How this
is measured
One question, five components, and every rating quoting the page it came from.
One question, five components, and every rating quoting the page it came from.
Clarity. Is the product stated plainly, early, in words from outside the company’s own vocabulary?
Specificity. Concrete claims versus unfalsifiable ones. “Cuts onboarding to two days” beats “drives transformation.”
Jargon. Density of terms that carry no information for a buyer. A higher score means less jargon.
Proof. Numbers, named customers, demos — anything a sceptic could check.
CTA. Does the next step follow from what was just promised?
Every component is scored on a fixed seven-step ladder — 5 · 20 · 35 · 50 · 65 · 80 · 95 — never a value in between. Each step is defined by an anchor example in the rubric the scorer reads, so two pages that match the same anchor get the same number. The composite is weighted: clarity 30%, specificity 20%, jargon 20%, proof 15%, call to action 15%. Clarity carries the most weight because it is the question the index asks.
A few anchors, so the numbers on a dossier mean something:
Clarity 95 “Invoicing software for freelancers” — product named, plain words. 50 “The developer platform for modern teams” — an abstraction where the product should be. 20 “Unlock your team’s potential” — could be any company.
Specificity 95 a number attached to an outcome. 65 named mechanisms, no numbers. 20 unfalsifiable transformation language.
Proof 95 a named customer with a number. 65 named customers, no numbers. 50 a usage number but nobody named. 5 nothing a sceptic could check.
Each dossier now shows, per component, which anchor the page matched, the exact words that drove it, and what the copy would need to do to reach the next step up. The quotes are checked mechanically against the page text; a rating whose quotes are not on the page is discarded, not published.
Each page is scored by three independent passes of the same model; the published number is the per-component median. When the passes disagree by more than ten points two more are run, and a spread above twenty is marked contested on the index itself. A pass whose quotes do not appear on the page is dropped before the median is taken.
What one rating is worth. On 12 October 2026 we re-scored 42 unchanged pages once, without publishing: the median move was 6 points, one in four moved more than 10, and 12 of 42 would have changed letter. So every score carries a band (score ± at least 5, wider when the passes disagreed), a letter only changes when a new score clears the boundary by 5, and a re-score that moves more than 10 points is not published directly: two more ratings are taken and the median of the three is published instead.
Not a judgement of the product, the company, or the people. Plenty of excellent products have incoherent homepages, and some of the clearest pages belong to mediocre tools. The index measures communication, nothing else.
A model scores every page; no person reviews, approves or edits a rating before it is published. That is stated on every dossier. What stands in for review is mechanical: quotes that are not on the page void the pass, struck words must appear in the headline and may never be concrete product nouns or brand names, and a translation that contradicts the clarity score is rejected. A score is a measure of what the homepage text gives a stranger to work with — not a verdict on the company, its product, or its people.
The rubric is versioned (7329c240, shown on every dossier). If the rubric
and a score disagree, the rubric wins and the page is re-scored; nothing is corrected by hand.
Pages get rescored on request when they change. Scores are never edited by hand and never altered for payment. The record of every change lives in a version-controlled repository.
No analytics, no tracking, no third-party script, no font from anyone else’s server, no email capture. A site that criticises other homepages for noise cannot ship any.