local-business-prospecting

Build and qualify local-business prospect lists by category and geography using Crawlora's Google Maps dataset, live Google Maps, Apple Maps, Yelp, and public business websites. Use when the user wants a deduplicated lead shortlist or CSV with business contacts and qualification evidence.

Produce a business shortlist with transparent selection criteria and source evidence. Researching a list does not authorize sending outreach or importing it into a CRM; perform those actions only when the user requests them.

Setup and requests

Set CRAWLORA_API_KEY to your key from crawlora.net. Run the bundled scripts/crawlora.sh from this skill's directory or by absolute path. It calls https://api.crawlora.net/api/v1 with x-api-key and prints JSON. Keep the key in the environment. Read reference/endpoints.md for the exact parameters; providers use different geography and pagination fields.

Workflow

1. Establish the business type, geography, requested count, and qualification criteria from the user's brief. If a material criterion is missing, ask; otherwise state a reasonable assumption and start with a small discovery page.
2. Search /datasets/google-map-businesses/search for bulk discovery. Use /facets?facet=category to discover exact, possibly localized category labels. Search supports page_size up to 100 and a 10,000-record result window. Radius searches require lat, lon, and radius_m (maximum 50,000 meters). Dataset counts describe this indexed corpus, not every business in the area.
3. For fresh discovery or sparse dataset results, use live Google Maps search. Apple Maps and Yelp can supplement coverage; read their individual parameters in the reference. Preserve source identities when merging results.
4. Deduplicate by provider ID first. Across providers, compare normalized name, address, domain, and phone; require consistent evidence before merging. A shared chain website alone does not make two branches the same business.
5. Refresh the most relevant candidates with live place detail. Inspect the business's own website only when contact or qualification evidence is needed. POST /web/scrape takes flat url and formats fields; POST /extract takes flat url and schema fields. Neither uses the catalog body label as a wrapper. Extract only publicly published business contact details and retain the exact source page. Do not infer email addresses or claim deliverability verification.
6. Apply the user's explicit rules, retain exclusion reasons, and stop when the requested count is met or the search bounds are reached. If scoring helps, explain the weights. A low rating is evidence of reviews, not proof of buying intent.

sh

# Start with a small indexed discovery page:
scripts/crawlora.sh /datasets/google-map-businesses/search \
  q=coffee city=Cambridge country=US page_size=10

# Live discovery; body fields are flat, not nested under mapSearchOption:
scripts/crawlora.sh -X POST /google/map/search \
  '{"keyword":"coffee cambridge massachusetts","language":"en","country":"us"}'

Output and limits

Return a table or CSV as requested. Useful columns are business_name, provider_ids, address, website, public_phone, public_email, rating, review_count, qualification_evidence, source_urls, retrieved_at, and refresh_status. Keep dataset refresh time separate from retrieval time.

• permanently_closed=false excludes known closures but still includes unknown status (null). It does not certify that every result is open.
• rating: null is unknown; do not coerce it to zero. The dataset has a minimum rating filter, not a maximum: apply a “below four stars” rule locally, excluding unknown ratings. Google's review endpoint returns a small first-page sample.
• Locationless service-area businesses can lack coordinates. A radius query cannot discover them; use text/area searches if they belong in the brief.
• Bound discovery pages and live enrichments to the requested task. Report a shortfall rather than padding the list with weak matches.
• Keep Google Maps calls at least one second apart. On 429, back off; retry a transient 5xx once. Stop on 401/403. Check application code too. Keep useful results when one source fails and mark which rows were not refreshed.