competitor-deep-dive

name

competitor-deep-dive

description

Deep-dive competitor analysis report: given a company name or website, researches their ads (Meta/TikTok/Google ad libraries), SEO, GTM strategy, product/pricing/reviews, and synthesizes recommendations — deployed as a live tabbed report on here.now with a per-tab CSV download. Use when asked to "analyze competitor X", "do a competitor deep dive", "research [company] as a competitor", or "build a competitor report/analysis".

Competitor Deep Dive

Produces one shareable report: a tabbed static site (Overview, Product Analysis, GTM, Ads, SEO, Recommendations), each data tab backed by a downloadable CSV, deployed to here.now.

Input: a company name or a website URL. If only a name is given, resolve the canonical website first (web search); if only a URL is given, resolve the company/brand name from the site itself (used to search the ad libraries, which key on advertiser name, not domain).

Scope discipline: cap ad results at 10 per platform (30 rows max across Meta + TikTok + Google). This is a competitive-intelligence snapshot, not an exhaustive archive — do not silently expand scope.

Sequential execution (one long agent turn, no Workflow orchestration by default): sections build roughly in this order because later ones reference earlier findings.

Setup (once per machine, not per run): cd ~/.claude/skills/competitor-deep-dive && npm install && npx playwright install --with-deps chromium. Scripts live in scripts/; symlink or point node_modules there from any scratch working directory if running scripts outside the skill folder itself.

1. Resolve identity

• Canonical website (follow redirects, prefer the marketing site over app./login. subdomains)
• Legal/advertiser name if it differs from the brand (ad libraries often require this — e.g. "Wise" → "WISE PAYMENTS LIMITED"). Check ad-library autocomplete/search suggestions to confirm the right entity before pulling ads.

2. Product Analysis tab

• Homepage: value prop headline, 3-5 headline features, hero screenshot if easy to grab
• Pricing: find /pricing (or nearest equivalent — /plans, footer link); capture tiers, prices, billing period, and whether it's self-serve or "contact sales"
• Reviews — always check these three, every run, in this order:
  1. Google Business Profile (google.com/maps/search/<company>) — real headless browser, domcontentloaded + a few seconds' wait. The star rating renders even in the signed-out "limited view"; the exact review count is gated behind sign-in — report the rating, note the count as unavailable rather than guess it.
  2. App stores (iOS App Store ?see-all=reviews&platform=iphone, Google Play play.google.com/store/apps/details?id=...) — use a real Playwright browser, not WebFetch. WebFetch cannot execute JS, and both of these are client-rendered — it will report empty/truncated content and look like the site is bot-gated when it isn't. A plain chromium.launch({headless:true}) with a desktop UA and waitUntil:'domcontentloaded' (not 'networkidle' — App Store never fully idles) gets real star ratings, review counts, and review text with zero friction. If Play Store shows no rating widget at all, that itself is a real, reportable finding (not every listing has one).
  3. G2 / Capterra / Trustpilot — start with a web search summary ("<company>" reviews site:g2.com OR site:capterra.com OR site:trustpilot.com); if a real rating doesn't surface that way, try the same real-browser approach as the app stores before concluding there's no listing — don't assume bot-gating without testing it.
  • Say plainly when a source has no listing or no visible rating — never estimate a count or score that isn't explicitly shown.
  • Synthesize overall sentiment + 3-5 notable quotes with source links once the above is gathered.
• Feeds content.json's product block (see step 10) — CSV is built automatically by build-report.js from content.product.csvRows: source, type(feature/pricing/review), text, rating(if any), url

3. GTM tab

Four signals, inferred from the site itself — do not re-collect data that belongs in Ads/SEO, just reference it qualitatively once those tabs exist:

• Sales motion: self-serve signup vs. demo-gated — read the primary CTA on homepage/pricing
• ICP / target segment: who the copy, case studies, and customer logos are speaking to
• Positioning & messaging: the core value-prop headline + 2-3 differentiation claims
• Channel mix signal: one line synthesizing what Ads/SEO show once those tabs are built (e.g. "TikTok-heavy, UGC-style creative; thin organic footprint")

Feeds content.json's gtm block — CSV built automatically from content.gtm.csvRows: signal, observation, source_url

4. Ads tab

Use the three scraper scripts — don't reimplement this by hand, every platform-specific gotcha (bot-gating workarounds, TikTok's mandatory country dropdown + autocomplete-only search, Google's GetCreativeById trick, retry logic) is already encoded in them:

node scripts/scrape-meta-ads.js "<exact advertiser name>" <country-code> 10 <data-dir>
node scripts/scrape-google-ads.js <domain> 10 <data-dir>
node scripts/scrape-tiktok-ads.js "<advertiser name>" 10 <data-dir>

• Meta: the advertiser name must match the "Sponsored" byline exactly (often the legal entity, not the brand — e.g. "WISE PAYMENTS LIMITED" not "Wise"). If unsure, run a keyword search first (the script will report 0 matches if the name is wrong) and read a real hit's byline before re-running with the exact string.
• Google: search by domain, not advertiser name — far more reliable, and resolves the legal entity automatically.
• TikTok: if the advertiser name has no autocomplete match, the script falls back to a plain keyword search to get a definitive "Total ads: 0" confirmation before reporting no presence — don't conclude absence from autocomplete alone.
• Each script downloads media locally (images/video) rather than linking platform CDN URLs directly — those are signed and expire in days; self-hosting keeps the report durable for its here.now lifetime.
• Outputs land in <data-dir>/{meta,google,tiktok}_ads.json and <data-dir>/{meta,google,tiktok}_media/ — build-report.js (step 10) reads these directly.

CSV (ads.csv, built automatically): platform, creative_id_or_url, format, headline_or_copy, media_filename, first_shown, last_shown, reach(TikTok only), library_url

5. SEO tab

• Fetch /sitemap.xml (or /sitemap_index.xml); count URLs, bucket by path segment (blog/, product pages, docs/, etc.)
• Homepage <title>, meta description, H1s, and (rare but check) meta keywords
• "Top keywords" here means on-page keyword signals — repeated terms in title/meta/headings/body copy — not real search-ranking data. State this scoping explicitly in the report; this skill has no access to an actual rank-tracking tool (Ahrefs/SEMrush/etc.), so never imply these numbers reflect real search volume or rank.

Feeds content.json's seo block — CSV built automatically from content.seo.csvRows: keyword_or_phrase, source(title/meta/h1/body), count_or_context

6. Recommendations tab (narrative, no CSV)

Synthesize across every tab above into concrete "where we can win" opportunities — gaps in their ads (platforms/formats they're not using), SEO (thin sections, missing keywords), product (pricing gaps, review complaints), and GTM (underserved segments). Ground every recommendation in a specific finding from an earlier tab, not generic advice.

7. Overview tab (narrative, no CSV, built last)

One-paragraph company summary + one key snippet pulled from each of the 5 tabs above (e.g. "5,000 active TikTok ads", "self-serve PLG motion", "142 indexed URLs, thin blog presence", "4.3★ average sentiment across review sites", "top opportunity: no Google Search presence").

Also always research and include, each with its own hyperlinked source (web search — Crunchbase, LinkedIn, press coverage, the company's own About/press page are the usual finds):

• Founders (names; link to the source that names them)
• Year founded
• Estimated employee count (LinkedIn's "N employees" figure is usually the most reliable single source)

If any of the three genuinely can't be found after a real search, say so explicitly in that spot — never estimate or guess a plausible-sounding number.

8. Sourcing discipline (applies to every tab, not just Overview)

Every factual claim that came from a specific external page — reviews, founder/employee info, academic or press citations, ad-library entries — gets a real <a href> hyperlink to that source directly in the report prose/table, not just buried in a CSV column. If you can name where a fact came from, link it; if you can't name a specific source, don't state it as fact.

9. Design

Neobrutalist by default: bold 2-3px black borders, hard offset box-shadows (no blur — 5px 5px 0 #000 pattern), flat vibrant accent colors (no gradients), a subtle dotted background texture, chunky "pressed" tab-switcher buttons, numbered/badged callout cards. This is baked into scripts/build-report.js's HTML template — don't reinvent it per run; only edit that script's CSS if the user asks for a different look (and if they do, that's a permanent change to make there, not a one-off tweak in a throwaway file).

10. Build and deploy

1. Write <data-dir>/content.json — the analyst-authored narrative and table content (Overview paragraph + facts + snippets, Product positioning/pricing/reviews, GTM signal rows, SEO on-page/sitemap rows, Recommendations). See scripts/content.example.json for the exact required shape — copy it and fill in real findings, don't improvise the schema from memory.
2. Run node scripts/build-report.js <data-dir> <dist-dir> — this reads content.json plus the three *_ads.json scraper outputs, copies media into <dist-dir>/media/{meta,google,tiktok}/, and writes index.html + all four CSVs. Don't hand-write the HTML/CSVs — this script is the single source of truth for both content assembly and the neobrutalist template together.
3. Publish via the here-now skill (bash ~/.claude/skills/here-now/scripts/publish.sh <dist-dir> --title "<Company> Competitor Analysis")
4. Verify before reporting done — load the published URL with a real headless browser, check for JS console errors, confirm ad media/videos actually loaded (readyState/naturalWidth), and click through at least one tab switch. Don't just trust that the build script ran without throwing.
5. Report the live URL, the claim URL (24h anonymous expiry — mention this explicitly), and a one-line summary of what's in each tab.
6. After every user-requested change to the report (design, new fields, new sourcing rules, anything), update the relevant script(s) and/or this SKILL.md in the same turn so the next run starts from the improved baseline — don't let fixes live only in one report's throwaway build script.