Last spring we shipped a portfolio site for Lumen Press, a small publishing studio in Lviv, and ran it through axe DevTools before launch. Ninety-eight out of one hundred. No critical errors, no serious warnings. Two weeks later a reader emailed Lumen’s editor to say she couldn’t tab past the hero section. The scanner had missed the one thing that mattered most: our custom focus ring was invisible against the animated gradient background we were so proud of.
That email changed how we test everything now. Automated audits catch missing alt text and insufficient color contrast on static elements. They almost never catch what happens when contrast shifts during an animation, when a focus outline gets painted under a moving layer, or when a transition fires before a screen reader has finished announcing the previous state.
What the scanner sees and what the user sees
Lighthouse and axe both evaluate a page in a frozen state. They check the DOM as it exists at one moment, usually right after load. Our sites rarely sit still. The Lumen homepage had a slow parallax drift on the header image and a focus ring styled with a 2px solid outline in a color that read as a pale gray, chosen to look quiet against the paper-white background. Against the gradient, which cycled through a cream-to-blush range over eight seconds, that same gray dropped below a 2:1 contrast ratio for roughly a third of the cycle.
The scanner ran during the cream phase and passed it. A keyboard user tabbing through the nav at the wrong second saw nothing.
Fixing it without killing the gradient
We didn’t want to remove the animation. Instead we gave focus states their own visual layer, independent of whatever was happening behind them. The new outline is a 3px solid line in Arche’s signature ochre, rendered with a drop shadow that gives it a dark halo regardless of background color. It no longer tries to blend in. It interrupts on purpose.
Oksana Tkach, who leads visual design on most of our client work, put it plainly during the postmortem: a focus indicator that looks elegant against every background is usually one that’s too faint against most of them.

Reduced motion is not a toggle, it’s a redesign
Every Arche site respects prefers-reduced-motion, but for the first two years we treated it as a switch: motion on, motion off. That’s lazier than it sounds. Turning off a transition entirely can leave content appearing in a jarring snap, or worse, leave state changes with no visual signal at all. A dropdown that used to ease open over 300ms would, with motion off, simply exist or not exist, with no cue that anything had changed.
For Variant Studio’s site, a branding agency in Kraków we worked with last autumn, we rebuilt every transition as a pair. The full version uses a custom cubic-bezier curve and runs for 240 to 400ms depending on the element. The reduced version keeps a minimal fade, usually 80ms of opacity change, enough to register as a state shift without any spatial movement. Nothing teleports. Nothing stutters.
Testing with the setting actually on
We keep a dedicated laptop in the studio with reduced motion switched on at the OS level, permanently. Dmytro Pavlenko, who writes most of our CSS, checks every new component on that machine before it goes near a client review. It’s a small ritual but it catches things code review doesn’t, like a loading spinner that was only animated, meaning that with motion reduced it just vanished and left users staring at a blank card with no indication anything was loading.
Contrast during transition, not just at rest
WCAG AA asks for a 4.5:1 contrast ratio for body text and 3:1 for large text and UI components. Most design tools check this against a single swatch. Our sites use animated backgrounds on at least a third of client projects, which means contrast is a moving target, not a fixed value.
For a recent project with Kadrova, a Kyiv-based recruiting platform, the hero text crossed from a light section into a darker one as users scrolled, with the background transitioning over roughly 600 pixels of scroll distance. We measured contrast at five points along that transition, not just the start and end, and found a dip below 3:1 right around the midpoint where the two background colors mixed into a muddy gray-green. The fix wasn’t a different color. It was a subtle text shadow that held contrast steady across the entire scroll range, regardless of what was happening underneath.
An audit tells you the page passed. It doesn’t tell you whether someone could actually use it while scrolling, tabbing, or reading at their own pace.
What a checklist can’t catch
We now run a short manual pass on every project before handoff, on top of the automated scan. It covers the things that don’t show up as line items in a report:
- Tab through every interactive element with the monitor brightness lowered to simulate glare conditions
- Check focus visibility at three points during any looping or scroll-triggered animation, not just at rest
- Turn on reduced motion and reduced transparency together, since many sites only test one at a time
- Read every ARIA live region announcement aloud, in order, to catch overlapping or redundant speech
- Resize text to 200% and confirm no focus ring gets clipped by an overflow: hidden container
The overflow problem nobody flags
That last point cost us a full day on the Variant Studio rebuild. Several cards used overflow: hidden to keep rounded corners clean during hover animations. Once text was resized, focus outlines on buttons inside those cards got cropped at the edge, sometimes to nothing. No automated tool flags this, because the outline technically exists in the DOM. It’s just invisible.
Where this leaves our process
Accessibility work used to happen at the end of a project, as a pass before launch. It’s now built into the same Figma components and CSS variables we use for everything else, which means focus states and motion-reduced variants exist from the first prototype, not bolted on afterward. It’s slower by maybe two or three hours per project. Given what one missed focus ring cost Lumen Press in frustrated readers, that trade feels obvious now.
We still run the automated scanners. A 98 is better than a 60. But we no longer treat the number as the finish line, just as one data point among several that include an actual hand on an actual keyboard, moving through the actual site.
Leave a Reply