Accessibility audit of Australian small-business websites

Status: not started. This study has not been run, no sites have been sampled, no tests have been performed and no result exists. What follows is the method, published in advance so that it cannot be adjusted afterwards to suit whatever comes out of it.

The accessibility question: website accessibility and WCAG conformance in Australian small business

How accessible are Australian small-business websites in practice, measured against WCAG 2.2, and where do the failures concentrate?

Two things about the phrasing matter. “In practice” means as the sites are actually built and actually maintained, not as a supplier described them at handover. “Where do the failures concentrate” is the part with any use in it. A headline proportion is a talking point. A ranked list of which success criteria fail most often is something a business owner or a developer can act on in an afternoon.

Why the Disability Discrimination Act (DDA) and web accessibility matter for Australians

The Disability Discrimination Act 1992 (Cth) makes disability discrimination unlawful in defined areas of public life, and section 24 covers the provision of goods, services and facilities. A website through which a business takes bookings, enquiries or sales is generally understood to sit inside that.

The Act contains no turnover exemption. This is the point most often assumed the other way around, because the Privacy Act 1988 (Cth) does currently exempt many small businesses by reference to an annual turnover threshold. There is no equivalent in the Disability Discrimination Act. Its obligations do not switch on at a revenue figure, which means the obligation and the population being studied here overlap completely. The legal background is set out at greater length on website accessibility in Australia, which is general information rather than legal advice.

What does not exist, as far as this site has been able to establish, is a published, method-stated measurement of how Australian small-business sites actually perform. The obligation is discussed constantly and the compliance rate is asserted constantly. The assertions do not come with samples.

Sampling frame and selection: businesses across Australia, outside government and enterprise

The population of interest is Australian businesses with a public website, trading in a consumer or small-commercial market. It excludes the enterprise and government segments.

No complete list of that population exists. The frame will be an approximation and will be named as one. The current design draws from a business listings dataset filtered to Australian locations, across a fixed set of industry categories and a spread of states and territories. The website field must be populated. Selection within the frame is randomised with a recorded seed, so the draw can be reproduced later.

Known biases in that frame will be stated on the finished page, not discovered by a reader. Businesses that maintain a listing are more likely to maintain a website. Some categories are far better represented than others. Businesses without a website are invisible to the study entirely. This means the finding describes sites that exist, not businesses in general.

The frame is not a random sample of Australian small business. It is a convenience sample with randomisation applied inside it, and it will carry that description.

Sample size: how many businesses and organisations to sample

The study is designed around a band of 200 to 400 sites. That is a design parameter, not a result. No site has been drawn.

The band is chosen so that per-criterion failure rates can be reported with a denominator that supports a whole-number percentage. It is also small enough that a manual component is actually deliverable, not just aspirational. If the achieved sample lands outside the band, the achieved number is what gets published, along with the reason why.

What would be tested against WCAG Level AA, per the Australian Human Rights Commission’s guidelines

Two layers, reported separately and never merged into a single score.

Automated layer. Run on the home page and two internal pages per site, using an open-source rules engine whose version and ruleset are recorded. It covers criteria a machine can genuinely decide: missing alternative text, form fields without programmatically associated labels, contrast below the threshold in 1.4.3 Contrast (Minimum), missing document language, and structural failures under 4.1.2 Name, Role, Value.

Manual layer. Run on a stratified subset, covering criteria that require judgement: 2.1.1 Keyboard and 2.1.2 No Keyboard Trap, 2.4.7 Focus Visible, 2.4.11 Focus Not Obscured (Minimum), 2.5.8 Target Size (Minimum), 3.3.2 Labels or Instructions, and the quality of alternative text against 1.1.1 Non-text Content rather than merely its presence. Level AA is the reference, consistent with the Australian Human Rights Commission’s April 2025 guidelines.

The task set for the manual layer is fixed in advance: reach the enquiry or booking form by keyboard alone and attempt to complete it. Most of what a small-business site exists to do runs through that one flow.

The limitation, stated before anything else is read

Automated tooling detects only a portion of the WCAG success criteria, commonly estimated at somewhere between about a third and a half. The remainder require human judgement. A machine can confirm that an image has alternative text, but not whether the text describes the image.

The consequence is not a caveat. It is the meaning of the number. Any automated figure produced by this study measures a floor, not conformance. A site recorded with no automated failures has not been shown to conform to anything. It has been shown to pass the subset a machine can check. The finished page will state this beside every automated figure, rather than once at the bottom. No headline will be written in the form of a conformance rate.

Who does the manual testing

The manual layer would be performed by the author, who is a practising web developer and not a certified accessibility specialist. That is a limitation on the finding and will be recorded as one. The testing protocol will be released, so that a reader can see exactly what was and was not done.

The design allows for a second reviewer with lived experience of assistive technology to test a subset independently, with agreement between reviewers reported. Whether that reviewer is engaged is not settled. If it does not happen, the page will say so directly, rather than omitting the sentence.

How results would be reported

Aggregate only. No business is named, no website is identified, and no screenshot of an identifiable site appears. The purpose is to describe a market, not to expose individual operators. In most cases they were sold a site by somebody else and have no idea what is wrong with it. Shaming a small business for a developer’s default template would be both unfair and pointless.

Findings would be reported as per-criterion failure counts with denominators, split by automated and manual layer, with the strata shown. No composite accessibility score is produced, because a composite hides which criterion failed and that is the only part with practical value.

What would be published

The full method as run, including any departure from this page and the reason for it. The instrument itself: the test script, the rules engine and version, the manual protocol, the scoring sheet. The randomisation seed and the frame filters. Per-criterion aggregate results, including criteria where nothing failed. The dates of the collection window.

Site-level raw results are not publishable, because a list of failures against a fixed sampling frame can identify individual businesses. That constraint will be stated with the release rather than left as a silent omission.

What would falsify or complicate the finding

Picture a high automated failure rate driven overwhelmingly by one platform’s default template. That would make the finding a statement about a platform, not about Australian small business — so the platform breakdown would need to lead the report.

Substantial disagreement between the automated and manual layers on the same sites would undermine the automated layer’s usefulness and would be published as its own result.

Sites that fail on a home page but pass on the flow that matters, or the reverse, would complicate any single figure. They would be reported as two figures.

And if a large share of sampled sites turn out to be defunct, parked or redirected, the achieved sample would differ from the drawn sample. That changes what the denominator means. That count gets published either way.

Conflict of interest

This study has not been run. There is no result to report.