Transparent review method
Control the capture. Describe the evidence.
Visual localization QA is most useful when a reviewer can separate language-driven changes from unrelated rendering noise and another person can reproduce the result.
Scope the question
Define the route, product state, target locale, viewport, build, theme, and expected source screen. Decide whether the review is looking for truncation, overlap, missing strings, changed hierarchy, bidirectional layout, glyph rendering, or a broader regression.
Do not describe the process as exhaustive unless the tested screen-state and viewport matrix is actually exhaustive.
Control capture variables
| Variable | Keep stable | Why it matters |
|---|---|---|
| Viewport and zoom | CSS width, height, scale, device pixel ratio | Reflow can otherwise look like a locale defect. |
| Content fixture | Names, totals, dates, inventory, images | Dynamic content creates unrelated pixels. |
| Theme and state | Colour mode, focus, hover, validation, expansion | State styling can dominate a difference view. |
| Fonts | Loaded family, weight, fallback, rendering host | A fallback font changes width, metrics, and glyph shape. |
| Timing | Animations paused, data settled, fonts ready | Intermediate frames are not comparable evidence. |
Use each view for one job
- Side-by-side review
- Establish overall structure, reading order, missing regions, and unexpected reflow.
- Overlay or reveal
- Trace displaced boundaries and inspect whether a component moved with its content.
- Pixel difference
- Locate change. Thresholds suppress small channel differences but do not determine correctness.
- Zoomed inspection
- Confirm clipping, collisions, punctuation placement, diacritics, and baseline behaviour at the source resolution.
Classify findings by observed impact
- Blocker: a critical action or required information is unusable in the tested state.
- High: important content is unreadable, misleading, obscured, or materially misordered.
- Medium: the defect impairs comprehension or polish but a practical path remains.
- Low: a visible inconsistency has limited task impact.
- Observation: evidence worth recording without enough support to call it a defect.
Severity is a team decision. The workbench records the reviewer’s assessment; it does not calculate product risk.
Close the loop with humans
Ask a qualified language reviewer to confirm meaning, grammar, terminology, tone, cultural fit, and locale conventions. Ask an accessibility specialist to assess semantics and assistive technology. Re-capture a fixed build under the original conditions, then attach the before-and-after evidence to the team’s issue system.