Should I try to get a perfect audit score?

Each finding carries an impact weight, so a critical failure counts several times what a low-impact one does.

The status matters: a failure counts fully, a warning counts partially, and a pass or an informational result counts nothing against you.

Confidence scales the result, so something inferred rather than directly observed moves the number less than something measured.

Repeated findings in the same group are damped, so twenty instances of one defect do not count twenty times — the first counts most and the rest progressively less.

Unverifiable results contribute nothing in either direction, which is why they affect coverage rather than the score.

The weighting means the expensive findings are fixed first, because they are the ones moving the number.

What is left at a high score is low-impact, frequently damped, and sometimes inferred — three multipliers all working against the effort.

Some of it is not fixable on your side at all: a third-party script you do not control, a platform behaviour, a header a provider sets.

And some of it is a deliberate trade you made, which the score cannot know about.

So the effort per point rises steeply while the real-world benefit per point falls.

Direction over time: the same site measured the same way, where a fall means something changed.

A release check: a score that drops after a deployment is a signal worth investigating.

A rough comparison between sections of one site, where the methodology is constant.

A conversation opener with people who do not read technical reports, which is a real use and not a technical one.

None of those requires the number to be high — they require it to be consistent.

Comparing two tools, which weight and group findings differently and will never agree.

Comparing two sites, whose sizes, stacks and audiences make the same number mean different things.

Deciding what to fix, which the findings list does better because it carries the reasons.

Proving anything to anyone: a score is a summary of an opinion about weights, and a reasonable person could weight differently.

The findings that describe something a visitor or a customer experiences — a broken page, a slow paint, an unreachable form.

The findings with a security or a data consequence, regardless of what they do to the number.

The findings that keep recurring after each release, which are a process problem rather than a page problem.

The coverage figure, since a clean result over more of the site is worth more than a higher number over less of it.

And stop when the remaining list is things you would not fix if no report existed.

A critical failure observed directly carries a weight of ten, multiplied by a full loss for a failure and a confidence factor of one.

A low-impact warning carries a weight of one and a half, multiplied by a partial loss of about a third — roughly half a point of weight, twenty times less than the critical failure.

The same low-impact warning inferred rather than observed is scaled down again, to three quarters of that.

And the fifth instance of it in the same group counts a fifth as much as the first.

So at a high score, fixing one remaining item can be worth a small fraction of what fixing one of the first items was — which is the arithmetic behind the advice, not an opinion about effort.