CRO audit
Find what stops a visitor from acting, in the order worth fixing. The score is a structured review, not a prediction; its value is the fix order.
Inputs
- The page (URL or file) and its screenshots at 390 and 1440 (landing-qa's
shot.py). - The page job and the traffic source. Table:
references/page-jobs.md. - The ad or post the visitor clicked.
- Any real numbers the owner has (conversion rate, bounce, scroll depth), with their source.
Steps
- Set the page job. Write it in a copy of
templates/cro-scores.json. A page is judged against its job; a waitlist with no price is fine, a cold sale with no price is not. - Run first-screen-check first. Its result feeds headline, above_fold, cta and message_match.
- Score the 12 dimensions 0 to 10 with the anchors in
references/rubric.md. Write a note for every score under 8: the block, what is wrong. - Total it.
python3 scripts/cro_score.py cro-scores.json. It applies the weights, the first-screen hard gate (under 60% of the first five dimensions is BLOCK), the grade, and the ordered fix list. - Write the fixes in three groups: quick wins, high impact, test ideas. Each names the dimension, the block, the exact change and, for copy, two alternative lines.
- Hand back to the build loop (landing-factory) or to the owner as a TEARDOWN report.
Gates
- The page job is set before any score.
- Every score under 8 has a note naming the block.
- Verdict not BLOCK and grade C or better before paid traffic.
- No outside conversion statistic in the report without its source next to it.
Output
CRO audit: site/index.html (page job: cold subscription, paid social traffic)
dimension score weight points note
headline 8 18 14.4
cta 8 15 12.0
...
trust 6 5 3.0 no real product photo yet (placeholder), no named roasters in About
first-screen gate: 53.8 of 67 (80%, needs 60%)
total 78.3 of 100, grade B, verdict B
fix list (fix order stage, then points lost):
1. [2 headline] headline: 3.6 points lost.
...
Full run: examples/northwind-roasters/out/cro_score.txt from examples/northwind-roasters/cro-scores.json.
Failure modes
| Trap | What it looks like | Instead |
|---|---|---|
| Generic ideal | a waitlist marked down for "no pricing" | score against the page job |
| Polish first | button colors debated while the headline is a slogan | the fix order: job, headline, sections, proof, friction, polish |
| Borrowed statistics | "red buttons convert 21% better" | only numbers with a source, ideally the owner's own |
| Score as forecast | "grade B means 4% conversion" | the score orders fixes; tests measure conversion |
| Vague fixes | "improve trust" | the block, the exact change, two alternative lines |
| Good FAQ hides a bad hero | total 70, first screen weak | the hard gate blocks it |
| One reviewer's taste | a 3 where a colleague gives an 8 | two reviewers on the big five; talk when they differ by more than 2 |