Search visibility was a counting problem: tools counted, analysts decided what the counts meant. The Jev tools in the catalog move the judgment into the run: code crawls, parses, and does the arithmetic; the model answers narrow typed decisions, and a threshold turns each answer into an action.
Code finds, Jev judges
Jev SEO Audit states the division in one line: code finds, Jev judges, the dashboard shows the probabilities. Paste a URL and a deterministic crawl fills a wall of findings; Jev answers narrow typed questions with Choice, Score, or Noul, and only decisive answers become ranked findings. Jev brings no index, volume data, or backlink graph, so nothing invents a monthly search count; it also ships UNLICENSED and publishes no benchmarks.
Jev GEO Audit puts measured numbers on the same idea. It audited 300 pages from 10 sites with nine decisions each, turning probabilities into a citability score and a fix list. The run cost $0.0228 for 300 pages versus $1.2063 for an LLM judge on 100 pages, 158 times more per page; median latency was 61.1 ms per page for Jev, against 2,837 ms for the judge. The two agreed on 85.8% of 900 decisions, rising to 98.2% when Jev's confidence was 0.90 or higher. That is agreement, not accuracy: the judge is another model, and the sample is 10 sites.
Three audits, three rulebooks
Three site-audit tools share a crawl shape, not a rulebook.
| Audit | Rulebook | What Jev adds |
|---|---|---|
| jev-seo (Claude Code skill or CLI) | 52 rules tied to Google Search Central | page type, intent, importance, trust, title fit |
| jev-seo (Rust CLI and MCP server) | 50 rules (R01–R50, nine areas) | intent classification, competitive gaps, AI visibility |
| Jev SEO Audit | none published | findings ranked only when decisive |
The 52-rule audit ranks repair work and never predicts rankings or traffic; its blind second judge is a separate model, so agreement shows consistency, not correctness. The Rust CLI reranks live DuckDuckGo SERPs and audits schema, robots.txt, sitemaps, and llms.txt without paid API keys; it scrapes rather than licensing SERP data, and only its Jev scoring needs a TYPESAFE_API_KEY.
Internal linking, judged link by link
jev-linkmap is the cleanest demo of the split. A Python pipeline picks candidate targets and in-copy anchor phrases, asks Jev whether each deserves a link, and places only those clearing the threshold. One request per page carries the copy once and asks 30 questions: a yes/no link probability and an anchor choice for 15 candidates. A recorded run: 566 pages, 8,460 decisions, 5.95 seconds, $0.27 total, versus about $67 for Claude Opus 5 on the same queue, with at most three new links per page. A System Two referee rewrote the rubric between two passes, moving anchor agreement from 71% to 88% for $15.51 once. Caveats: one site, one date; a referee that is a strong reader, not ground truth; NOASSERTION license.
Ian Nuttall's Internal linking tool sells the same job per page: crawl up to 2,000 pages, map every internal link with anchor text and position, and let Jev classify each page and its search intent before recommending placements. Free up to 100 pages; bigger runs cost 1 credit ($1) per 500 pages, refunded on failure, though purchases were paused at review time. Its crawler honours robots.txt and skips canonical duplicates; no accuracy figures are published. The economics have their own read in The Cost of a Decision.
Scoring search results before you click
LinkScout scores the SERP itself. Each result gets one systemOne request with six questions (relevance, depth, SEO filler, sales page, page type, key passage), then a 0-100 badge computed in code: (0.65*relevance + 0.35*depth) * (1 - 0.7*seo) * (1 - 0.5*sales*(1-transactional)). The extension re-ranks Google, Bing, DuckDuckGo, and 104.com.tw, caching results for 24 hours per query and URL and fetching pages through Cloudflare Browser Run with a Jina Reader fallback. Fast mode: browser-fetched pages 0.06 to 0.4 seconds, Jev about 0.25 seconds, against 3 to 30 seconds for server fetchers. MIT-licensed, but it needs Cloudflare, TypeSafe, and optional Jina keys. The formula is composite scoring in miniature: the same score-each-candidate shape as retrieval pipelines.
Whether AI engines mention you
Notra attacks the AI-answer side: run buyer prompts across ChatGPT, Claude, Gemini, and Perplexity, then compare mention rates, recommendation positions, and share of voice by engine and prompt. It ships a REST API, an MCP server at mcp.usenotra.com/mcp, and the @usenotra/geo traffic SDK for Next.js, Nuxt, Astro, and SvelteKit that separates training crawlers, search indexing, assistant browsing, and human referrals. Jev sits inside a broader application, and setup is heavy: AGPL-3.0, commercially hosted and paid, with local development wanting Bun 1.4.0, Node.js 24.11.1, PostgreSQL, WorkOS AuthKit credentials, and provider keys.
Free tools that stop at one judgment
Not every visibility tool needs a crawl. slopcheck is a CLI and Claude Code Stop hook that reports which AI tells a passage contains, not whether AI wrote it: regex checks 18 tells for $0, one Jev call judges 15 more, and thresholds live in slopcheck/tells.yaml. Answers are line-level — "line 4 is a binary contrast, p=0.91" — and a Noul between 0.40 and 0.60 is treated as an uncertainty band: suppressed as the model not answering, not read as a weak yes. Caveat: the numbers come from nine unnamed drafts.
Slop Mop takes the per-item judgment to a LinkedIn feed: a free Chrome extension, built by Tom Frazier, that scores every post out of 100 before you reach it and folds anything past your threshold into a 44px strip you can unfold in one click. Its Slopprint shows nine tells, six from Graphite's AI-tells research and three the author's own, and your own vote (Not sure, No, Maybe, Probably) overrides the model. It calls itself a bad-writing detector, not an AI detector. No accuracy figures are published.
One judgment per question, arithmetic in code, probabilities with caveats deciding what changes.
