Analysis · 17 November 2026
We counted 2,649 images on 48 store homepages. Most of the “missing” alt text wasn’t missing.
The standard finding in an accessibility audit is “images without alt text”. When you actually look at the markup, that phrase hides two very different problems — and the bigger one is not the one people expect.
What we did
On 17 November 2026 we fetched the homepages of 48 online stores that serve their markup to a plain HTTP request, parsed every <img> element, and counted. We excluded tracking pixels, inline SVG icons and images with no source. One page per site, no JavaScript rendering, no judgement calls about whether a picture is decorative — the markup says what it says.
The sample is 48 stores, not 48,000, and it is one page per store. Treat the percentages as a description of this sample, not of the web. The full table, including the 34 sites that refused our requests, is in the measurement.
The numbers
- 2,649 images across 48 homepages.
- 970 had no text alternative — 36.6% of all images. 88% of the homepages had at least one.
- Of those 970: 129 had no alt attribute at all. 841 had alt="".
- 225 images shared their alt text with at least two other images on the same page.
- 98 alt strings were longer than 125 characters; 69 repeated a word three or more times; 39 were “Image”, “Image unavailable” or “Banner title icon”; 22 were filenames.
- 1,317 images had no flag at all — someone had written something specific. That is the part the audits never mention.
Finding one: empty is not absent
Eighty-seven per cent of the images without a text alternative had an alt attribute that existed and was empty. That is not the classic story of a developer forgetting the attribute. It is the story of a value being set to “nothing” — by a component that takes a prop it never receives, by a CMS field left blank that still renders the attribute, by a lazy-loading template that writes alt="" as a placeholder, or by someone who read “decorative images get empty alt” and applied it to a collection tile.
This matters because the two cases fail differently. A missing attribute is visible in a code review and gets read aloud as the filename. An empty attribute is silent: validators that only check for the presence of the attribute pass it, and screen readers skip the image entirely. The user hears nothing where the page put a product.
An empty alt is correct for exactly one kind of image: one that adds no information — a divider, a background texture, an icon that sits next to a text label. If the image is the only place the page says “linen shirt, sand”, the empty alt deletes that sentence for anyone not looking at the screen.
Finding two: the template writes the sentence
225 images had duplicate alt text within a single page. The repeats are revealing:
- “Everyday Seamless Leggings” — six times, on six different colours.
- “Everlane pic description” — the literal name of a CMS field, shipped to production.
- “Visual Category Module Image” — a component name doing duty as a description.
- “Mercer-Olive-Corner-Sectional-web.jpg” — the filename of a sofa, read aloud as a sentence.
None of these are careless in the way audits imply. They are what happens when alt text is generated by the template from the product title, or when a placeholder is never replaced because nobody is looking at the page with a screen reader. The page passes a “has alt text” check and still tells six different images the same thing.
The part that is not broken
Half the images in the sample — 1,317 — had alt text with nothing wrong with it: specific, short, describing the image rather than the product title. Somebody wrote those, one at a time. The gap is not ignorance; it is volume. The stores that got it right had one thing in common: someone had been through the catalogue with the images in front of them.
How we write them, and what to be sceptical about
Our tool sends each image to a vision model with a fixed set of rules we wrote: one sentence, under 125 characters, name the subject and its distinguishing details, never invent a product name or a price, do not append sales language like “showcasing”. The model classifies each image first — informative, decorative, logo, chart, screenshot, text-in-image — because “is this decorative?” is the decision that causes the empty-alt failures above.
Disclosure, since it is relevant to how much you should trust this post: we sell alt text remediation, so the measurement is vendor data, and the writing is machine-generated and reviewed by rules rather than by a room of editors. What we can offer in exchange is the method and the raw table, both public, plus a free audit that shows the output on your own images before you pay anything. The rules are also the reason we mark existing alt text as “keep” rather than rewriting it: 1,317 of those images did not need us.
If you take one thing away
When you next look at an audit that says “missing alt text on 400 images”, ask how many of them are empty strings rather than absent attributes, and how many share a sentence with five neighbours. Those are the rows that pass the checklist and still leave a screen reader user with a blank where the product should be — and they are the cheapest to fix, because the sentence does not exist yet and nothing has to be untangled.
Data: the alt text gap, 2026 · Method questions: hello@alttextfix.com · Written by AltTextFix (Pixelforge LLC), which sells alt text remediation. Corrections welcome and published.