The specification exists. The buyer still has to read it.
A muted caption names the operating limit. A help sentence explains the required file. A heading crosses a photograph. Each puts different text against a different surface.
Start with that pair.
A brand palette gives you possible colors. A contrast review records the colors behind a particular statement, in a particular page state. Keep the evidence close enough that another reviewer can repeat the comparison.
The worksheet below is a proposed review method. The product page, measurements, and revisions are fictional. They describe no customer engagement, live site change, or business result.
Set the Review Boundary
Choose a page and a reading task before opening a color tool.
For this walkthrough, a fictional supplier publishes a product specification page. The buyer needs to identify the supported file format, read a diagram caption, understand an upload instruction, and find the installation guide heading.
Write those four statements in the review record. Assign each an evidence identifier. A screenshot of the whole page helps locate them. A close view shows the relevant text and its immediate background.
Record the page version, browser version, viewport, theme, and interaction state. Use the actual values during a real review. Leave an unknown browser version unresolved until someone checks it.
This article isolates text contrast. Review keyboard interaction separately. Use the website zoom and reflow review when enlargement or layout hides information.
A single contrast record cannot establish accessibility for the whole page. Report its boundary with the result. Name the text and states examined rather than labeling the entire website accessible.
Select the Criterion Before the Verdict
WCAG 2.2 SC 1.4.3, Contrast (Minimum), is Level AA. It requires at least 4.5:1 for ordinary text and images of text. Large-scale text has a 3:1 minimum. W3C criterion.
Large-scale means at least 18 point, or 14 point bold, with equivalent sizing for Chinese, Japanese, and Korean fonts. Document size and weight. A heading tag alone does not establish large text. W3C definition.
The criterion exempts logotypes and qualifying incidental text. Incidental cases include inactive controls, pure decoration, text invisible to everyone, and text within a picture containing other meaningful visual content. Ordinary product instructions do not become incidental because the designer makes them faint. W3C exceptions.
The installation heading in this walkthrough is page text placed over an image. Keep it in scope for the reading task.
W3C's Understanding page explains that thresholds must not be rounded upward. A calculated 4.499:1 fails a 4.5:1 minimum. It also directs reviewers toward underlying foreground and background colors rather than antialiased letter edges. W3C explanation.
Put the chosen criterion in every worksheet row. Keep a separate field for an organization's stronger design target. A design preference and a criterion threshold answer different questions.
Determine What Sits Behind the Words
For each text sample, inspect the foreground and the surface visible behind it. Trace transparent containers until the visible background is identified. Record where the values came from.
For the fictional specification, assume opaque foreground text on an opaque white panel. The record names both CSS values and confirms that no overlay changes the pair. That gives the calculation a reproducible input.
If opacity, blending, a gradient, or an image complicates the surface, leave the row open until the displayed pair is established. Do not substitute the nearest palette swatch and call the measurement complete.
Use the browser's computed styles to locate authored colors. Inspect parent layers where needed. Save the relevant declarations with the evidence. Note any assumption the measurement still depends on.
For imagery, identify the area behind the letters in the actual crop. Inspect each background variation that the text crosses. A clear corner elsewhere in the image says nothing about the area behind a word.
Treat an animation, responsive crop, or theme change as another condition to examine. A result for one fixed frame cannot establish every frame. If the possible backgrounds remain uncontrolled, record that limit and assign a design decision.
Work Through Four Fictional Text Samples
Assume all pairs below use opaque sRGB colors. The ordinary text examples use 16 CSS pixels at normal weight. The image heading uses 24 CSS pixels at normal weight.
At the CSS unit conversion, 24 CSS pixels equals 18 points. That places this normal-weight heading at the large-text threshold. Physical screen pixels are a different measurement. W3C size guidance.
These are invented inputs. The ratios are calculated from them using WCAG's relative-luminance and contrast-ratio definitions. Display values below are shortened; verdicts use the unrounded calculations. W3C calculation definitions.
Product specification. The sample reads “Supported input: CSV.” Its foreground is #333333 on #FFFFFF. The panel's computed background is white. The ratio is approximately 12.63:1. Record a pass against the ordinary-text threshold for that pair.
Diagram caption. The sample reads “Connector locations, rear view.” It uses #777777 on #FFFFFF. The ratio is approximately 4.478:1. Record a failure against 4.5:1. Displaying the rounded value as 4.5 would conceal the failed comparison.
Muted help text. The sample reads “Remove account identifiers before uploading.” It uses #999999 on #FFFFFF. The ratio is approximately 2.849:1. Record a failure. Keep the instruction in scope because the reading task requires it.
Text over an image. The heading reads “Installation guide.” Its foreground is #FFFFFF. For this simplified synthetic image, the crop behind the letters contains only flat #666666 and #AAAAAA regions. The ratios are approximately 5.742:1 and 2.323:1. The lighter region fails the large-text threshold.
A single average background would erase that difference. Record both regions and identify the letters over the weaker pair. In a real photograph, inspect the full relevant area instead of assuming two sample points exhaust its variation.
Use a Contrast Review Worksheet
Give each row one sample in one state. Keep long evidence notes in a referenced record so the worksheet remains readable.
| Page and state | Text sample | Color pair | Ratio | Criterion | Evidence | Retest result |
|---|---|---|---|---|---|---|
| Product specification, default | Supported input: CSV | #333333 / #FFFFFF | 12.63:1 display | SC 1.4.3, 4.5:1 | C01, computed opaque pair | No revision proposed |
| Diagram, default | Connector locations, rear view | #777777 / #FFFFFF | 4.478:1 display | SC 1.4.3, 4.5:1 | C02, caption styles | Proposed revision untested |
| Upload instructions, default | Remove account identifiers before uploading | #999999 / #FFFFFF | 2.849:1 display | SC 1.4.3, 4.5:1 | C03, text and panel styles | Proposed revision untested |
| Image heading, wide crop | Installation guide | #FFFFFF / #666666 and #AAAAAA | 5.742:1 and 2.323:1 display | SC 1.4.3, 3:1 | C04, crop regions and type size | Proposed revision untested |
| [Page, release, state] | [Exact words] | [Foreground / background] | [Full calculation reference] | [Criterion and threshold] | [Capture and source values] | [Version, observation, open limits] |
Add size and weight to each evidence record. Include the calculation's full precision there. The table's short ratio is a reading aid, never the stored basis for a borderline verdict.
Keep exception decisions explicit. If a reviewer excludes text, name the exception and the evidence supporting it. Leave a disputed classification open for review.
Propose a Revision Without Claiming a Live Fix
For the fictional caption and help text, propose #666666 on white. That synthetic pair calculates to approximately 5.742:1. It clears the ordinary-text minimum under the stated opaque-color assumptions.
The proposal does not prove that a stylesheet was changed. It also does not prove that the new color reaches every instance. Keep the worksheet's retest field unresolved until a reviewed revision can be inspected.
For the image heading, propose a solid #333333 panel behind the complete text area. White text on that panel calculates to approximately 12.63:1. Inspect the panel's coverage at each supported crop before accepting the proposal.
Avoid recording “darkened background” as complete evidence. Name the resulting pair. If a translucent scrim remains in the design, establish its resulting background across the relevant image regions.
The content owner can identify the failed statement and propose an outcome. The implementation owner chooses a change within the site's design process. Keep those responsibilities visible instead of treating the worksheet as permission to edit production.
Recheck States That Share a Token
A shared token is a useful inventory clue. It does not establish that every use has the same background, size, or opacity.
List where the caption token appears. Inspect a caption inside a white card, a tinted callout, and an expanded panel if those conditions exist. Give each distinct pair its own row.
Include hover, focus, selected, error, and disabled states where they contain relevant text. Record whether each state exists. Do not invent a state to make the checklist look complete.
If an inactive control falls under the criterion's exception, keep that classification attached to the specific state. Its enabled state needs its own review. A label that looks disabled may still describe an available action.
Follow the information into related content. A comparison table may reuse the muted label style while changing the cell background. The comparison table review covers the relationships that make those cells meaningful.
Likewise, a readable help sentence still needs to explain the correct field. Use the lead form review for labels, instructions, and error relationships. Contrast settles one property of the text.
Keep the Evidence Usable for the Next Reviewer
Store the exact text with its location. A record named “gray copy” gives the next reviewer several possible targets. A record naming the diagram caption identifies the statement to inspect.
Record how the pair was obtained. For a flat panel, keep the computed foreground, computed background, and relevant opacity values. For an image, identify the crop and the background area behind the text. Preserve the original asset reference where the review process permits it.
Keep a short uncertainty note when the evidence is incomplete. “Background varies beyond the captured frame” tells the implementation owner what still needs investigation. A guessed ratio would hide that work.
Choose a retest owner before closing the initial review. Give that person the evidence identifiers and the proposed version to inspect. A change request without a return check leaves the original reading task unresolved.
Close the Record at the Same Boundary
Retest the original failed sample after the proposed change reaches a reviewable version. Reconstruct the page state. Record the new pair, type size, threshold, and calculation evidence.
Compare the surrounding content too. A revised panel may cover a caption or move text beyond its intended area. Record any new observation separately rather than stretching the original contrast row.
For the fictional walkthrough, the revised colors remain proposals. A truthful close reads: four text samples examined, three samples needing revision, and all proposed changes awaiting rendered retest.
Do not turn that narrow result into a website certificate. Non-text contrast concerns other visual information. Color-only meaning concerns whether the palette carries information without another cue. Neither question disappears because a paragraph passes its text ratio.
Hand the record to the responsible owner with the failed pairs and the unresolved conditions. Preserve the earlier evidence so the later result can be traced.
Check the words against the surface behind them.
Sources
- W3C WCAG 2.2: Contrast (Minimum). Criterion, definitions, and exceptions. Checked 2026.10.28.
- W3C Understanding SC 1.4.3. Threshold precision and foreground/background evaluation guidance. Checked 2026.10.28.