What does the review responder do?

A small serverless system that watches every new review across Google Business Profile, Facebook, and Yelp; reads each one with three small extractors (rating with sentiment cross-check, themes, and specifics like names and dishes); picks one of four moves — auto-reply, draft, escalate, or ignore — and writes replies in your voice from your own Drive policies.

How much does it cost to run?

About $3/month at typical small-business review volume (around 50 reviews/month). The fixed cost is essentially zero — quiet weeks bill nothing. The variable cost is pennies per review, dominated by Bedrock tokens for the three extractors and the composer. At 500 reviews a month it lands under $15 total.

Which review platforms are supported?

Google Business Profile (push via Pub/Sub, full read/reply via the Business Profile API), Facebook page recommendations (read-only via a third-party aggregator like Birdeye, Yext, or ReviewTrackers — Meta deprecated their own page recommendations webhook in Graph API v22.0 in January 2025), and Yelp (hourly poll of the Fusion API; reply API is enterprise-partner-gated, so Yelp is draft-only).

How does it avoid replying to a 1-star review with a generic thank-you?

Auto-reply is reserved for 4- or 5-star reviews with no specifics, no named staff, no themes flagged for review, and high confidence on all three extractors. Anything 1- or 2-star, anything that names a staff member, anything that hits a safety or legal keyword, and anything the responder isn't confident about lands in your inbox or Slack as a draft or escalation — never as an auto-reply.

How does the reply stay in our voice?

Three short Drive docs ground every reply: a voice file (tone, opener templates, signature, length), a policies file (refund window, replacement policy, contact phone, hours, staff roster), and a never-say file (banned phrases, comparative claims, legal/medical language). Four guardrails sit between the model and the post button: citation required, no personal-info echo, no fabricated specifics, and no off-policy promises. Edit a Drive doc and the system picks up the change on the next refresh.

What happens with a 1-star or safety-flagged review?

Safety and legal keywords (food poisoning, injury, lawyer, refund-stalled, named-staff complaint) bypass scoring and escalate immediately, regardless of star rating. The responder doesn't try to compose a public reply at all; it drafts a single-sentence acknowledgement ("We're sorry, we'd like to look into this — please reply with your visit details or call the number on your receipt") and sends the whole package — original review, proposed acknowledgement, matched policy excerpt — to the manager via SNS with optional SMS.

Which AWS services does it use?

Lambda (with Function URLs for webhooks), SQS, DynamoDB on-demand, S3, EventBridge (hourly Yelp poll, weekly themes roll-up), SNS, SES outbound, Secrets Manager, CloudWatch Logs (7-day retention), AWS Budgets, and Bedrock (Claude Haiku 4.5 via Global cross-Region inference, plus Titan Text Embeddings v2 with a Bedrock Knowledge Base backed by Amazon S3 Vectors). No API Gateway, no NAT Gateway, no always-on compute.

Series · 7 parts Published May 4, 2026

Review responder

A serverless responder that watches every new review across Google, Facebook, and Yelp, drafts replies in your voice from your own policies, posts the safe ones automatically, and quietly hands you the rest with everything you need to respond in one sitting. Seven posts on the same system — one diagram at a time — with an engineering reference at the end.

01

A review responder on AWS for a few dollars a month

The whole system on one page — a review intake, a responder, and a dispatch-and-learning piece, plus the four moves they share for every review.
02

How a review reaches the responder

Three lanes at the door: Google Business Profile push, Facebook page reviews push, Yelp poll. Dedupe and screen happen before any AI runs.
03

How the responder reads a review

Three small extractors run in parallel — rating with sentiment cross-check, themes, and specifics like names and dishes — each with a confidence score the move-picker reads before it decides anything.
04

How the responder picks a move

Four moves, one pick per review: auto-reply, draft, escalate, ignore. Safety keywords bypass everything; named staff and specifics become drafts.
05

How a reply stays in your voice

Three Drive docs ground the composer; four guardrails sit between the model and the post button — citation, no PII echo, no fabricated specifics, no off-policy promises.
06

What the review responder costs

A coffee a month at SMB review volume. Pennies per review, dominated by Bedrock tokens for the extractors and the composer.
07

Engineering reference: the review responder architecture

Same system, drawn purely for engineers. Service names, resource identifiers, region, Bedrock model IDs, Knowledge Base wiring, Google Business Profile and Meta Graph API specifics.

What does the review responder do?: A small serverless system that watches every new review across Google Business Profile, Facebook, and Yelp; reads each one with three small extractors (rating with sentiment cross-check, themes, and specifics like names and dishes); picks one of four moves — auto-reply, draft, escalate, or ignore — and writes replies in your voice from your own Drive policies.
How much does it cost to run?: About $3/month at typical small-business review volume (around 50 reviews/month). The fixed cost is essentially zero — quiet weeks bill nothing. The variable cost is pennies per review, dominated by Bedrock tokens for the three extractors and the composer. At 500 reviews a month it lands under $15 total.
Which review platforms are supported?: Google Business Profile (push via Pub/Sub, full read/reply via the Business Profile API), Facebook page recommendations (read-only via a third-party aggregator like Birdeye, Yext, or ReviewTrackers — Meta deprecated their own page recommendations webhook in Graph API v22.0 in January 2025), and Yelp (hourly poll of the Fusion API; reply API is enterprise-partner-gated, so Yelp is draft-only).
How does it avoid replying to a 1-star review with a generic thank-you?: Auto-reply is reserved for 4- or 5-star reviews with no specifics, no named staff, no themes flagged for review, and high confidence on all three extractors. Anything 1- or 2-star, anything that names a staff member, anything that hits a safety or legal keyword, and anything the responder isn’t confident about lands in your inbox or Slack as a draft or escalation — never as an auto-reply.
How does the reply stay in our voice?: Three short Drive docs ground every reply: a voice file (tone, opener templates, signature, length), a policies file (refund window, replacement policy, contact phone, hours, staff roster), and a never-say file (banned phrases, comparative claims, legal/medical language). Four guardrails sit between the model and the post button: citation required, no personal-info echo, no fabricated specifics, and no off-policy promises. Edit a Drive doc and the system picks up the change on the next refresh.
What happens with a 1-star or safety-flagged review?: Safety and legal keywords (food poisoning, injury, lawyer, refund-stalled, named-staff complaint) bypass scoring and escalate immediately, regardless of star rating. The responder doesn’t try to compose a public reply at all; it drafts a single-sentence acknowledgement (“We’re sorry, we’d like to look into this — please reply with your visit details or call the number on your receipt”) and sends the whole package — original review, proposed acknowledgement, matched policy excerpt — to the manager via SNS with optional SMS.
Which AWS services does it use?: Lambda (with Function URLs for webhooks), SQS, DynamoDB on-demand, S3, EventBridge (hourly Yelp poll, weekly themes roll-up), SNS, SES outbound, Secrets Manager, CloudWatch Logs with seven-day retention, AWS Budgets, and Bedrock (Claude Haiku 4.5 via Global cross-Region inference, plus Titan Text Embeddings v2 with a Bedrock Knowledge Base backed by Amazon S3 Vectors). No API Gateway, no NAT Gateway, no always-on compute.

All posts

Frequently asked questions

Other series