nika-site-audit

A full SEO + GEO site audit as one workflow file.
Discover, crawl bounded, score deterministically, one narrated verdict, artifacts as receipts. Checked before a single request leaves your machine.

The whole audit, drawn by nika itself

This diagram is generated by nika graph site-audit.nika.yaml --format mermaid, not drawn by hand. Eleven tasks, five waves:

graph TD
  robots["robots · invoke · nika:fetch"]
  llmstxt["llmstxt · invoke · nika:fetch"]
  sitemap["sitemap · invoke · nika:fetch"]
  crawl["crawl · invoke · nika:fetch"]
  rubric["rubric · invoke · nika:read"]
  geo_score["geo_score · invoke · nika:jq"]
  seo_score["seo_score · invoke · nika:jq"]
  scorecard["scorecard · invoke · nika:chart"]
  verdict["verdict · infer · mock/echo"]
  report["report · invoke · nika:write"]
  done["done · invoke · nika:log"]
  crawl --> seo_score
  geo_score --> report
  geo_score --> scorecard
  geo_score --> verdict
  llmstxt --> geo_score
  report --> done
  robots --> geo_score
  rubric --> verdict
  scorecard --> report
  seo_score --> report
  seo_score --> scorecard
  seo_score --> verdict
  sitemap --> seo_score
  verdict --> report

Loading

One model call in the whole pipeline (verdict). Everything else is deterministic: fetches, jq scoring, a chart, a report. That is the point.

Quick start

brew install supernovae-st/tap/nika     # single binary

nika check site-audit.nika.yaml         # the audit BEFORE the audit:
                                        # plan · cost floor · permits · secret flows
nika run  site-audit.nika.yaml          # offline verdict (mock model) · artifacts in out/
nika test site-audit.nika.yaml          # the committed golden guards the outputs

Live run against your own site, with a real model · local first, nothing leaves your machine but the audit itself:

nika run site-audit.nika.yaml \
  --var url=https://your-site.com \
  --model ollama/qwen3:8b        # any provider/model works the same way:
                                 # mistral/mistral-medium-latest ·
                                 # anthropic/claude-sonnet-4-6 (needs its env key · `nika doctor`)

Note the permits: block in the workflow pins the network to the audited host. Widen it when you change vars.url; everything else stays default-deny.

What it does

Wave	Tasks	What happens
1	robots llmstxt sitemap crawl rubric	GEO discovery in parallel: robots.txt, llms.txt, sitemap.xml, plus a bounded same-origin crawl (8 pages, robots respected). A missing file is a finding, not a crash
2	geo_score seo_score	deterministic scoring in jq: AI-crawler access, llms.txt presence, titles, canonicals, status codes
3	scorecard verdict	a chart artifact from the scores, and the one LLM call: the narrated verdict, grounded in the rubric
4-5	report done	everything folds into out/report.html + out/scorecard.svg

Every run writes a hash-chained trace. nika trace verify exits 0 or names the first broken link: the receipt comes from the runtime, not from the model's prose.

Why this shape

The interesting failures in an audit pipeline happen before the model is called: a crawl that hammers a site, a secret that leaks into a prompt, a cost that surprises you. nika check catches those statically, and the permits: boundary makes the blast radius readable in the file itself. The model gets exactly one job: narrate the verdict from scores computed deterministically.

The brouillon-era legacy showcase

legacy/site-audit-v10.nika.yaml preserves the original 1072-line pipeline (April 2026, the brouillon era — the draft engine the Diamond rewrite retired): 50 tasks, 22 layers, locale detection across 38 locales, an interactive 10-tab dashboard, AI-generated imagery and a podcast narration. It ran against the pre-rewrite nikab binary in the retired pre-v1 dialect and does not parse on current engines, by design. Recorded legacy-era results:

Site	Pages	Locales	Tasks
qrcode-ai.com	1,654	38	50/50
htmx.org	278	1	50/50
tailwindcss.com	~1,900	1	50/50

Keep it as evidence of scale and a design archive. Do not learn the syntax from it; the current language is the public spec. The full era model — exploration · brouillon · diamond — is the machine-verified timeline.

Files

File	What it is
site-audit.nika.yaml	the v1 workflow: 11 tasks, declared permits, offline by default
site-audit.nika.yaml.golden.json	the committed golden: nika test compares typed outputs against it
skills/	the audit rubric the verdict is grounded in
legacy/	the brouillon-era showcase (April 2026), read-only
prototypes/	exploration, not the product

_{Docs: docs.nika.sh · Engine (AGPL-3.0): nika · Templates: nika-starter · nika-actions-starter}