Here’s a common accessibility scenario for a team working on an e-commerce site:
The team runs an automated accessibility scan on the site.
The results show a 100% score and zero detected issues.
The team is thrilled and moves on to the next thing on their priority list.
Two days after that, a lawsuit arrives alleging the site is inaccessible because someone can’t make a purchase.
The fact is, an automated scan only reports on a portion of the issues. As soon as a user with a screen reader tries to navigate, issues emerge that can block them from accomplishing their goal. The same can be true for other users who have disabilities.
This has always been the problem because too many teams and organizations think a perfect scan is all they need. Today, with AI, it’s also easier and faster to reach that and be convinced that everything’s fine.
In this article, I’m going to cover how to think about the numbers and how AI factors into it, including:
The difference between compliance, conformance, and accessibility.
Why you can’t treat all issues the same.
How AI can help you, and how you can help AI help you.
Why you must have a sustainable accessibility program.
A score from an automated scan measures partial conformance to the WCAG, not actual accessibility or compliance. These three things are not the same.
Compliance means meeting applicable legislation, such as the European Accessibility Act or the Americans with Disabilities Act Title III. Compliance often requires conformance.
Conformance is adherence to a technical standard, in this case the Web Content Accessibility Guidelines (WCAG).
Real accessibility focuses on two things:
Can a person actually complete what they came to do?
Can you show that you are managing accessibility as an ongoing effort?
A scan can tell you one thing: You passed the checks a machine can run, which covers a portion of conformance.
A scan cannot tell you:
That a shopper who uses a screen reader added the right size to the cart.
That someone who navigates using their keyboard made it through checkout.
For a firsthand example, read Michael Taylor’s How I shop an unfamiliar website with a screen reader.
A scan also will not stop a demand letter or a lawsuit from arriving. UsableNet’s ADA Accessibility Lawsuit Tracker shows the latest monthly filing trends.
Before going further, it’s important to say this isn’t legal advice, but this information is based on more than 25 years of UsableNet experience in accessibility.
Further, where conformance is concerned, it’s not often talked about that the WCAG standard was written as a technical guideline, not a legal standard.
WCAG treats every violation the same way: a failed success criterion is a failed success criterion.
Reality doesn’t work like that. A decorative image missing an alt attribute and a keyboard trap that blocks someone from reaching your navigation are both “failures,” but they are nowhere near the same problem.
This is where the definitions of conformance, compliance, and accessibility matter.
If “accessible” means “zero automated issues” to your team, that definition is far too small. With this, you’re neither compliant nor accessible.
Zero automated issues can absolutely be a meaningful milestone to achieve. It might even reduce legal risk. But you must take it as one step, not the whole scope.
WCAG covers far more than automation can report. You need a manual audit to fully understand where you stand with all WCAG 2.2 AA success criteria and the actions you need to take to make your site accessible. Pretending otherwise is how teams talk themselves into a false sense of conformance.
Even if you are fully WCAG conformant, you shouldn’t state you’re accessible until you know two things:
A person with a disability can complete what they came to do.
You’ve got a process in place to show you’ve got a sustainable accessibility program.
If the automated scan number is misleading, why do so many e-commerce teams chase it? Because AI has made it easy to improve the score.
AI-assisted coding with prompts like “just make the accessibility errors go away” can create a clean report while the underlying experience gets worse.
I wrote about one version of this in a recent UsableNet article, “AI and Accessibility for Ecommerce,” so I won’t retell the whole story here. The short version: a team can push an automated score to 100% by suppressing or hiding content from assistive technology. The report looks perfect. The page is less accessible than it was before.
The point isn’t that AI is the villain. It’s that zero detected issues in a scan has never meant “accessible,” and AI just lets you reach that number fast and in ways that may not be valid.
The work isn’t chasing the easy score. It’s deciding to find and fix what matters most first, and building the program to make that consistently possible.
The most useful thing a team can do with an audit is stop treating it as a flat checklist of issues and start prioritizing by impact.
Start with severity: what does a given issue do to a person trying to use the site? As I outlined in my earlier example, a keyboard trap in your main navigation is not the same as a decorative image missing its alt attribute. Ideally, the audits you’re using indicate the severity of an issue to help you with your prioritization.
Severity is the parameter that tells you how likely a user is to complete a task.
High severity issues are blockers to accomplishing a task. These are issues you’d want to treat as urgent to fix.
Medium severity issues are likely not blockers but could still cause trouble for customers. You’d want to tackle these to provide a smoother experience for users, especially in your primary flows.
Low severity issues aren’t blockers, but they also add friction. They could add a lot of friction, especially if they cluster on pages or in flows.
Beyond severity, there are other things that can re-order your remediation list.
Legal exposure: Is the issue something that’s showing up in demand letters and claims in your industry? Does it sit in a core flow like checkout, login, sign-up, or search? This is an area where something like a missing alt attribute might get fixed sooner than later. It’s definitely something a plaintiff would lock on to in order to make a claim even if the reality is that it’s not blocking anything.
Product context: The core journeys of your site deserve more urgency than a rarely visited blog post.
Detection method: Some issues are easy to catch at scale, and others only surface when a real person tries to use the page or an expert reviews the code. This is where getting automation down to zero issues is useful since you’re remediating the easy-to-find issues.
Momentum is also a key to success. Don’t let your first move be a massive remediation effort that stalls for a quarter.
Perceivable progress matters, and internal teams deserve to see it and celebrate it, and your customers deserve to have those fixes available. A good severity-and-priority framework is what turns an overwhelming audit into an actionable roadmap.
It’s the difference between a proactive maturing program and reactive checkbox thinking.
Here’s the encouraging part. Prioritization at scale is exactly the kind of work AI is good at. It can:
Analyze a whole site.
Cluster issues back to the shared component or template behind them.
Compare one release against the last.
Surface the most recurrent problems, so you know where to look first.
But there’s a condition. AI is only reliable inside a perimeter.
My colleague Michele Lucchini, UsableNet’s Vice President of Product, describes it as giving the model a solid guardrail: a vetted body of knowledge that is the only thing it’s allowed to reason from.
Inside that perimeter, AI can be reliable. Outside of it, a general-purpose model will give you a confident, credible-sounding answer that is simply wrong. If you don’t have the expertise to catch it, that wrong answer ships for your customers to struggle with.
We’ve seen this in our own testing.
We took a simple newsletter sign-up form—one email field with a placeholder and a submit button—that was already accessible, and asked three different setups to check its conformance.
A general AI assistant flagged failures and told us to add a label the form didn’t actually need.
A second, more cautious AI coding tool did something similar and asked us to go fix it rather than fixing it itself.
Only the model constrained to our vetted body of knowledge, provided by UsableNet’s AQA, got it right: the form already passed, and no work was required.
That example flips the assumption many people start with. We expected AI’s risk to be missing real issues. However, just as often, an unconstrained model invents issues that aren’t there and sends you off to “fix” things that were already fine.
In prioritization terms, that is noise that eats the exact time you were trying to save. This is why context has to come from a trusted source so the AI has that to use.
A good expert audit already carries severity ratings and page context, so AI-assisted triage builds on vetted human judgment instead of guessing at it. It’s the same body-of-knowledge idea, applied to deciding what to fix first.
It’s also what we’re working toward with AQA, our developer-led, AI-assisted platform. Years before we added AI, we were describing the test methodology for every WCAG success criterion in detail. To give you a sense of that depth, our platform has more than 70 separate tests just to validate the single WCAG criterion 1.1.1 covering alternative text for non-text content.
Giving that kind of precise, vetted knowledge to an AI coding tool is what keeps the AI’s recommendation grounded rather than guessed.
Keep in mind though, even a solid body of knowledge isn’t a silver bullet to solve all your accessibility issues. A competent human who understands accessibility is required to work alongside the AI. The body of knowledge makes a competent human faster.
If a score isn’t the goal, what is? A defensible, ongoing program. And this is where the legal picture and the accessibility picture actually agree.
The European Accessibility Act doesn’t just name a standard with WCAG 2.2 AA. It also expects you to demonstrate an accessibility program and to treat issues as tracked, managed work.
Some US courts are starting to reflect the same logic. In Jones v. Moscot.com, LLC, decided in the Southern District of New York in 2026, a legally blind shopper who uses a screen reader sued under the ADA after hitting barriers trying to buy a pair of sunglasses. The court dismissed the claim, not because the site scored perfectly, but because the retailer could show an ongoing, documented effort.
Moscot could prove that it:
Worked with an accessibility partner since 2023.
Had taken real steps toward WCAG conformance.
Tested against the specific barriers alleged.
Fixed the one low-severity issue that remained.
Showed it intended to keep accessibility central.
That evidence, continuing program, cleared a legal bar that a one-time fix or a promise to do better would not have. (It’s important to keep in mind that this is the result of a single case out of thousands, but it is notable nonetheless.)
AI has a role here too, where it can help:
Generate reports.
Summarize audits.
Track progress release over release.
That’s exactly the kind of evidence a managed program produces. The people still own the program and the decisions.
The thread running through all of this is intellectual honesty. Be honest about:
Who you are as an organization.
What your teams can realistically sustain.
Where you actually are in your accessibility journey.
Don’t let a tidy automated scan report tell a story the reality can’t back up.
A 100% automated score is a starting or intermediate point, not a destination. The e-commerce teams that get the most out of AI aren’t the ones chasing a cleaner scan. They’re the ones who:
Know what good accessibility looks like.
Prioritize by real impact.
Keep knowledgeable humans in the loop.
Build a program they can stand behind.
Used that way, AI helps you get there faster. It just won’t get you there on its own.
If you’d like a clear picture of where your site actually stands, and a plan to prioritize what matters most, request a consultation with UsableNet.