- Detected accessibility errors per home page rose 10.1 percent in 2026, reversing six years of slow improvement.
- Six failure types account for almost all of it, and only two of them are the kind a tool can close on its own.
- A regulator ordered an overlay vendor to pay 1 million dollars over claims that its AI made sites compliant. The order bars the claim itself, not the product.
- Online shops trading into the European Union have been in scope since June 2025, with a narrow exemption for the smallest service businesses.
- Fix in revenue order. A checkout nobody can complete costs more than a thousand images with no description.
Somebody sells you a line of JavaScript. Paste it into your theme, and by Thursday your shop is accessible. It scans, it patches, it adds a little person icon in the corner with sliders for contrast and font size. The pitch is irresistible for a business with one developer and no developer, which is most of them.
In January 2025 the Federal Trade Commission ordered the marketer of one such widget to pay 1 million dollars, and the reason it matters to you is not the money. It is what the complaint said the product had failed to do.
Is the web getting more accessible, or less?
Less, for the first time in years. The 2026 edition of the annual scan of the top million home pages found detected failures on 95.9 percent of home pages, up from 94.8 percent the year before, with 56.1 errors per page against 51 in 2025. That is a 10.1 percent increase in detected errors and a reversal of six years of gradual improvement.
One number moved the other way. Missing alternative text fell to 53.1 percent of pages from 55.5 percent, which is the metric most likely to have been touched by automation, since generating a description for an image is exactly what a model does well. Meanwhile low contrast text rose to 83.9 percent of pages, with 34 separate instances on an average page, up 15 percent in a year.
Put those two movements side by side and you have the shape of the whole problem. The failure that automation can address improved slightly. The failure that comes from design decisions got worse.
Which of the six can a machine actually fix?
Two of them properly, two of them partially, two of them not at all. The distinction is whether the correct answer exists in your data or has to be decided by a person.
| Failure | Pages affected | Can automation close it? | Why |
|---|---|---|---|
| Low contrast text | 83.9% | No | A tool finds every instance. Changing them is a brand decision somebody has to make once. |
| Missing alt text | 53.1% | Mostly | A model drafts a description from the image plus your product data. A human checks the decorative ones stay empty. |
| Missing form labels | 51% | Partly | Labels can be generated, but only the person who built the form knows what the field means. |
| Empty links | 46.3% | Partly | Icon links can be described automatically. Ones whose meaning depends on the row they sit in cannot. |
| Empty buttons | 30.6% | Partly | Same problem, higher stakes: a mislabelled button in checkout is worse than none. |
| Missing page language | 13.5% | Yes | One attribute in your template. Five minutes, once. |
Read the right hand column and the widget pitch collapses on its own terms. Two of the six are a template edit and a batch job. The rest need somebody who knows what your form fields mean and what your brand colours are for.
Can a widget make a site compliant?
No, and a regulator has now said so in an order. The complaint against the vendor of an accessibility plug-in alleged it had claimed its product could make any website compliant with the accessibility guidelines, and that the tool did not achieve it. The consent order bars representing that automated products can make any website conformant without evidence to support the claim.
The detail worth carrying away is what the plug-in was said to have missed: navigation menus, form fields and image descriptions. Those are not exotic corners of a site. They are the parts a customer uses to find a product and buy it, and they map onto three of the six rows in the table above.
There was a second allegation, about third party articles and reviews formatted to look independent while the financial connection went undisclosed. It is a useful reminder when you are researching any tool in this category: some of the comparison pages you will read were bought. The rules about what a vendor may claim are the same rules that apply to your own marketing, which we set out in what you may claim about AI in your own copy.
An overlay is not automatically useless. Some of them provide genuinely helpful reading tools for some users. The order was about the compliance claim, not the existence of the product. Buy it as a convenience feature if you want one; do not file it under legal risk closed.
What is the standard you are actually being measured against?
The Web Content Accessibility Guidelines, at conformance level AA. The W3C publishes three conformance levels, A, AA and AAA, with 2.2 the current version, published in October 2023 and updated in December 2024. Content meeting 2.2 also satisfies the earlier versions, so working to the latest saves you from tracking which regulation names which.
Level AA is the practical target because it is what most laws point at. Level AAA is not intended to be met across a whole site and no reasonable regulator expects it. A shop that reaches AA on its home page, category pages, product pages and checkout has done the work that matters commercially, whatever remains on the blog archive.
Does the European rule apply to a small shop?
If you sell online to consumers in the European Union, probably yes, and it has applied since 28 June 2025. The European Accessibility Act covers e-commerce services, meaning buying and selling products or services online through a website or an app, and penalties are set nationally, ranging up to significant fines and, at the extreme, withdrawal of a non compliant service.
The exemption people reach for is narrower than the summaries suggest. Microenterprises with fewer than ten employees and under 2 million euros of turnover are exempt from some requirements when providing services. Run the test honestly. Count employees rather than full time equivalents you find convenient, use turnover rather than profit, and remember that the exemption evaporates the moment you cross either threshold, with no transition period built in for the year you grow.
Two more things about that exemption. It is written for service providers, so a business manufacturing or distributing covered products is in a different position. And it exempts you from obligations, not from customers: an exempt shop with an unusable checkout still loses the sale.
Where should a small catalogue start?
At the checkout, always, and not because it is the most broken. Because it is the only page where a failure costs you the whole order rather than a visit.
Work outward from the money. Checkout and cart first, including every form field, every error message and the payment step. The buy box on a product page second, which means the variant selector, the quantity control and the add to cart button. Category listing and search third. Then everything else, in whatever order suits you, which for most shops means the image descriptions that a batch job can draft.
That ordering matters more than it sounds, because the default advice runs the other way. Run a site wide scan and it will hand you three thousand image warnings and eleven form issues, and the three thousand look like the emergency. They are not. Eleven form issues in a checkout is a business that cannot take money from some of its customers.
What a scan sees, and what it misses
An automated scan sees the failures that live in markup. It cannot tell you whether the reading order of your product page makes sense, whether a keyboard user can reach the variant picker, whether an error message explains what to do, or whether your carousel traps somebody who tabs into it. Those are found by a person using the site the way an affected customer would.
The cheapest version of that test costs an afternoon and no money. Put your mouse in a drawer and buy something from your own shop using only the keyboard. Tab through it, and note every place where you cannot tell what is focused or cannot reach a control. Then turn on the screen reader already built into your operating system and try the product page. You will find things no tool reported, and you will find them in the first ten minutes.
This is the same lesson as the alt text problem, where the useful move was triage rather than volume. We set out how to sort a catalogue by image role in deciding which product images actually need alt text, and the principle carries: generating three thousand descriptions is easy, deciding which two hundred carry information a buyer needs is the work.
What should an accessibility statement say?
What is true, and nothing beyond it. A statement that claims full conformance you have not tested is a marketing claim of exactly the kind that got a vendor fined, made by you this time, on your own domain.
Four short paragraphs cover it. What standard you are working to, naming the version and the conformance level. What you have assessed and how, including whether it was an automated scan, a manual review or both, with the date. What you know is not yet conformant, stated plainly, because a known gap disclosed reads as competence while the same gap discovered reads as neglect. And how somebody reaches a human if they hit a barrier, with a route that does not itself require the broken part of the site.
That last point catches people. A contact form is a poor fallback for a customer who cannot use forms. Publish an email address and a phone number alongside it, and answer both.
What does fixing it actually cost?
Less than the audit industry implies and more than the widget subscription, and the split between the two is where most small shops get their budgeting wrong.
An automated scan of your key templates costs nothing beyond the time to run it, since free tools cover the markup level failures perfectly well. The keyboard and screen reader pass costs an afternoon of your own time, repeated once a quarter. Those two together find the majority of what a customer would hit. Where money starts is remediation on a theme you did not write, which is developer time measured in days rather than hours, and it scales with how unusual your theme is rather than with how many products you sell.
Two costs are easy to miss. The first is the contrast decision, which is cheap to implement and slow to approve, because it touches brand. Get that decision made early or it will block a week of otherwise finished work. The second is regression: a shop that fixes everything in March and installs three apps in April is back where it started by summer, unless a scheduled scan tells somebody. Budget for the schedule, not only for the sprint.
Set that against the upside, which is not only legal. Form labels, focus states, contrast and error messages that explain themselves are conversion work under another name. The changes that let a screen reader user complete checkout are the same changes that help somebody on a phone, in bright sunlight, in a hurry.
What if a complaint arrives?
Answer it like a customer service problem before it becomes a legal one, because at the point of first contact that is usually what it still is.
Acknowledge quickly and specifically. A person telling you their screen reader cannot get past your variant picker has handed you a bug report more precise than any scan, and treating it as a support ticket rather than a threat changes the tone of everything that follows. Offer an immediate alternative route to buy, by phone or by email, which costs you nothing and solves their actual problem today. Then fix the thing, and tell them when you have.
Keep a record of what you were told, what you changed and when. Where a matter does escalate, the difference between a business that logged a complaint and shipped a fix in three weeks and one that has no record at all is substantial, in every jurisdiction, whatever the underlying rule. It is the same evidence discipline that decides disputes about anything automated on your site, and the file is only useful if somebody started it before the letter arrived.
Where AI genuinely earns its place
In drafting, in detection and in triage, which is three real jobs rather than the one it is sold for.
Drafting: a model that can see the image and read the product record writes a serviceable description of a product photo faster than a person, and the failure mode is blandness rather than error. Detection: a scan run on a schedule catches the regression somebody introduces next month, which is the failure pattern nobody plans for. Triage: sorting ten thousand findings into the fifty that sit on a purchase path is a sorting problem, and sorting is what these tools are good at.
What none of them do is decide. Whether your brand blue on white passes at text size is a fact a tool reports; changing it is a choice about your identity that a person makes once and then applies everywhere. The tool that promises to make that choice for you is the tool that got fined.
The same thinking applies to how the page reads to a machine rather than to a person, which is a related but separate discipline. Descriptions, structure and labels that help a screen reader also help the assistants your customers now ask about products, which we looked at in how AI shoppers read a product page. If you are rebuilding rather than patching, the accessible version is much cheaper to get right at the template stage, which is one of the arguments for starting from a store built to current standards rather than retrofitting a theme bought in 2019.
The honest summary
There is no line of JavaScript that ends this. There is a checkout you can fix this month, a contrast decision you can make once, a scan you can schedule, and a batch of image descriptions a model can draft while you do something else. Nineteen in twenty home pages fail an automated check today, which means the bar to being noticeably better than your competitors is lower than the scale of the problem implies. It also means nobody gets to feel comfortable about it, including the businesses that bought the widget.