jev-seo skill

Run the binary. Prefer JSON output when chaining commands. Never invent scores; every number below comes from a command.

Fast paths

• Full site picture: jev-seo crawl <url> --max-pages 50 then jev-seo llms <domain>.
• Own-site truth: jev-seo gsc auth, then jev-seo gsc query --site <url> for free Search Console clicks and positions.
• Docs folder gate: jev-seo audit docs/ --min-pass 60 --html report.html --pdf report.pdf --md report.md.
• One question, one answer: jev-seo geo <file> --query "<query>", jev-seo robots <domain>, jev-seo schema <file>, jev-seo sitemap <url>.
• Internal links: jev-seo link docs/ --limit 10 suggests one contextual target per page, or no_link.
• Environment check first on a new machine: jev-seo doctor.

Reading output

• Health, Score, and Grade lines are deterministic 0-100 values with A-F grades.
• Top Actions are ranked by impact with effort bands; ids like RULE-R01 map to the rule registry (R01-R64) in src/rules.rs.
• New in v0.4.0 (unreleased): check + verify, .jev-seo/policy.toml, regional --country/--language/--device, GSC inspect/winners/losers, bing sites/queries, --sarif/--junit, R59-R64, citation struct, seo_index, 4 new MCP tools (19 total), jev:// resources, evidence snapshots. Portable manifest (plugin.json + mcp.json + .app.json), annotated tools, bundled skills/jev-audit/SKILL.md, bridge GET /health. New in v0.3.0 (2 Oct 2026): serve, capabilities --json, watch --diff-only, fix-plan, crawl sitemap seeding (closes #3), R32/R08/R56 breadth, JS feature gate, GEO trends with sparkline SVG, offline --rescore. v0.1.2 shipped 26 Sep: explain, report --baseline, crawl --vitals, query --provider dfs, seo_cite_check, seo_gap, keyless GEO, drift gates.
• Jev contract: Choices offer insufficient_context, Flags print the runner-up, every fan-out appends to ~/.jev-seo/eval.jsonl. Question version rides the spend line.
• Completeness says what was skipped (caps, missing sitemap, missing robots). A capped crawl is a sample, not a verdict.
• Needs review marks Jev answers below the confidence bar. Treat those lines as unverified.
• --json keeps stdout machine-clean; progress goes to stderr.

MCP server (19 tools)

Run it over stdio: npx -y @akashpriyadarshii/jev-seo (or jev-seo mcp with a local build). Client config:

{ "mcpServers": { "jev-seo": { "command": "npx", "args": ["-y", "@akashpriyadarshii/jev-seo"] } } }

No local build? Set JEV_SEO_BIN=/path/to/jev-seo and the wrapper uses it instead of downloading. First npx run needs network for the one-time download; musl/Alpine systems cannot run the glibc builds and must use JEV_SEO_BIN with a local cargo build.

seo_cite_check reports whether an AI answer names the target domain. Three paths, first one that works wins: pass answer to score your own words (works everywhere); omit it and sampling-capable clients answer via sampling/createMessage; otherwise you get needs_answer: true plus the prompt, answer it naturally as you would a user, and re-call with answer. Result fields: cited (bool), since_last (previous verdict or null), source (answer/sampling/engine), excerpt (300 chars).

Agentic loop: seo_gap ranks the Search Console queue (needs site auth once via gsc auth); work it top down; seo_cite_check each query; seo_report surfaces drift alerts (citation lost/gained, geo drops). seo_geo scores keyless without any key; TYPESAFE_API_KEY upgrades it to Jev judgment. Paid engines (Perplexity, OpenRouter, Ollama) plug in via JEV_SEO_LLM_KEY + JEV_SEO_LLM_MODEL (+ optional JEV_SEO_LLM_URL) and answer cite checks server-side. Deploy gate: drift baseline --report audit.json --label v1 before a change, drift compare --report new.json --label v1 after. One-file handoff: audit --digest digest.md, canonical close: bundle audit.json --out bundle.md.

Rules for agents

1. Crawl before judging. Never score pages the crawl has not fetched.
2. Quote the completeness line alongside any score.
3. Gate deploys with audit --min-pass or crawl health deltas via --diff, not vibes. --fail-on blocking (default) fails only on deterministic rules; judgment calls warn. --forbid names non-negotiables.
4. Export findings with --csv for spreadsheets; rebuild past runs with --rescore, no spend.
5. Jev needs TYPESAFE_API_KEY. Without it the tool still runs; semantic sections report local-only output.
6. Keep crawls polite: default 50 pages, 8 workers, robots.txt honored. Raise --max-pages deliberately.

Capabilities

jev-seo capabilities --json reports {version, rule_set, rules, mcp_tools, mcp_tool_names[], commands[]}. This is the machine source of truth. Do not parse counts from the README.

Plugin bridge

jev-seo serve --addr 0.0.0.0:8787 exposes Streamable HTTP at POST /mcp plus GET /plugin.json and /extensions/* for ChatGPT Plugin Extensions.

Fix plan

jev-seo audit docs/ --json > audit.json && jev-seo fix-plan audit.json --json: deterministic rule to patch to verify per finding (fact or heuristic, confidence capped), sorted blocking first. One hint per finding. Verify with jev-seo audit <file> --json | jq '.findings | no RULE-Rxx'.

Watch

jev-seo watch docs/ --once --no-jev or jev-seo watch https://example.com --once --json; loop mode jev-seo watch docs/ --every 30. Compares against watch-<slug> sqlite baseline, emits clean|regressed|improved|baseline with delta, blocking counts, drift alerts. Polling, not inotify.