Field audits: the scanner measured on the real web.
A benchmark on fixtures shows how a scanner behaves on pages built to test it. A field audit shows how it behaves on the web people actually ship. Each edition scans one set of live websites through production twice, before and after a round of accuracy fixes, then sends every finding the fixes removed back to the live page for a re-check.
Latest edition · September 2026 field audit V
Findings
1,263
from 1,438 · −12%
Removals judged live
577
of 3,741 removed · 19 genuinely lost
Sites compared
53
of 56 scanned, 5 pages each
Read the September 2026 field audit V in full: every engine, every site, and what it does not claim.
All editions
Each edition audits its own set of sites: 173 different websites across 5 editions, 69 of them in more than one. Compare before and after within a row, not totals between editions.
| Edition | Findings before | After | Sites compared | Ran | Builds |
|---|---|---|---|---|---|
| September 2026 field audit V | 1,438 | 1,263 | 53 | 2026-09-26 to 2026-09-27 | 3e26cc9 → 0bc595c |
| September 2026 field audit IV | 823 | 860 | 51 | 2026-09-22 to 2026-09-23 | 5364104 → 2d89faa |
| September 2026 field audit III | 926 | 928 | 50 | 2026-09-15 to 2026-09-16 | ee7f57b → 115e5b2 |
| September 2026 field audit II | 890 | 873 | 47 | 2026-09-14 to 2026-09-15 | d79703c → ee7f57b |
| September 2026 field audit | 1,158 | 800 | 45 | 2026-09-07 to 2026-09-08 | e400fec → edc3638 |
An edition's figures never change after it is published. A new audit is a new edition, and a correction to an edition's notes is dated on its page, so a link to one keeps showing what it said on the day.
How an edition is produced
- The same sites are scanned through production on the build before a round of fixes and on the build after.
- Sites that blocked the scanner in either run, or changed reachability between runs, are excluded and counted.
- Every finding present before and absent after goes back to the live page to be re-measured by a method independent of the scanner. The ones it cannot judge are counted as such, not dropped.
- Every table and total is computed from the raw result files by scripts and committed with the site. The notes that explain individual rows are written by hand.
What a field audit does not claim
- Not a false-positive rate: that needs every remaining finding re-verified, not only the ones that changed.
- Not "fewer false positives" by the percentage findings fell: not every removed finding was false.
- Not deterministic: live sites change markup and block scanners between runs.
The controlled half of the evidence is the precision and recall benchmark. The scanner audited here runs on every pull request through the accessibility GitHub Action