Skip to content

examples.md

Latest commit

 

History

History
151 lines (117 loc) · 6.63 KB

examples.md

File metadata and controls

151 lines (117 loc) · 6.63 KB
Raw

Examples — real-world workflow

The README shows one-liners ("analyze Nike"). This page is the repeatable procedure the server was actually built for: map who advertises in a niche, read their copy, see which creatives scale, then jump to their websites to understand how they close and what backend they run. No login, no token — the two tools scrape the public Ad Library web app, so they work for commercial ads in any country. (Meta's official ads_archive API only returns political/issue ads worldwide plus all ad types for the EU/UK, which is why this server doesn't use it.)

Tool What it does
search_ad_library keyword search of the Ad Library for a country
scrape_ad_library_url scrape a specific Ad Library URL you already have

Both render the SPA with crawl4ai + headless Chromium and parse the ad cards.

Phase 1 — Discovery (search_ad_library)

1. Pick 3–5 queries

The terms a prospect would actually search, or that a competitor would put in their own copy. Example, "AI automation consulting" in Mexico:

"implementación de IA en tu negocio"
"automatización con inteligencia artificial"
"IA para empresas"
"consultoría inteligencia artificial"

2. Run each query

search_ad_library(
    query="implementación de IA en tu negocio",
    country="MX",
    active_status="active",
    wait_seconds=12,      # raise if results come back empty — the SPA is slow to hydrate
    scroll_rounds=8,      # infinite-scroll cycles; 0 = first page only, 15+ = deep sweep
)

3. What you get back, per ad

field meaning
advertiser, advertiser_handle name + numeric Page id (feed the id back in as advertiser_page_id)
library_id, ad_details_url stable id + deep link to the ad's detail view
started_running first seen date — months old = validated copy, not an experiment
ads_using_creative how many active ads share this exact text — proxy for budget/scale
landing_url, landing_domain where the click goes
cta the button label (Learn more, Sign up, Send message, Shop now, ...)
link_text the headline/caption strip under the creative
body full ad copy

4. Read the results

  • Group by advertiser, sort by frequency → who dominates the niche.
  • High ads_using_creative → that advertiser is scaling; their copy is the benchmark.
  • landing_domain:
    • fb.me / api.whatsapp.com / wa.me → funnel goes straight to a DM, no website. Usually a solo operator or a CRM reseller. Nothing is measured or pre-qualified.
    • a real domain → they run a landing page; worth the Phase 2 teardown.
  • Read the body: does it sell the tool or the outcome? Generic ("automate your processes") or a concrete pain ("double data entry between your ERP and CRM")? One vertical or "companies with 30–50 employees" in general?

Pull every ad from one advertiser

Once you have a Page id, drop the keyword and target their "all ads" view:

search_ad_library(advertiser_page_id="100094954977054", country="MX", scroll_rounds=12)

The Page view is more heavily client-rendered and does not always hydrate under the headless browser. If it comes back empty, fall back to a keyword search of the advertiser's name — that path is reliable.

Phase 2 — Website teardown (outside this MCP)

For each competitor that has its own domain:

1. Resolve the real domain

The ad's landing_url is often a useless fb.me redirect. Open their Facebook Page (facebook.com/<advertiser_handle>) and read the "About" link, or try candidate domains.

2. Fingerprint the stack with curl

curl -sI -L -A "Mozilla/5.0" https://DOMAIN/          # server, x-powered-by, CDN, security headers

curl -s  -L -A "Mozilla/5.0" https://DOMAIN/ | grep -oiE \
  "wp-content|/_next/|__NEXT_DATA__|framer|webflow|lovable|_l5e|leadconnector|msgsndr|\
   gohighlevel|calendly|cal\.com|hubspot|hsforms|typeform|tally\.so|supabase|\
   G-[A-Z0-9]{9,11}|GTM-[A-Z0-9]{6,8}|AW-[0-9]{9,}|fbq\('init'|wa\.me/[0-9]+" | sort -u
marker means
leadconnector / msgsndr GoHighLevel — all-in-one CRM + booking + funnel + SMS
hubspot / hsforms HubSpot CRM
cal.com / calendly dedicated scheduling
lovable / _l5e / lovable-uploads built on Lovable (no-code); pair with supabase for its backend
supabase Supabase backend (forms, auth, DB)
wp-content WordPress · /_next/
AW- they also run Google Ads, not just Meta
G-... / GTM-... GA4 / Google Tag Manager

3. Read the funnel with a fetch/summary tool

Hero copy, the offer, where the CTA goes (form / WhatsApp / calendar), stated implementation time, pricing, and the type of social proof (concrete numbers vs. vague vs. placeholder logos).

4. Open the conversion page separately

/demo, /book, /prueba, /getdemo — this is where you see whether booking is a native calendar, a form to a CRM, or just a WhatsApp link.

5. SPA gotcha

Sites that are 100% client-rendered (Vite/React with no SSR) return almost nothing to curl. Use a real browser automation tool and read the network requests — that reveals embedded booking widgets, chat widgets, which pixels fire, and form endpoints (e.g. a supabase.co XHR tells you the backend without any server-rendered HTML).

Phase 3 — Output & cadence

  • Keep a running doc of advertisers, their angle, and who is scaling creatives.
  • Keep a teardown table: stack / funnel / close, one row per competitor.
  • Re-run Phase 1 every 2–4 weeks and diff advertisers / copy / ads_using_creative.

Notes & limits

  • 403 is expected. Facebook flags the headless browser and returns HTTP 403, but usually still serves the rendered ad cards — the tools judge success by parsed content, not status code. If a call comes back with success: false and no ads, you are likely being rate-limited: wait a few minutes, raise wait_seconds, or fall back to a real browser.
  • scroll_rounds trades time for completeness. Each round is a scroll-to-bottom plus a 2.5s wait. 0 returns just the first paint (~20–26 cards, fast); 8 (default) is a good balance; 15+ for a full historical sweep of a busy query.
  • The scraper parses the public SPA markup, so a Facebook layout change can break field extraction. raw_markdown is always returned so you can re-parse by hand if needed.
  • No spend or impression numbers — Meta only publishes those for political ads. This is creative / cadence / landing-page intelligence, not a budget estimator.