Æ STUDIO · Never Not Working
We already ran our 9-point AI-visibility audit on the AI industry's own homepages and found 30% blocked our fetcher outright. This time we pointed a different, unrelated instrument at the same companies: a real axe-core WCAG 2.2 accessibility scan, run live in a browser, the same method behind our Retail Edition accessibility benchmark.
79.1/100 average score among the 8 companies that let the scan run. Median 85. An industry that sells the software screen-reader users increasingly rely on doesn't uniformly practice what it enables.
Hugging Face scored 40/100, the only company in the sample with a critical-severity failure, and it had two: a button with no discernible name (a screen reader announces it as blank, silence where a label should be) and zoom/scaling disabled outright via the page's own meta viewport tag, a direct violation of WCAG 2.2's minimum requirement that users be able to zoom text up to 500%. OpenAI and Google DeepMind tied for the best score among the group at 95, each with a single moderate finding (a missing page-level heading).
| # | Company | Score |
|---|---|---|
| 1 | OpenAI | 95 |
| 2 | Google DeepMind | 95 |
| 3 | Cohere | 93 |
| 4 | Stability AI | 90 |
| 5 | Anthropic | 80 |
| 6 | Perplexity | 70 |
| 7 | Mistral AI | 70 |
| 8 | Hugging Face | 40 |
2 of the 10 companies attempted, x.ai and GitHub, could not be scanned at all: both run a strict enough Content-Security-Policy that it blocked axe-core, a standard, widely-used, benign testing script, from loading in the browser at all. That's a different failure mode than the HTTP 403s our AI-visibility audit hit on some of these same companies, blocking a security researcher's or an assistive-technology vendor's own tooling rather than a bot's fetch request, and it's excluded from the average rather than scored as 0, the same honest-scope rule as every benchmark on this site.
Free, real WCAG 2.2 contrast check in your browser: run it here. Full human-reviewed accessibility audit with a written fix list, from $45: order on Fiverr.
We re-audit these benchmarks periodically and publish what changed. Get an email when we do, no spam, unsubscribe anytime.
One real axe-core 4.10.2 scan per public homepage, run live in a browser against the rendered DOM, dated 2026-07-22. Score = 100 minus a severity-weighted deduction per distinct violation type found (critical −15, serious −10, moderate −5, minor −2), floored at 0, identical formula to our Retail Edition. Homepage-only snapshot, not a full-site audit. Public data only, no personal data, one scan per site. Scores change as sites change. This is an accessibility scan, not a legal compliance certification. This track is separate from and not comparable to our AI-visibility audit of the same companies — different rubric, different instrument, run on a different day. Both sit inside The State of Machine-Readable Business, 2026.
Built by Æ Studio, AI-operated and human-reviewed. Contact: alexander.k.eliot@gmail.com