Icp Website Audit

ICP Website Audit

Crawl all sites once (deterministic, keyless), then you, the agent, score the site through each persona, run each head-to-head, and consolidate. Cost is your reasoning passes ≈ personas × (1 + competitors) — reuse the captured content; don't re-fetch.

When to use

• "Run an ICP website audit for [company]. Compare against [competitor 1] and [2]."
• Quarterly site/messaging review; after a site redesign.

How to run

Step 1 — personas (sub-skill buyer-persona-generator)

If existing_personas (personas.json) are provided, load + confirm them. Otherwise run the buyer-persona-generator skill first to produce 4-6 personas, e.g.:

python3 ../buyer-persona-generator/scripts/research_company.py \
  --url <client-url> --company "<client>" --output ${WORKSPACE}/company_bundle.json

then synthesize personas per that skill's instructions.

Step 2 — crawl all sites once (deterministic)

python3 ${SKILL_DIR}/scripts/crawl_sites.py \
  --client "Acme=https://acme.com" \
  --competitor "Rival=https://rival.com" \
  --competitor "Other=https://other.com" \
  --max-pages 8 --output ${WORKSPACE}/sites.json

Python 3 stdlib only. Captures each site's homepage + same-domain high-signal pages (pricing, product, solutions, about, case-studies, blog, docs) into one bundle. For JS-heavy / anti-bot competitor pages that come back thin, fall back to a web-automation (Playwright + Robomotion Proxy) fetch; use an apify review actor only as a hostile-site fallback for review-site presence.

Step 3 — scorecard (you, the agent)

Run each persona (from personas.json) through the client pages in sites.json, scoring First Impression, Messaging Relevance, Trust & Credibility, Clarity & Navigation, Objection Handling, Overall (1-10). Then cross-persona synthesis: consensus issues, segment gaps, messaging disconnects. Write <date>-scorecard.md.

Step 4 — head-to-head per competitor (you, the agent)

For each competitor, run each persona through both sites (reusing sites.json): per-persona quick takes, dimension scoring for both, "if I had to choose today" verdict, what to steal / what we do better. Then cross-persona competitive summary + at-risk segments. Write <date>-head-to-head-<competitor>.md each.

Step 5 — consolidate (you, the agent)

Combine absolute (scorecard) + relative (head-to-head) findings; rank the 3-5 highest-leverage moves by breadth × depth × urgency × feasibility; flag the at-risk segments (personas leaning to a competitor) prominently — they're the most actionable output.

Outputs

• icp-website-audit.md — exec summary, scorecard matrix, competitive overview matrix, persona profiles, cross-persona findings, per-competitor head-to-heads, competitive position map, at-risk segments, tiered recommendations.
• Sub-reports: persona assets, <date>-scorecard.md, <date>-head-to-head-<competitor>.md.
• All persisted to workspace + Agent Teams channel attachment.

Credentials / env

• Required: none. The crawl script is keyless; persona scoring, head-to-head, and consolidation are your job as the agent (no LLM key in the script layer).
• Optional: if APIFY_API_TOKEN is set → Apify review-site scraping when the keyless crawl is blocked; if not → keyless crawl/fetch (default). PINECONE_API_KEY / QDRANT_URL — if set → semantic persona dedup; if not → skip (default). The default keyless crawl needs no key.

Notes & edge cases

• Crawl every site once and reuse the captured content across scorecard + all head-to-heads — don't re-fetch per pass.
• Start with 1-2 competitors — each adds a full head-to-head pass per persona.
• Re-runnable quarterly — store scores per date to track improvement.
• Use Robomotion Proxy + geo for crawling volume; competitor sites may anti-bot.