How the Health Score works
Methodology last updated 6 September 2026.
Accessibility scores are easy to inflate and easy to game. This page sets out exactly what our free Alt-Text Health Score checks, how the number is built, and what it deliberately does not claim.
What we look at
The scanner fetches your page's HTML once, exactly as it leaves your server. Where the page looks JavaScript-built, it opens the page in a full web browser, lets the scripts run, and reads the result. Either way it reads every image tag: plain, lazy-loaded, responsive and noscript variants. Images marked decorative with a presentation role, or with an empty alt on something that looks like page furniture such as a spacer or border, are correctly skipped. Screen readers should stay silent on those, and marking them is good practice, not a fault. An empty alt on what looks like a photograph scores nothing, because a listener never learns it is there. Everything else is a content image and gets assessed.
How each image earns points
Every content image scores one point, half a point, or nothing. The score is the points earned divided by the points available, as a percentage.
- Nothing for alt text that fails outright: no alt attribute at all, or junk such as filenames, placeholders, URLs, labels like "image" or "logo", and text that just repeats the file name.
- Half for alt text that exists but falls short: too thin to describe anything, longer than a listener can take in (over 250 characters), marketing filler with no concrete content, written in all capitals, copy-pasted across three or more images, or unrelated to the page's own subject and brand.
- Full points for alt text that says what the image shows, concretely and in context.
An alt that names a person, place or brand is treated as legitimate, including when the image file is named after the same person. Names are good alt text.
What the page could score
Beside the score, every result says what the same page would score once the flagged images are described. That is not a forecast. It is the same arithmetic run forwards: an image scoring nothing is short of one whole point, one scoring half is short of half a point, so describing them adds exactly what they lack. Nothing else about the page changes.
We show it because a number on its own tells you where you stand and not what to do. The gap between the two figures is the work.
The AI check
Text rules cannot tell whether "red jumper on a chair" is true of the picture. So for a sample of images on every scan, up to eight, a vision model reads the image itself and judges the alt text against it: accurate and useful, partly right, or wrong. Failing images are checked first, then a few that passed the text rules, because alt that reads well and describes the wrong thing is the most dangerous kind. Where the text falls short, the model writes the description NarraSight would ship for that image, in your page's own subject and tone. Those verdicts are marked "AI checked" on the result and replace the text-rule verdict for that image in the arithmetic.
The business case, and what it is allowed to say
Under the result we set out what the flagged images mean for your business: who they shut out, what that means for customers and image search, and where you stand legally. That text is assembled from two things and nothing else. The figures measured on your page, and a fixed list of published facts, each naming its source.
Before any of it reaches you, every number in it is checked back against those two sources. If a figure turns up that we did not measure and cannot cite, the draft is thrown away and a plainer version is sent in its place. This tool cannot tell you that you are losing a particular sum of money each month, because no such figure exists in anything it is allowed to read. That is a deliberate limit, and it is the one we would want a supplier to put on themselves before selling to us.
Three habits follow from it. We do not tell you that you are breaking the law. We state the duty and leave you to weigh it. We describe the search benefit as image search, because that is where alt text is a confirmed ranking factor and general web ranking is not. And where the European Accessibility Act comes up, we say in the same breath that service providers with fewer than 10 staff and turnover or balance-sheet total of EUR 2 million or less are exempt from the services requirements (Article 4(5)), because your headcount is not visible from a web page and it would be careless to imply otherwise.
Which legal paragraph you see depends on what the page tells us about you. A council, an NHS body or a government site gets the Public Sector Bodies (Websites and Mobile Applications) (No. 2) Accessibility Regulations 2018, because those are what bind it. Most businesses get the Equality Act 2010 anticipatory duty. A hire business, or a small site we cannot place, gets no statutory paragraph at all, only a pointer to this page: the duty is set out here, but quoting it at what may be a one-person business on a free scan reads as a threat, and that is not how we want to open a conversation. The sector wording (a shop's product photos, a hotel's rooms) is chosen from phrases on your own page, and where we cannot tell, we say "content" rather than guess.
What this page sounds like
Above the business case, the result now reads your page's images out in order, the way a screen reader announces them: the alt text where there is one, the file name where there is not, and any placeholder word exactly as written. Nothing is added or paraphrased. It is your markup, spoken. Where our judge wrote a better description during the scan, a second button plays the same run with those descriptions in place. Images the judge did not look at are read unchanged in both versions, so the "after" never contains a description that does not exist. The audio is generated by your own browser and nothing is sent anywhere to produce it.
What the score is not
Here is what the number cannot tell you.
- It reads one page per scan. Where the page looks JavaScript-built, the scanner opens it in a full web browser, lets the scripts run, and scores the rendered result. When that is not possible within its time budget, it scores the raw HTML and says so on the result.
- A page with no visible content images gets "not assessed", never a perfect score.
- A few sites refuse every automated visitor outright. We try twice, first with a direct fetch and then in a full browser, and if both are refused we tell you so rather than guessing. NarraSight itself is unaffected: it works inside your CMS, not from outside.
- The AI checks a sample, not every image, and some image hosts refuse automated fetches, and the result states how many images were checked against the picture itself.
- It measures image text alternatives under WCAG 2.2 Success Criterion 1.1.1 only. It is not a whole-site conformance audit and we never present it as one.
- Scanning reads your public HTML and changes nothing on your site.
The legal context
Text alternatives for images are the first thing WCAG 2.2 asks for (Success Criterion 1.1.1). The European Accessibility Act requires listed services to present non-text content in an alternative form (Annex I, Section III), and its technical baseline in practice, EN 301 549, incorporates WCAG 2.1 Level AA. The UK Equality Act 2010 sets no technical standard, and no UK court has yet given judgment on website accessibility. A low score is not a finding of unlawfulness, and a high score is not a compliance certificate. It is evidence about one requirement, stated plainly.
Why it is built this way
NarraSight was founded in Glasgow by Abdul Eneser, who is blind and uses a screen reader every day. The score exists because most accessibility numbers flatter the site owner and tell the screen reader user nothing. We would rather show you a hard number with its working than a reassuring one without.