JevForAgents
⌘K
PricingSubmit a buildSponsors
Examples / Jev SEO Audit by Daniel Agrici
FIRST-HAND TESTMetric claims: AUTHOR REPORTEDSOURCE AUDITEDπŸ“ˆ Jev for SEO⚑ Jev Skills

Jev SEO Audit by Daniel Agrici

A Claude Code skill and CLI that audits a live site, keeps rule checks in code, and uses Jev for typed page judgments.

𝕏 View Original Post on X β†—πŸ™ Open Repository on GitHub β†—
MEDIA PREVIEW
Jev judgment pages from the original project report

Jev judgment pages from the original project report. Original screenshot in the project's README, showing judgment cards, quality heatmap and investment priorities.

Overview & Result

Daniel Agrici's jev-seo takes a homepage URL, crawls the site, applies documented rule checks and PageSpeed data, then asks Jev typed questions about page type, intent, helpfulness and related judgments. It ranks actions in code and renders PDF, XLSX and Markdown from one audit file. The public repository includes a Claude Code skill, CLI, example output and offline tests.

How Jev fits in the loop

  1. A live crawl and PageSpeed Insights collect page and performance evidence; optional DataForSEO calls add keyword and backlink data.
  2. Deterministic code applies 52 SEO rules and assigns action IDs, while Jev returns probabilities for semantic page and site judgments.
  3. The audit stores uncertain answers for review and renders PDF, XLSX and Markdown reports from the same audit.json file.

How to reproduce

  1. Clone https://github.com/AgriciDaniel/jev-seo and install its requirements in a Python 3.10+ environment.
  2. Run the repository's offline unit tests, then use bin/jevseo doctor to check local dependencies and keys.
  3. With a TypeSafe API key, run bin/jevseo run against a site you control; review uncertain judgments and generated action IDs before changing pages.

Why this build matters

Shows a reproducible separation between measured SEO facts, model judgments and human review, with a concrete skill, CLI and inspectable outputs.

Limitations

  • The default audit samples at most 60 pages on large sites and labels that cap in its report.
  • Jev thresholds have not been calibrated against human-labelled ground truth; the repository's blind second judge is another model.
  • An audit does not predict rankings or traffic, change the site, or track results over time. DataForSEO full mode incurs separate charges.
  • This site's review ran only the repository's 38 offline tests; it did not run a live audit or independently reproduce author-reported costs and evaluation scores.

Patterns

Agent EvaluationMCP / Skills