Scoring Methodology

The marker answers exactly one question: is this page better than it was last time? A team runs a scan, gets a list of findings, fixes some of them, and runs the scan again. If the work landed, the figure goes up. That is the entire job. It is a progress indicator for a single site against its own history, and it was built to be read in a series rather than on its own.

It is presented as "Progress marker (0-100)" with the range in the label and the figure standing alone underneath. That is deliberate and it is worth explaining, because it looks like a small formatting quirk and is not one. A number printed against a denominator, or as a percentage, reads instantly as a mark — the same shape as an exam result — and readers carry that reading straight into conversations about conformance where it does not belong. Putting the scale in the label keeps the range visible without ever rendering the figure as a fraction.

Everything else people want to do with a number like this, it is bad at. It cannot rank your site against a competitor's, because two pages with different templates, different element counts and different third-party embeds produce findings that are not on the same footing. It cannot be set as a contractual target, because a team optimising the figure will fix whatever is cheapest per point rather than whatever is worst for a person — and the section below on what this method costs makes that concern sharper, not softer. Used as intended, on one site, one template, a baseline and a rescan, it does a useful and modest job.

Here is the whole calculation, in the order the code performs it. It is short, and the shortness is the entire design goal.

A report groups the deducted findings by rule, because that is how they get fixed: one rule is usually one change. The grouping is presentation, and it does not touch the arithmetic. Eighteen findings from a single rule and eighteen findings spread across eighteen rules both deduct the same eighteen points, and you can add up the rule groups or count the raw rows and land in the same place.

Every deducted finding still carries one of four impact labels — critical, serious, moderate, minor — and the report still prints it beside the rule. That label is what tells you which thing to fix first. It no longer changes the figure, and the section below is honest about what that trade cost.

Take a page with three distinct rule failures, plus two kinds of finding that get reported without being priced. A brand colour fails the contrast rule on eighteen elements, each finding labelled serious. Three form fields have no programmatic label, each finding labelled critical. One link has no accessible name, labelled minor. Alongside those, the scan reports twelve findings from the rule that wants every block of content inside a landmark — one of axe-core's best-practice rules — and two contrast checks it could not decide, because the text sits on a background image.

The deduction is eighteen plus three plus one, which is twenty-two, and the marker reads seventy-eight. The findings list has thirty-six rows in it: the twenty-two that were deducted, the twelve best-practice findings and the two review items. All thirty-six are in the report, in the counts and in the export. Fourteen of them moved the figure by nothing at all, and the page says so rather than quietly folding them in.

Notice how little there is to get wrong. The contrast rule dominates the deduction because it genuinely produced most of the failures, not because a weight decided it should, and one CSS change clears all eighteen at once.

You can run this yourself on any AccessWiser scan. Export the findings, set aside the rows marked as needing review and the rows marked best practice, count what is left, subtract that count from one hundred and floor at zero. There is no step you have to take on trust and no grouping you have to guess at. If you do not land on the figure we printed, that is a bug and we want to hear about it: a published method that cannot be reproduced from the output is not really published.

The marker used to be a weighted figure, and it stopped being one on the second of September, 2026. The retired model deducted ten for a critical finding, five for a serious one, two for a moderate and one for a minor, then capped each rule's total contribution at twenty so that one repeated problem could not consume the whole scale. It ranked severity correctly. We replaced it anyway, and the reason is the property that matters most on a page like this one.

It could not be reproduced from anything on screen. The report printed four severity totals beside a formula that works per rule, and the cap applied to a grouping the totals did not show — so a careful reader, capping the three severities in front of them, would arrive at fifty-six where the answer was twenty-two. The person who built it did exactly that. A figure a customer cannot check is a figure they are entitled to distrust, and publishing the method does not rescue a method the output will not let you apply.

Counting has a real cost, and we would rather print it than have you discover it. It treats every finding as equal, so it rewards volume over severity. On a real site we scanned, clearing a single contrast rule — one CSS change, eighteen findings — moved the marker by eighteen, while clearing a keyboard trap that appeared exactly once moved it by one. A team optimising this figure will fix the numerous thing before the blocking thing. That is a genuine defect in the marker as an incentive, it is the price of a figure anyone can reproduce, and it is the strongest argument on this page for reading the findings list rather than the number printed on top of it.

The severity is not lost, only unpriced. Every finding still carries its impact, the list still shows it, and the report still groups by rule with the impact beside each group. What changed is that the figure no longer encodes it, so the order you work in has to come from the list and from what a barrier does to a person.

Counting is also still sensitive to page length, and only partly cured of it. A longer page has more elements, so one template mistake repeated down it produces more findings and costs more points than the same mistake on a short page. That much is a real failure being counted more often, which we can defend. The next section is about the half of the problem that was not defensible.

One consequence of changing arithmetic, handled the same way every time: a scan that has already been issued is never re-scored. The figure a scan was given is the figure it keeps, so a history spanning a change of method shows a step rather than a smooth line. We would rather show you the step than quietly restate what an older scan found.

The most recent change to this method, made on the ninth of September, 2026, was to stop deducting for findings that are not failures of anything. It deserves the space, because the bug it fixed made the marker actively misleading in a way no reweighting could have corrected.

Two kinds of finding are advice rather than conformance failures. The first is axe-core's own best-practice set: around thirty rules that the engine tags that way and gives no success-criterion tag at all — the rule that wants every block of content inside a landmark, the rule that wants heading levels in order, the rule that wants exactly one main landmark. They are good advice and worth acting on. axe does not claim they are WCAG failures, and neither will we. The second is our own rule pack, which has always been able to mark a rule as advisory so that a consumer can present it as advice; the marker was that consumer, and it charged a point regardless.

Here is what that cost, measured rather than argued. The landmark rule reports every top-level block of content sitting outside a landmark, so the number of findings it produces grows with the length of the document. A page whose only defect was a missing main element scored ninety-seven with one block of content, sixty-eight with thirty, and zero with one hundred. The same page with a main element added read ninety-nine at every length. Four lines of markup moved a page from zero to ninety-nine, and a long, carefully built page could come out below a short, broken one. In that band the marker was a function of page length, and it fell as a site published more.

No change to weights fixes a term that grows with the size of the document. The term had to stop being priced. Under the current model that same page reads one hundred at every length, and the landmark findings are all still there — a hundred and fifty-one of them on a page of a hundred and fifty blocks, listed, exported and fixable. Nothing was hidden and nothing was dropped from the report. What changed is that we stopped deducting for something we were not prepared to call a failure.

The test is the marking, not the absence of a citation. Treating "this finding names no success criterion" as the test would have been the tempting shortcut and it is the wrong one: a rule whose mapping was missing or broken would silently stop being priced, turning a mapping bug into a quiet under-report. A finding is treated as advice only when something that knows it is advice has said so.

This decision cuts against us commercially, and it is worth saying so plainly. It raises figures, and a higher figure is a weaker argument for buying remediation work. It is the right arithmetic anyway. A number that falls because a page got longer is not measuring accessibility, and deducting for a best-practice suggestion means asserting a conformance failure that the check itself never claimed — which is the one thing a tool in this category must not do.

This is the part most vendors leave out, so it goes here in plain words: there is no accessibility standard anywhere that defines a numeric marker, a severity weighting, or a threshold to clear. WCAG 2.2 is a set of success criteria, and each one is either satisfied or not satisfied for a given page — the standard has no arithmetic in it and offers no way to trade one criterion off against another. Section 508 incorporates WCAG by reference and works the same way. So does EN 301 549 in Europe. Conformance under all three is a per-criterion determination, most of which requires a person to make it.

So when any tool shows you a figure out of a hundred, that figure is the tool vendor's opinion about how to add up its own findings. Ours is our opinion. The arithmetic on this page was chosen by us — the starting figure, the cost of one finding, the choice to price advice and review items at nothing, the floor at zero — and every one of those could have been chosen differently by reasonable people. Nothing about them is derived from a standards body, and any tool implying otherwise is overstating what it has.

What we can honestly claim is the difference you are reading right now. Our opinion is published, versioned with the product, and reproducible from the output of any scan. If you disagree with a decision here, you can name which one and work out exactly what it does to your figure, because the inputs are rows you already have in front of you. That is not possible with a number whose method is private, and it is the only reason a figure like this deserves any trust at all.

One practical consequence for anyone comparing tools: if another product gave your home page a much higher number than ours did, neither of us has caught the other in an error. We have applied different arithmetic to different rule sets with different severity labels. The productive comparison is not between the two figures — it is between the two findings lists, item by item, on your actual page. Those you can check.

There is exactly one outside anchor we consider legitimate to quote alongside a scan, and it is not a score of any kind. The WebAIM Million is an annual automated analysis of the home pages of the top one million websites, run by WebAIM at the Institute for Disability Research, Policy, and Practice at Utah State University. It has been repeated every year since 2019 with a consistent method, which is what makes it useful.

In the 2026 edition, the average home page carried 56.1 detectable errors, up from 51 the year before — the figure has risen in most years the study has run. The average home page contained 1,437 elements. Those two numbers together are the reason a raw finding count means something: they tell you what an ordinary page on the open web actually looks like when the same kind of automated detection is pointed at it.

So if your scan comes back with ninety findings, the useful sentence is not about a marker at all — it is that a typical home page has around fifty-six detectable errors and yours has more, and here they are, listed. That is a comparison against a published, repeated, independent measurement rather than against our own arithmetic. Source: WebAIM Million, webaim.org/projects/million/. We are not affiliated with WebAIM and quote the study with attribution.

One side effect of counting is worth noting here, because it makes that comparison easier to hold in your head. The marker and the deducted finding count are now the same statement rearranged: a page with ninety-two deducted findings reads eight. That is a pleasant coincidence rather than a design goal, and it is still not a reason to lead with the marker instead of the list.

WebAIM are careful about the same limitation we are, and it should be repeated here: their study detects errors automatically, which means it finds the machine-decidable subset. A page with no detectable errors is not therefore a usable page. The study measures what automated tooling can see, across a very large sample, consistently — no more than that, and no less.

The limits below are not disclaimers bolted onto a good number. They are the actual boundaries of what this kind of arithmetic can do, and knowing them is what lets you use the figure well.

Treat the marker as an internal instrument. Set a baseline on the templates that matter — home, a category page, a product or booking page, checkout — and rescan the same URLs after each round of remediation. Movement on your own baseline is meaningful. A single reading on its own is close to meaningless, and a reading from a different tool is a different instrument entirely.

Read it beside the findings list, never alone. The figure can move for reasons that have nothing to do with your remediation work: a new promotional banner, a third-party embed updating itself, a marketing page gaining a carousel. When the number changes, the findings list tells you which of those it was.

Order your work from the impact labels, not from the arithmetic. Because every finding costs the same point, the cheapest way to move the figure is to clear whichever rule fired most often, and that is very often not the thing standing between a person and your checkout. The list is sorted so you do not have to make that mistake.

When you report progress upward or outward, report the findings. "We closed the eleven distinct rule failures our scans found on checkout, and here is the dated record of when each was fixed" is a true, checkable statement about work performed. The marker is a reasonable footnote to that sentence and a poor replacement for it, which is the same order of priority the reports themselves use.