Exactly what we check.
And what we don't.

Every free scan runs 16 rule-based checks against your page's HTML, each mapped to a WCAG success criterion, each weighted by severity. Nothing here is guessed — a check either fires or it doesn't, based on parsing the markup your server actually sends.

Every one of these is a property you can read straight off the parsed HTML: an attribute is missing, a heading number jumped, an id repeats. That's also their limit — a check can only exist here if the answer is decidable by parsing markup alone, with no judgement calls and no rendering.

What the deep AI audit judges

The deep audit is a different kind of check. Instead of reading text, it downloads the actual image file behind each described image and sends it to a vision model alongside the alt text, then asks a narrower, harder question: is this true? That costs us roughly 5–15¢ per page in vision-model calls, depending on image size — vision models are billed by pixel area, so a page of large hero images costs several times what a page of thumbnails does, which is why it isn't unlimited and isn't free at scale. It opens up to six meaningful images per page, not all of them — the rule checks still cover every image; the capped step is specifically the "is this true?" judgment. Paid plans run it weekly on the pages you nominate; the free tier caps it at one audit a day.

This only works because the image file is actually opened. Every check in the table above it works from text alone and would score "IMG_4021.jpg" as present and non-empty. It takes looking at the picture to know that's not alt text, it's a filename.

What GotAlt cannot detect

This is the part most scanners bury in a footnote. We're putting it in a heading, because a tool that won't tell you what it misses is more dangerous than no tool at all — it lets you believe a passing report means something it doesn't.

How much of WCAG this actually covers

Sources disagree on the exact number, but nobody credible claims automation covers most of WCAG. Karl Groves, who has spent longer measuring this than almost anyone, put it precisely: an automated tool "can definitively test for approximately 25–29% of best practices for WCAG 2.0" and "cannot test for approximately 40%" — the rest falls into a grey zone that needs a human judgement call either way. karlgroves.com/automated-lies-with-one-line-of-code

Different tools and researchers land on different figures because they're measuring different rule sets against different versions of WCAG, but roughly a third machine-detectable is the defensible, honest framing — not a round "half" or "most." We didn't build 16 rule checks plus a vision-model alt text audit because that gets you to 100%. It doesn't. It gets you further than reading HTML alone, and it's honest about the rest.

How the score is calculated

Every page starts at 100. For each rule check that fails, we count how many times it fails on that page, capped at 5 instances so one repeated issue — say, 40 images missing alt text — doesn't drown out everything else on the page. That capped count is multiplied by the check's severity weight: critical is worth 10 points, serious 5, moderate 2, minor 1. All of those deductions are summed and subtracted from 100, floored at zero. A site's overall score is the average of its page scores.

So a page with one critical issue (an image missing alt, capped at one instance for this example) loses 10 points and scores 90. A page with five instances of a serious issue loses 25 points and scores 75.

Say this plainly: the score is a prioritisation aid, not a compliance measure. It tells you where to spend an hour first. It cannot tell you whether a site meets WCAG, the EAA, the ADA or any other legal standard — no automated score can, because roughly two-thirds of what those standards require isn't something a machine can check at all.