Green Scans Aren’t Enough: What Expert Accessibility Audits Catch That Tools Miss

A green accessibility score feels like closure. Lighthouse is happy. axe is quiet. Your automated WCAG scanner stops yelling about contrast. Then a shopper opens VoiceOver on a product listing, hits Load More, and the page goes silent—or worse, dumps them back on the button they already pressed.

That gap is not a tooling bug. It is a methodology limit. Automated scanners are excellent at catching markup problems. Expert audits—especially screen-reader walkthroughs against WCAG 2.2—catch interaction and announcement problems. On headless Shopify Hydrogen storefronts, where drawers, filters, carousels, and third-party widgets compose the purchase path, that second layer is where conversion and compliance actually live.

This post opens a four-part craft series drawn from real remediation on furniture ecommerce brands running sibling Hydrogen storefronts. For the expert layer, the retailer contracted Converge Accessibility—a digital accessibility firm known for pairing technical WCAG testing with legal and policy expertise—to deliver a deep WCAG 2.2 audit that goes beyond what SortSite, axe, and Lighthouse typically surface. We will stay practical: what tools miss, how we triage findings, and why manual screen-reader acceptance criteria belong in your definition of done.

Two layers, not one checkbox

On a recent national mattress retailer’s Hydrogen program, accessibility work ran as a deliberate two-layer stack:

• Automated WCAG scanning — contrast failures, many Level A/AA markup issues, illegal ARIA naming patterns. Useful, fast, and incomplete.

• Expert manual / screen-reader audit — Converge’s WCAG 2.2 report, itemized into engineering tasks, then retested with VoiceOver by client and agency teams.

The Converge deliverable became a task factory: each spreadsheet line turned into a tracked remediation item with severity, page area, and a retest path. That is the opposite of “fix whatever Lighthouse flags this sprint.” It is a backlog you can prioritize, port across sibling brands, and defend in stakeholder reviews.

Kickoff conversations also split creative vs developer ownership and flagged vendor outreach early—especially for hosted checkout HTML and chat widgets. That triage, done before the first pull request, saved weeks of false “we own this” optimism.

What scanners miss on Hydrogen storefronts

Expert findings clustered around behaviors automated tools rarely model well:

• Selected and expanded state. Size buttons, multi-select options, and store pickers that look “active” visually but announce as nameless buttons—or never declare that a choice stuck.

• Focus after dynamic UI. Load More that leaves keyboard and VoiceOver focus on the control instead of the first newly revealed product; filters and overlays that trap focus behind a dialog.

• Announcement ordering. Cart drawers that interrupt their own dialog labels with a live “N items in cart” blurt; polite live regions that fire once and go quiet on the next append.

• Landmark noise. Multiple banner or contentinfo landmarks from nested React headers and footers—valid HTML in isolation, confusing document structure for screen-reader users.

• Silent commerce chrome. Star ratings without accessible names; sale prices read as an ambiguous “From $A $B”; skip links that exist in the DOM but never become visible on focus.

None of those are exotic. They are the everyday surfaces of a furniture PLP/PDP/cart journey. Scanners may pass the same pages because the DOM is “complete enough.” Screen-reader QA fails them because the experience is incomplete.

The loop goes both ways. After a VoiceOver pass on one promo pattern, an automated scan later flagged an illegal aria-label on a generic role—proof that a “passing announcement” can still violate ARIA authoring practices and create a useless tab stop. Remediation is a dual-tool conversation, not a single green badge.

Triage: Critical, Major, and Out-of-Access

Severity alone is not a plan. Ownership is.

We bucket Converge items into paths the team can actually execute:

1. Ship in the accessibility wave — storefront-owned markup and behavior: labels, focus moves, landmark consolidation, skip-link visibility, cart dialog semantics.

2. Design change needed — scanners and screen readers agree something is wrong, but the fix is visual: heading structure that requires content redesign, focus rings that fail on dark imagery, active-tab shading below 3:1.

3. Hard to fix / regression-prone — decorative image alt treatment that flips back across Oxygen previews and sibling brands; card semantics that fight Tab vs VoiceOver browse modes.

4. Out-of-Access / vendor-owned — Shopify checkout HTML (discount/payment order vs progress controls, quantity icon confusion, payment radios that hide expandable content), chat widget keyboard and live-message gaps, cookie-banner timing quirks.

Out-of-Access does not mean “ignore.” It means escalate with evidence, keep the ticket Ready-for-QA so you verify vendor changes, and stop burning sprint capacity on HTML you cannot ship. Constructive vendor letters beat silent debt.

Why screen-reader acceptance criteria matter

For Load More–style patterns, acceptance was not “focus moved somewhere.” It was a sequence: announce how many products loaded, announce which item received focus, repeat on every subsequent Load More, and scroll the focused card into view. Client VoiceOver feedback caught previews that only announced a count once, then went silent—exactly the kind of intermittent polite-region failure automated tools will not schedule for you.

Practitioners also have to decide how they QA. The same product-card markup can announce a title twice under Tab and once under VoiceOver rotor navigation. Debating which method is the acceptance test is craft, not pedantry. Write it down. Otherwise “VoiceOver passed” means whatever the last person happened to try.

One line of team discipline that keeps showing up in this work: improve the accessibility experience, not just the scan. And when ARIA fights you, remember the craft rule—no ARIA is better than bad ARIA.

Methodology you can steal this quarter

1. Run your automated WCAG layer on a schedule—but treat green as a floor, not a finish line.

2. Commission (or schedule) an expert WCAG 2.2 screen-reader audit on critical journeys: PLP, PDP, cart, account, store locator.

3. Turn the report into tasks with severity, surface area, and ownership—including an explicit Out-of-Access bucket.

4. Define VoiceOver (and, where relevant, NVDA) acceptance criteria per pattern before code review, not after deploy.

5. Retest on shared previews with client screen-reader users; port proven fixes to sibling Hydrogen storefronts through a shared package, not copy-paste forks.

6. Re-scan after manual passes—illegal ARIA and contrast regressions love to sneak in behind “good” announcements.

What’s next in this series

Post 2 goes deep on Load More that survives VoiceOver: focus, live regions, and announcement order. Post 3 covers ARIA discipline—promo badges, product cards, and illegal names. Post 4 is the multi-brand story: one Converge backlog, two storefronts, without forking forever.

If you are mid-Hydrogen build—or mid-remediation after a letter—and your scans look fine while real shoppers still struggle, we should talk. Endertech pairs expert audit findings with remediation on headless Shopify stack work. Start at endertech.com/contact, explore our Shopify Hydrogen solutions, or skim our overview of ADA-minded practices on Hydrogen before the deeper craft posts land.

Series note: Client brand names are anonymized. Converge Accessibility is named with approval as the contracted expert auditor. Tool and pattern details are paraphrased from internal remediation work—not pasted from tickets.