This is a thought piece, not a definitive result. It presents a hypothesis, tests it honestly against real data, reports an inconvenient methodological complication we found along the way, and ends with an open question rather than a clean answer.
The claim we set out to test
“Accessible design is good design” is one of the most repeated sentences in the accessibility field. It’s usually offered as encouragement: a way of telling teams that fixing accessibility isn’t a tax on the product, it’s an investment in the product, because the same changes that help a screen reader user also help everyone else. It is intuitively appealing and rhetorically useful. It is also, as far as we can tell, rarely tested against a real dataset where accessibility and usability were scored independently for the same person doing the same task.
We had one. Across 15 independently run usability studies, every tester-task session in this dataset produces two separate numbers: an accessibility score and a usability score, both on a 0-100 scale, generated from the same observed session. That gave us 1,580 paired data points to actually check whether the two move together, and, if they do, how strongly, and for whom.
The first pass: a surprisingly strong result
Run a straightforward correlation across all 1,580 paired scores, and the answer looks emphatic. The Pearson correlation coefficient between accessibility score and usability score is r = 0.907, meaning accessibility scores statistically explain about 82% of the variance in usability scores (R² = 0.822). Taken at face value, that’s about as strong a validation of “accessible design is good design” as a real-world behavioral dataset is ever likely to produce. If we’d stopped there, this would be a short, triumphant piece.
We didn’t stop there, because a correlation that clean in messy behavioral data is itself a reason to look harder, not a reason to celebrate.
The complication: almost half the “pairs” aren’t really pairs
Looking at the underlying values rather than the summary statistic revealed the problem. Of the 1,580 tester-task sessions, 787 (49.8%) have an accessibility score and a usability score that are exactly, digit-for-digit identical. Not close. Identical.
This is not plausible as an emergent empirical finding: it’s a signature of how the scoring pipeline works. In almost half of all sessions, the scoring methodology is not independently assessing accessibility and usability at all; it is calculating one number and writing it into both fields, most likely because the underlying rubric for that session didn’t ask a differentiated question, or because a single combined signal (e.g., “could the tester complete the task”) was used as a proxy for both constructs when a more granular breakdown wasn’t available. This pattern appears across the large majority of the 21 distinct scoring frameworks represented in the data and across all 13 projects in this subset: it isn’t a single project’s quirk.
That matters enormously for the correlation calculation. A statistic like Pearson’s r treats every one of those 787 duplicated pairs as evidence that accessibility and usability move together, because by construction, they moved together perfectly. Removing them isn’t cherry-picking; it’s the only way to ask what the relationship actually looks like among the sessions where the two constructs were, in fact, independently evaluated.
The second pass: what’s left when you remove the duplicates
Restricting the analysis to the 793 sessions where accessibility and usability received genuinely different scores, the correlation collapses to r = 0.249 (R² = 0.062). That is a real, statistically meaningful, positive relationship, but it explains roughly 6% of the variance, not 82%. It is the difference between “these are functionally the same thing” and “these are related, but mostly independent.”
Breaking the differentiated subset down by cohort tells a more specific story than a single overall number can:
| Cohort | n (differentiated sessions) | r (accessibility vs. usability) | Mean gap (usability − accessibility) |
|---|---|---|---|
| Physical disability | 34 | 0.15 | −9.1 |
| Indigenous | 34 | 0.20 | −12.0 |
| Neurodivergent | 119 | 0.24 | +4.6 |
| Blind | 102 | 0.23 | +4.1 |
| Deaf | 120 | 0.29 | +5.4 |
| Low vision | 135 | 0.33 | +0.7 |
| Limited English proficiency | 154 | 0.44 | +10.5 |
| Over-65 | 95 | 0.45 | −4.7 |
Two things stand out. First, every single cohort shows a weaker correlation than the naive full-sample figure once duplicated scores are removed: the strongest cohort-level relationship (over-65, at r = 0.45) is still less than half the strength of the misleading r = 0.907 headline number. Second, the direction of the gap between the two scores is not random: physical disability and Indigenous testers show usability scores meaningfully lower than their accessibility scores (by 9 and 12 points respectively), while limited-English-proficiency and deaf testers show the opposite pattern, rating usability higher than accessibility.
What this might mean: three competing hypotheses
We think there are at least three plausible, non-exclusive explanations for the negative-gap cohorts (physical disability, Indigenous, over-65), and we want to be upfront that we can’t fully adjudicate between them with this dataset alone.
Hypothesis A: accessibility conformance doesn’t address these cohorts’ actual barriers. This is consistent with a separate finding from this same testing program: physical disability findings are disproportionately about interaction effort and interface responsiveness (not screen-reader operability), Indigenous findings are disproportionately about content findability and trust, and over-65 findings are disproportionately about hesitancy and reassurance. None of those barriers are well captured by conformance-style accessibility scoring, which historically weights assistive-technology compatibility heavily. If that’s right, a product can score reasonably on “accessibility” as conventionally measured while still being genuinely hard to use for these specific groups, which is exactly the -9 to -12 point pattern we see.
Hypothesis B: the accessibility score itself is more lenient for these cohorts by construction. It’s possible the scoring rubric weights certain accessibility sub-criteria (e.g., basic screen-reader compatibility) that these three cohorts don’t heavily rely on, inflating their accessibility score independent of their actual experience, while the usability score captures friction the accessibility rubric was never designed to detect. This is a measurement-instrument explanation rather than a real-world one, and we can’t fully rule it out without auditing the specific scoring weights applied to each session.
Hypothesis C: small samples are doing some of the work. Physical disability (n=34) and Indigenous (n=34) are the two smallest differentiated-subset cohorts in this table. We report the finding because the direction is consistent with independent evidence elsewhere in this program (the friction-type analysis referenced above), which makes us more confident it’s a real pattern and not noise, but 34 sessions is not a large sample, and this finding should be treated as a hypothesis worth testing further, not a settled result.
Where this leaves the original claim
“Accessible design is good design” survives this analysis, but in a much more modest form than the aphorism usually implies. There is a real, positive, statistically detectable relationship between accessibility and usability scores once you look only at sessions where the two were independently assessed (r = 0.249): the claim isn’t false. But it is far weaker than a surface-level read of this same dataset would suggest (r = 0.907), and the strength of the relationship is not uniform across the population the maxim is usually invoked to reassure people about. For at least three cohorts in this dataset, meeting a reasonable accessibility bar did not translate into a correspondingly good usability outcome.
The more actionable version of the claim, based on what we found, might be: accessible design tends toward usable design, on average, but it is not a guarantee, and the gap between the two is largest for exactly the cohorts whose primary barriers aren’t well captured by conventional accessibility criteria in the first place.
A methodological note we think matters beyond this dataset
The duplicate-score finding is arguably more important than the correlation result itself, and we suspect it isn’t unique to this dataset. Any usability or accessibility program that reports both an “accessibility score” and a “usability score” should check whether those two numbers are ever actually generated independently, or whether one is silently derived from (or copied from) the other in a meaningful share of cases. If they’re not independent, correlating them tells you about the measurement instrument, not about the product. We only caught this because we had access to the raw paired values rather than pre-aggregated averages; a dashboard showing only mean scores per cohort would never have revealed it.
Open questions for further research
We’d genuinely like to see this tested against other testing programs’ raw data, not just aggregated dashboards. Three questions we don’t think this dataset alone can answer: does the negative usability-minus-accessibility gap for physical disability, Indigenous, and over-65 cohorts replicate at larger sample sizes? Is it consistent across different scoring frameworks, or an artifact of a subset of them? And most importantly, if accessibility conformance genuinely under-serves these cohorts’ real barriers, what would a scoring rubric look like that captured trust, findability, and interaction effort as first-class accessibility dimensions rather than leaving them to a separate, loosely correlated “usability” score?
About this analysis
Figures in this piece are drawn from an anonymised aggregation of scoring data across 15 independent usability testing projects conducted by See Me Please between late 2025 and mid-2026. All correlation and gap statistics are computed on tester-task-level scoring records; cohort sample sizes are reported throughout, and readers should weight the smaller cohort figures (n<40) accordingly. All client and participant identities have been removed.
