Audit Website AeoSAFE

From the repository's own README, as read at the audited commit. Badges and raw HTML are left out.

/audit-website-aeo

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: audit-website-aeo
description: Audits a live website for AI-engine discoverability (AEO/GEO). Crawls the site, runs 16 deterministic checks plus a 6-dimension content evaluation, and produces a scored report (A-F) with prioritized fixes. Use to get a baseline before improve-aeo-geo, or to measure progress after changes.
---

# Audit Website AEO/GEO Skill

You audit a **live website** the way an AI agent would — crawling its pages, parsing structure, and judging whether the content is citation-worthy for ChatGPT, Claude, Perplexity, and Google AI Overviews.

The audit has two halves:

- **Foundational (50%)** — 16 deterministic pass/fail checks run by a script. Reproducible, no judgment.
- **Intelligence (50%)** — 6 content-quality dimensions **you** score by reading the pages, using the rubric below.

Final score = `0.5 × foundational + 0.5 × intelligence`, mapped to an A-F grade.

This skill produces a **diagnosis**. To then fix a codebase, hand off to the **`improve-aeo-geo`** skill.

---

## Workflow

Follow this sequence exactly.

### Step 1: Get inputs

Ask the user for:

1. **Website URL** (required) — the live site to audit.
2. **Crawl depth** (optional) — how many pages to crawl. Default 10, max 30.
3. **Output location** (optional) — where to save the report. Default: current directory, or `workspace/<customer-name>/` if working a customer project.

If the user already gave a URL when invoking the skill, don't re-ask — just confirm crawl depth and proceed.

### Step 2: Run the deterministic audit

Run the bundled script from this skill's `scripts/` directory. It requires only Node 18+ — no `npm install`.

```bash
node <skill-path>/scripts/aeo-audit.mjs <url> --max-pages=10 --out=<output-dir>/aeo-audit.json
```

The script crawls (sitemap + robots.txt + internal links), runs the 16 checks per page, aggregates site-wide, and writes a JSON report. It also prints a human-readable summary. Tell the user the foundational score and the failed checks.

If the script errors (site unreachable, 0 pages crawled), report the error and stop — don't fabricate a score.

### Step 3: Read the JSON report

Read the `aeo-audit.json` file. The key fields:

- `scoring.foundationalScore` — the deterministic score (0-100). This is **final** — do not change it.
- `checks` — the 16 site-wide checks with pass/fail and details.
- `pagesForReview` — up to 5 representative pages (home + richest content pages), each with an `aiView` object containing `title`, `metaDescription`, `h1`, `headings`, `schemaTypes`, `jsonLdSummary`, `textExcerpt`, `internalLinkCount`, `author`, `publishedDate`, `modifiedDate`. **Use these for Step 4.**
- `prioritizedFixes`, `worstPages`, `coverage`, `heuristicIntelligenceSignals` — supporting context. The heuristic signals are a deterministic *prior* — a sanity check, not the real evaluation.

### Step 4: Score the 6 intelligence dimensions

You are an AI agent that just found this site via web search. A user asked you a question and you landed here. **Decide: would you cite this site in your answer?**

Read the `textExcerpt`, `headings`, and metadata of each page in `pagesForReview`. Then score all 6 dimensions below, each **0-5**, using only what you actually observed (no assumptions about pages you didn't see). Write the rationale **before** the score.

#### Rubric (0-5 each)

**Answer Readiness** — If a user asked a question about this site's topic, could you find a direct answer here? The #1 factor — content answering questions in the first paragraph gets 4.8x more citations.
- 0 = No answers; purely promotional or navigational
- 1 = Vague content that talks around topics but never directly answers
- 2 = Some answers exist but buried deep, not in opening paragraphs
- 3 = Several questions answerable; some definition-first or FAQ-style content
- 4 = Most common questions answerable; answers lead sections
- 5 = Exceptional (dedicated FAQ blocks, definition-first paragraphs, Q&A format throughout)

**Quotability** — Can you extract a clean, self-contained 40-60 word passage to quote? Comparison tables get 2.8x citations; FAQ blocks +156%.
- 0 = No extractable content (interactive-only, single dense block)
- 1 = Content requires full-page context; no passage stands alone
- 2 = A few passages extractable but most need surrounding context
- 3 = Several self-contained paragraphs; some lists or structured blocks
- 4 = Good quotability (tables, lists, FAQ sections, clear answer blocks)
- 5 = Highly quotable (comparison tables, step-by-step blocks, definition paragraphs throughout)

**Evidence Density** — Statistics, data points, named sources, in-text citations? Adding in-text citations = +115% visibility; statistics = +40% citation rate.
- 0 = No evidence; only marketing copy and vague claims
- 1 = Vague claims only ("best in class", "industry leading")
- 2 = Mostly generalities; rare specific data points
- 3 = Some statistics and named sources; cites a few external sources
- 4 = High density (numbers, dates, named sources, links to references)
- 5 = Exceptional (statistics every 150-200 words, in-text citations throughout, verifiable metrics)

**Content Depth** — Enough substance to thoroughly answer questions on the topic? Long-form (2000+ words) gets 3x more citations.
- 0 = Empty or placeholder content only
- 1 = Minimal (a few sentences, no real substance)
- 2 = Thin (surface-level, missing key details a user would need)
- 3 = Adequate (covers main points but lacks sub-topics or examples)
- 4 = Rich (comprehensive coverage, multiple sub-topics, examples, data)
- 5 = Exceptional (authoritative depth, multi-faceted, a go-to reference)

**Freshness** — Current enough to cite confidently? 76% of ChatGPT's most-cited pages were updated in the last 30 days.
- 0 = No date signals; content appears abandoned or timeless-generic
- 1 = Dates present but clearly outdated (2+ years, stale references)
- 2 = Moderately dated; no "last updated" indicator
- 3 = Reasonably current OR exp

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

declared (1 observation(s))

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.