Web accessibility

A dated record of what actually happens when an agent requests public pages: how many answer, how many refuse, and on what grounds. The numbers are counts from real campaigns, not estimates.

They are published so that a future run — or a future model — can compare rather than guess, and so that anyone can contradict them with their own measurements.

2026-06-27 — Commercial search campaign, 23 sites (FR-77)

Context: real purchase research (reconditioned game consoles, budget ≤ 200 €), francophone e-commerce sites. Diwall v1.14.0, no --stealth (introduced later, v1.15.0).

23 sites were targeted; 22 results were recorded. The four counts above sum to 22 and the shares to 95.7 % — the outcome of the twenty-third was never written down. The URL list was not kept either, so it cannot be recovered. The shares are shares of the panel targeted, left as they are rather than rebased on 22: nothing establishes that the twenty-third site went untested, only that its result went unrecorded.

Accessible sites were both SSR without a WAF. Blocked sites were major retailers and refurbished-goods marketplaces. Individual site identities were not logged in this run — only the aggregate is available (see the field notes, FR-77).

2026-07-02 — Stealth benchmark, quantitative (FR-79)

Context: --stealth was found non-functional in production since v1.15.0 (a playwright-stealth 2.x API break silently no-op’d it). After the fix (v1.16.0), measured on bot.sannysoft.com (scenarios/test_stealth.json).

The two columns do not describe the same number of checks: 12 + 18 = 30 without --stealth, 0 + 31 = 31 with it. The benchmark page renders some checks only when a signal is present, which plausibly accounts for the difference — but that was not verified at the time, so these are reported as counts and no total is claimed. In particular this is not a “31 of 31” result.

What this confirms: --stealth, once actually applied, clears every basic JS/browser fingerprint check on this benchmark. What it does not answer: whether this would have changed FR-77’s 39% block rate — that panel’s 23 URLs were never individually logged, and reconstructing a commercial panel without a genuine purchase intent was declined as indistinguishable from a load test on third-party infrastructure (see the field notes, FR-79, for the full reasoning). TLS fingerprinting (JA3/JA4) and Cloudflare Enterprise behavioural analysis remain untested by this benchmark — playwright-stealth does not claim to cover either.