenses; no vendor payment for placement.
90-day snapshot; platform policy and browser kernels change.

8. Ethics and legal

The test used profiles for legitimate research into fingerprint consistency and account survival. It did not automate platform logins, bypass access controls, or publish account credentials, cookies, or personal data. The profile-level dataset is anonymized. Readers are responsible for complying with platform terms and local law.

9. Reproducibility

Questions or corrections: research@antidetectdata.com.