Skip to content
AnglinAI Docs

TheAccessible.org Audit + PDF — Custom GPT

You can run full WCAG accessibility audits, remediate PDFs, and generate Accessibility Conformance Reports (ACRs) directly from a ChatGPT conversation — without leaving the chat.

What you can do

Scan websites for WCAG violations

Give the GPT a URL and it scans the page in headless Chromium across desktop, dark-mode desktop (when the page supports it), and mobile viewports using axe-core + WCAG validator. For multi-page audits you can request a site crawl (up to 25 pages).

Convert PDFs to accessible deliverables

For every PDF you convert, the GPT produces three things:

  • Accessible HTML — a semantic, tagged web rendition.
  • Accessible (tagged) PDF — proper structure tree, reading order, alt text, language, and document title.
  • PDF/UA-1 conformance report — automated veraPDF check.

Fix accessibility issues

For websites: the GPT identifies violations, provides CMS-specific fix instructions (WordPress, Squarespace, Webflow, Wix, Shopify, Canvas, Drupal, Ghost), and can walk you through a bulk alt-text review.

For PDFs: the GPT edits the underlying HTML and regenerates the accessible PDF automatically — no Acrobat required.

Generate an Accessibility Conformance Report (ACR)

After a URL scan, the GPT walks you through each WCAG criterion that needs human review, presents an AI pre-grade with reasoning, and asks you to confirm or override one at a time. Once all criteria are decided, it produces a signed ACR HTML + PDF you can hand to legal or procurement.

It can also produce a focused Alt-Text Remediation Report — a polished document your content editors can use to apply alt-text changes in their CMS.

Prerequisites

  • A ChatGPT account that can use Custom GPTs (Plus, Team, Enterprise, or Free with limits).
  • An API key. Get one from pdf.theaccessible.org/settings?tab=api-keys — click Create key, give it a name, and copy the value. The key is shown only once. Paste it when the GPT first asks.
  • For URL scanning: the URL must be publicly reachable (the scanner runs in a cloud headless browser).
  • For PDFs: the PDF must be reachable by a public URL or small enough for ChatGPT to send inline.

Open the GPT

Open TheAccessible.org Audit + PDF in ChatGPT.

A direct link will be added once the GPT is published in the GPT Store. Until then, use the MCP server directly.

Scan a website

Say something like: “Scan https://example.com for WCAG 2.2 AA issues.” The GPT will:

  1. Start the scan and give you an estimated wait (typically 30–90 seconds for a single page).
  2. Poll progress in the background.
  3. Present the findings: violations grouped by WCAG criterion, a VPAT conformance score, and (for CMS sites) editor-specific fix steps rather than raw HTML diffs.
  4. Offer to walk you through fixing alt text, then generate an ACR.

For multi-page audits add: “…and crawl the whole site, up to 20 pages.” The GPT will discover URLs via your sitemap, charge one credit per page, and present aggregate results.

Convert a PDF

Just send the GPT your PDF — drag-and-drop, paste a public URL, or say “convert this for me”. It will:

  1. Confirm the source.
  2. Start the conversion (usually 30 seconds to a few minutes).
  3. Hand you: a download link for the accessible PDF, the HTML preview, and a one-paragraph summary of the PDF/UA-1 conformance verdict.

Fix accessibility issues in a website

After a scan, tell the GPT what to fix:

  • “Walk me through fixing the alt text on this page.”
  • “The nav landmark is missing a label — how do I fix that in Squarespace?”
  • “There are 12 color contrast failures — what are my options?”

The GPT translates each violation into steps inside your specific CMS editor when the platform is detected with high or medium confidence.

Fix accessibility issues in a PDF

Tell the GPT the issue:

  • “The chart on page 3 should have alt text ‘Q3 sales by region’.”
  • “That decorative divider image is being announced — mark it decorative.”
  • “Set the document language to Spanish.”

The GPT will locate the element, show you the proposed change, apply the edit, regenerate the accessible PDF, and re-run PDF/UA-1 automatically.

Generate an ACR / VPAT

After a URL scan is complete, say: “Let’s generate an ACR.”

  1. The GPT fetches the WCAG criteria needing human review.
  2. For each criterion it presents: the artifact extracted from your page (heading outline, form labels, image alt text, etc.), an AI pre-grade with its reasoning, and asks you to Confirm or Override. It goes one at a time — this is intentional; the audit trail requires individual decisions.
  3. Once all criteria are reviewed, it asks for your sign-off details (name, role, organization, date).
  4. It produces a signed ACR HTML document and (if PDF rendering is available) a PDF — both as clickable download links valid for 1 hour.

Troubleshooting

“I don’t have a public URL for my PDF” Drag-and-drop the file into the chat instead.

“The scan is stuck” Ask “what’s the status?” and it will re-poll. If stuck for more than 5 minutes, ask it to start over.

“The GPT approved all ACR items at once without asking me” That’s a system-prompt drift issue. Report it via [email protected] and we’ll fix the GPT instructions.

“acr_getScreenshot returned an error” Your scan predates the element-screenshot feature (Phase 6). Run “rescan this URL” and try the ACR queue again.

Limits

  • PDF max size: 25 MB.
  • URL scan: 1 credit per page. Multi-page crawl: 1 credit per page in the crawl.
  • Alt-text batch: 1 credit per 10 images (AI suggestions via Claude Haiku).
  • Download URLs: expire 1 hour after they’re generated. Re-call the result endpoint to get fresh URLs.
  • Conversion time: typically 12 seconds per PDF page; URL scans typically 30–90 seconds.

Privacy

  • Your PDF and derived HTML/PDF are stored in our private R2 bucket, scoped to your user account.
  • Scanned page HTML is stored per-job for remediation purposes.
  • ACR decisions are stored in our database under your user ID.
  • Your API key is stored by ChatGPT; it’s not visible to other users.

For the full data-handling story see the privacy policy.

Power users: skip the GPT

Use the OpenAPI spec directly, or connect Claude Desktop / another MCP client to the MCP HTTP server at https://api.theaccessible.org/api/mcp. Same 21 tools, same deliverables, no chat in the middle.