Multi-Location NAP Consistency Crawler

AGENT ROLE: Autonomous NAP-integrity agent. Extract each location's Name/Address/Phone from every source, normalize, diff against the canonical record, and emit the JSON in references/output.schema.json.

OBJECTIVE

For each business location, establish a canonical NAP and detect every place (own site pages, citation directories) where it differs — even by formatting — plus flag directories where the location is missing.

INPUTS

• locations (REQUIRED): array of { id, canonical: { name, address, phone }, page_urls: string[] }.
• citation_urls (OPTIONAL): map of location_id → [directory listing URLs] to check (Yelp, BBB, Apple Maps, industry directories).
• strict_phone (OPTIONAL bool, default false): if true, formatting differences in phone count as mismatches; default normalizes to digits.

AUTHENTICATION / RUNTIME

• No API key. Keyless HTTPS GET, UA seoskills-nap/1.0, honor robots. Some directories block bots — treat blocks as unverifiable, not mismatch.

EXPECTED TOOL CALLS

• Run scripts/nap_crawl.py --locations locations.json [--citations citations.json].
• Per source URL: GET; extract NAP from LocalBusiness/PostalAddress JSON-LD first, then microdata, then visible-text heuristics.

PROCEDURE (deterministic, per location)

STEP 1 — CANONICAL: normalize the supplied canonical NAP (see normalization below). STEP 2 — For each page_url and citation_url: fetch, extract NAP (schema → microdata → regex fallback), normalize. STEP 3 — NORMALIZE: name → lowercase, strip legal suffixes/punctuation; address → USPS-style abbreviations (Street→St, Suite→Ste), collapse whitespace; phone → digits only (unless strict_phone). STEP 4 — DIFF each source's normalized NAP vs canonical; classify per field match | mismatch | missing. A source with any field mismatch = inconsistent. STEP 5 — MISSING CITATIONS: any citation_url that returns no detectable listing for the location → missing_citation. STEP 6 — SCORE consistency = matched_fields / total_checked_fields per location; EMIT inconsistencies grouped by location, each with the exact source URL and the field-level before/after.

RATE LIMITS & ERROR HANDLING

• Politeness: ≤ 3 concurrent per host, ≥ 250ms spacing. Timeout 15s.
• IF a source returns 403/429/CAPTCHA THEN mark that source unverifiable (NOT a mismatch) and continue — never assert inconsistency from a blocked fetch.
• Retry transient 5xx ≤ 2.

MISSING / INSUFFICIENT DATA

• IF no NAP can be extracted from a source THEN extraction="failed" for that source (distinct from a real mismatch).
• Formatting-only differences (e.g., "Suite 200" vs "Ste 200") normalize to a match unless the raw strings are requested; always return both normalized_match and the raw values.
• Never guess an address; only report what was extracted.

OUTPUT

One JSON object per references/output.schema.json. No prose.

FILES

• scripts/nap_crawl.py — multi-source NAP extraction, normalization, diff, scoring.
• references/output.schema.json — output contract.