Ruleset 2026.11.3 · updated 10 November 2026

How the Craft score works

We open your page in a real browser and run 62 checks. The Craft score comes from the 21 that proved, on real pages, to separate unmodified templates from the home pages of established product companies. This page says what the score measures, how we test it, what a review can't see, and how much the score moves when the same page is scanned again.

What it measures, and what it doesn't

• It measures how much of a page still looks the way a starter template or an AI site builder leaves it. It doesn't try to tell whether AI was used, and it isn't a judgment of taste.
• Nothing else goes into it: not an AI's opinion, not ours, and not what the page's owner tells us about the page.
• A high score means few of those signs. It doesn't prove a page is clear or persuasive. That takes a reader, which is what the editor's read is for.

How the score is built

• The Craft score starts at 100. Each scored finding lowers it, and a finding counts for more when its check separates templates from established pages more sharply.
• A check is scored only if it earned it: on real pages it had to fire clearly more often on templates than on established product pages.
• Accessibility and Performance problems are real, but they turn up as often on well-made pages. They are scored in their own two areas and are not part of the Craft score.
• Everything else a report lists is a note: an observation that didn't separate the two kinds of page. Notes don't move any number.

How we test it

We scanned two sets of real pages: 41 home pages of established product companies, and 101 pages that are a template's or a generator's own output (54 unmodified starter-template demos and 47 pages hosted by AI site builders). Each set was split in half. The rules were chosen on one half, then tested on the other half, which they had never seen.

On that unseen half (24 company pages and 54 template pages, ruleset 2026.11.3, 10 November 2026): take one page from each group at random, and the company page scored higher 97% of the time. A coin flip would be 50%.

Read this as a ceiling. These are two extremes; most real pages sit in between, where the score will separate less cleanly. “Established product company” is our judgment, not proof that every choice on those pages is right. AI-builder pages are identified by where they are hosted, so we can't tell how much each was edited afterwards. And some of what the score rewards is simply a page being finished.

The editor's read

After the checks, an AI editor reads the page in context. It can add a finding only by quoting the page or pointing at a screenshot, and it can say when a pattern the checks flagged looks like a deliberate choice. Its findings are marked “Editor's read” and don't change the score. Patterns it reads as deliberate stay in the list and still count; they're left out of the fix prompts. Its own number, shown next to the area scores, is an opinion and can differ from one run to the next.

What a review covers

• One page per review, loaded without signing in. Anything behind a login isn't seen.
• Two sizes: a 1440 × 900 desktop window and a 390 × 844 phone. Most checks read the desktop version; the phone pass checks layout, the headline and the buttons.
• The page is scrolled to the bottom (up to 30,000 px) before it's read, so content that loads on scroll is included.
• Up to the first 6,000 elements, 60 headings, 50 buttons and 30,000 characters of visible text. Color contrast is checked on the first 400 pieces of text.
• Screenshots cover the first three desktop screens (2,700 px) and the first phone screen. Findings further down the page have no screenshot.
• Keyboard focus is checked on the first four tab stops, and one made-up address is requested to see the 404 page.
• The copy checks know English and Spanish phrases. On pages in other languages they find less.
• Performance comes from Google PageSpeed's phone test. When that isn't available, the Performance area is scored on search and sharing basics only, and the report says so.
• Cookie banners are closed when they have an obvious reject or close button. A banner we can't close is reviewed as part of the page.

How much the score moves between scans

On 10 November 2026 we scanned a fixed set of 50 public pages 5 times each, minutes apart, and scored every scan with ruleset 2026.11.3. 47 pages loaded at least three times and are counted here; 3 answered with an error (blocked or not found) and are left out.

This run did not include Google PageSpeed. That only affects the Performance area, which is not part of the Craft score.

Changelog

Review a page