GBP Review Velocity & Sentiment Monitor

AGENT ROLE: Autonomous reputation-monitoring agent. Pull reviews per location, compute velocity/sentiment/response metrics, and emit the JSON in references/output.schema.json.

OBJECTIVE

Quantify each location's review health — how fast reviews arrive, the star trend, sentiment and recurring themes, and how many reviews go unanswered — and flag deteriorations that need attention.

INPUTS

• account_id + location_ids (REQUIRED): GBP resources (accounts/{a}/locations/{l}).
• window_days (OPTIONAL, default 90): velocity + trend window.
• sentiment_backend (OPTIONAL enum lexicon|openai): default lexicon (keyless); openai if OPENAI_API_KEY set.
• benchmark (OPTIONAL): prior run's aggregates for trend deltas.

AUTHENTICATION (Google Business Profile API)

1. REQUIRE OAuth token, scope https://www.googleapis.com/auth/business.manage (GBP_OAUTH_TOKEN or delegated SA).
   • IF absent THEN STOP error.code="AUTH_MISSING_TOKEN"; 401 → STOP error.code="AUTH_EXPIRED".
2. Requires Google-approved Business Profile API access. 403 SERVICE_DISABLED/PERMISSION_DENIED → STOP error.code="API_ACCESS_NOT_APPROVED".
3. Reviews live on the v4 endpoint: GET https://mybusiness.googleapis.com/v4/{account_id}/locations/{loc}/reviews (paged via pageToken).
4. openai sentiment backend: POST https://api.openai.com/v1/chat/completions with OPENAI_API_KEY.

EXPECTED TOOL CALLS

• Run scripts/gbp_reviews.py --account accounts/1 --locations loc1,loc2 --window 90.
• Per location: page all reviews in the window; each review has starRating (ONE..FIVE), comment, createTime, and reviewReply (present iff answered).

PROCEDURE (deterministic, per location)

STEP 1 — FETCH reviews within window_days (page until createTime older than the window). STEP 2 — VELOCITY: reviews_per_week = count / (window_days/7); compare first-half vs second-half of the window to detect acceleration/deceleration. STEP 3 — STAR TREND: mean star of older half vs recent half; star_delta. STEP 4 — SENTIMENT: map starRating to polarity, refine with text sentiment (lexicon or openai); extract recurring THEMES from negative reviews (top noun phrases / lexicon categories: service, wait, price, cleanliness, staff). STEP 5 — RESPONSE RATE: answered / total; unanswered_recent = unanswered reviews in the last 14 days (highest priority). STEP 6 — FLAGS: velocity_drop (recent < 0.5× earlier), sentiment_decline (star_delta <= -0.4), low_response_rate (< 0.5). EMIT per-location metrics + flags, sorted worst-first.

RATE LIMITS & ERROR HANDLING

• GBP v4 strict QPM. 429 → backoff 2^attempt (max 5) then STOP error.code="RATE_LIMITED" with partial.
• openai sentiment 429 → backoff (max 4); on persistent failure fall back to lexicon and set sentiment_backend="lexicon".
• Per-location failure → mark status="fetch_error", continue.

MISSING / INSUFFICIENT DATA

• IF a location has < 5 reviews in the window THEN report raw counts but mark trend metrics low_confidence (too few to trend).
• Star rating is authoritative for polarity; if a review has no comment, sentiment = star-derived only (note text_available=false).
• Never invent themes; if no negative reviews, themes = [].

OUTPUT

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

FILES

• scripts/gbp_reviews.py — GBP v4 reviews client, velocity/star/response metrics, sentiment + theme extraction.
• references/output.schema.json — output contract.