- The WebAIM Million found missing alt text on 53.1 percent of home pages in February 2026, but it is not the most common failure. Low contrast text is, at 83.9 percent.
- WCAG Success Criterion 1.1.1 has six exception classes. A decorative image is correct with an empty alt attribute, which no image model will ever produce on its own.
- An average home page in the sample carried 66.6 images, up 13.6 percent in a year, so the volume argument for automation is real.
- Sort by role before you generate anything. Product shots and functional icons split cleanly, and the split decides which images a person still has to write.
- If you sell online to consumers in the European Union, e-commerce services have been in scope of the European Accessibility Act since 28 June 2025, with an exemption for microenterprises providing services.
Missing alternative text sits on 53.1 percent of the top million home pages, measured in February 2026. That number gets quoted as proof that alt text is the web's biggest accessibility problem. It is not. In the same WebAIM Million analysis, low contrast text appeared on 83.9 percent of pages, and it has been the top failure every year the study has run. If your first accessibility project is an alt text sweep, you have picked the second problem.
Pick it anyway if you sell things, because the second problem is the one your catalogue creates at scale. The sample averaged 66.6 images per home page, up 13.6 percent in a single year, and a shop with a real product range is far above that average. This is the shape of work where generating text automatically looks obvious and where doing it badly is worse than doing nothing.
What does WCAG actually require?
Less than people assume, and more precisely. Success Criterion 1.1.1 asks that non text content have a text alternative serving an equivalent purpose. The operative words are equivalent purpose, not accurate description.
The criterion then lists the situations where a full description is not what is wanted. Reading the W3C explanation of Non-text Content, there are six of them, and each one changes the answer:
- Controls and input. If the image is a control, it needs a name describing its purpose, not a description of the picture.
- Time based media. Audio and video need descriptive identification, with the fuller requirements living in Guideline 1.2.
- Tests. If a text alternative would invalidate the exercise, only descriptive identification is required.
- Sensory. Content meant to create a specific sensory experience gets descriptive identification rather than a transcript of the feeling.
- CAPTCHA. Identify and describe the purpose, and offer an alternative form using a different sensory mode.
- Decoration and formatting. Pure decoration must be implemented so assistive technology can ignore it.
That last one is the reason a blanket generation pass fails. A vision model asked to describe a decorative background gradient will describe it, fluently and at length. The correct answer is alt="", an empty string, which tells a screen reader to skip the element entirely. No model produces that unprompted, because producing nothing looks like failure.
An empty alt attribute and a missing alt attribute are not the same thing. alt="" says this image carries no information. No attribute at all leaves the screen reader to guess, and many will read the file name aloud.
Which images can a model write, and which cannot it?
Split the catalogue by role first. Once an image has a role, the right alt text is often a short rule rather than a judgement, and the roles that need judgement are a small minority.
The W3C publishes an alt decision tree that walks the same split as a sequence of questions, starting with whether the image contains text and whether that text also appears nearby. It is worth running once by hand on ten images from your own store, because the answers are rarely the ones people expect.
| Image role | Typical example in a shop | What a model can produce | What a person still decides |
|---|---|---|---|
| Informative | The main product photograph | A usable draft describing colour, material and shape | Which attribute matters to the buying decision |
| Functional | A cart icon, a social link, a logo linking home | A description of the picture, which is the wrong answer | The destination or action, written as the name |
| Decorative | Section dividers, background texture, spacer | A confident description of nothing | That the correct alt is empty |
| Image of text | A promo banner reading twenty percent off | An accurate transcription of the words | Whether the text should be real HTML instead |
| Complex | Size chart, care symbols, comparison graph | A caption of what the chart looks like | The finding, plus a full alternative nearby |
| Duplicate | The same photo used beside its own caption | A second description of the same thing | That repeating the caption is noise |
Two of those rows are where generated alt text goes wrong in a way nobody notices in review. Functional images get described instead of named, so a screen reader user hears a paragraph about a shopping trolley illustration rather than the word Cart. And complex images get captioned rather than explained: a model writing about a size chart tends to say that it is a table with measurements, which is exactly the information a sighted user already has and none of the information they need.
Why is a wrong description worse than nothing?
Because a missing alt attribute is a known gap and a wrong one is a confident lie. A screen reader user who hits an unlabelled image knows something is missing and can go looking. One who hears a plausible description that does not match the product has been told a fact, and they will act on it.
The Bureau of Internet Accessibility made this point about image recognition tools before generated alt text became a routine feature: describing an image accurately is not the same as serving its purpose on the page. Their example is an image whose visual content and page function are simply different things.
The WebAIM data quantifies how common that already is without any AI involved. Of images that did have alternative text, 10.8 percent carried questionable or repetitive text, meaning things like alt="image", a file name, or a copy of the caption sitting next to it. Roughly one in four images on popular home pages has alt text that is missing, questionable or repetitive. Generation done badly moves those images out of the missing column and into the questionable column, which scores better on an automated scan and is not better for anybody.
A triage that fits an afternoon
The point of triage is to stop treating eight thousand images as one job. Most of them are not a job at all.
- Find the decorative ones and empty them. Anything inside a styling wrapper, any divider, any spacer, any icon that sits beside its own text label. This is usually the largest bucket and it needs no model. Set the alt to an empty string.
- Name the functional ones from the destination. If the image is inside a link or a button, take the name from where it goes. Your template already knows this, so it is a code change rather than a content project.
- Generate for informative product shots only. This is the bucket where a model earns its place, and it is also the bucket where you have structured data to feed it. The product title, colour, material and category should go into the prompt.
- Route complex images to a person. Size charts, care instructions, ingredient panels, comparison graphs. There are rarely more than a few dozen, and each needs a real alternative rather than a sentence.
- Sample and check. Pull thirty generated descriptions at random and read them out loud. Anything you would not say to a customer on the phone gets rewritten.
The generation step improves enormously when the model is given the product record rather than only the pixels. A photograph of a navy wool coat can be described from the image alone, but the description that helps a buyer says which of the three navy coats in your catalogue this is. Our piece on what a generated product image can honestly do covers the same asymmetry from the other direction, and the reason is identical: the model sees the picture and not the catalogue.
The prompt that catches the two common errors
Generation quality is mostly decided by what you put in the prompt, and the two failures worth engineering against are predictable enough to write a rule for. The first is describing what is visible instead of what matters. The second is captioning an informative graphic instead of stating what it shows.
A prompt that works has four parts. It states the role of the image, so the model knows a product shot is not a chart. It supplies the product record, so colour and material come from your data rather than from a guess about the pixels. It sets a hard length ceiling, because without one the output drifts toward a paragraph. And it forbids the phrases that waste the listener's first second, which are the ones beginning with image of, photo of, or picture showing, since a screen reader has already announced that this is an image.
Then read the output as audio, not as text on a screen. A description that scans fine in a spreadsheet column often turns out to be unreadable aloud, because it stacks adjectives in an order nobody speaks in. This single check finds more bad descriptions than any automated scoring, and it takes a minute for thirty samples.
How do you know the pass actually worked?
Not from the scanner. An automated tool can see whether an alt attribute exists and whether it is obviously junk like a file name. It cannot tell whether the sentence describes the right product, and that is the failure mode generation introduces.
Three checks are worth building into the routine. Count the images your scanner reports as having alt text and compare it to the count of images you deliberately marked decorative, because those two numbers together should account for every image on the page. Spot check twenty product pages against the actual products, looking for a description that belongs to a different colourway. And listen to one full product page with a screen reader turned on, start to finish, which takes about three minutes and reveals ordering problems that no per image check can.
The number to watch over time is not a percentage of images with alt text. It is how often a description sends someone to the wrong product, because that is the one that costs a return.
What does the law require of a shop?
It depends where your customers are, and for European consumers the answer changed recently. The European Accessibility Act treats e-commerce services as in scope, and Bird & Bird's guide notes that the obligations apply to any provider selling to consumers in the EU from 28 June 2025, whether or not the provider is established there.
The scope is wider than a checkout page. It reaches the steps leading to a transaction, so product pages, search and the account flow are all inside it. The requirements point at the same perceivable, operable, understandable and robust language that WCAG uses, and there is an additional duty to publish an accessibility statement about the service.
There is a real exemption, and small sellers should know it exists. Article 4(5) of the directive states that microenterprises providing services are exempt from the service accessibility requirements and from the obligations attached to them. A microenterprise is under ten employees with turnover or balance sheet total at or below two million euro. Note the shape of it: the exemption covers services, and manufacturers of covered products do not get it regardless of size.
An exemption is not a reason to skip the work. It is a reason not to panic about a statement and a conformity procedure while you fix the images.
One more scope detail catches sellers by surprise. The accessibility statement the directive asks for is a public document about the service, not an internal note, and it has to describe how the service meets the requirements rather than assert that it does. Writing it forces the inventory work anyway, which is a reasonable argument for doing it early even where the exemption applies.
Questions that come up
How long should alt text be?
Short enough to be heard in one breath, which in practice is under about fifteen words for a product photo. Length is not a WCAG requirement, but a screen reader reads the whole string before moving on, and a paragraph in an alt attribute is a paragraph the user cannot skim, skip or re-read easily.
Does alt text help SEO?
It gives a crawler text where there was only an image, which is a genuine benefit, and it is the wrong reason to write it. Alt text written for a search engine reads like keywords and sounds like keywords out loud. Write it for the person listening and the crawler gets a good sentence for free.
Can I just install an accessibility overlay?
An overlay cannot know why an image is on your page, so it faces exactly the problem described above with less context than you have. It also cannot decide that an image is decorative. The parts of accessibility a script can fix are the parts you can fix properly in about the same time.
Do I need alt text on images in emails?
Yes, and it does a second job there. Many mail clients block images by default, so the alt string is what a subscriber reads when the picture never loads. A product email whose images are all decorative from the screen reader's point of view is a blank email for a significant slice of the list, which is a deliverability and revenue problem long before it is a compliance one.
What about images inside product descriptions?
Those are the ones nobody audits, because they arrive with the copy rather than through the image pipeline. If your description editor lets a supplier paste an image, it should require an alt value or mark it decorative at the moment of pasting. Retro fitting them later is the expensive version of the same task.
The part worth keeping
Automating alt text is a good idea applied to about a third of the images it usually gets applied to. The value is not in the model writing better sentences than you would. It is in the sorting step that happens before the model runs, which is cheap, mechanical, and the thing every overlay vendor skips.
If your store's images live in a system that knows the product behind each one, that sorting is close to free, and the difference in output quality is large. Our notes on what AI product photography holds up under and on what vision actually adds to a language model both land on the same practical point: the model performs in proportion to the structured context you hand it. If you are choosing the platform underneath, an AI ecommerce store builder that generates code you own at least lets you see and change the template that emits the alt attribute, which is where this problem is really solved.