Live in Production
2024
Responding to Google reviews manually is a constant, low-skill drain on business owners. Reviews arrive unpredictably across dozens of locations. A negative review left unanswered for 24 hours signals indifference to every prospective customer who reads it. A positive review without a reply is a missed loyalty-building moment.
The typical manual process — monitoring the dashboard, judging tone, drafting a response, getting approval, posting — takes 10–20 minutes per review and happens inconsistently. Businesses with high review volumes simply give up and leave most reviews unanswered.
The challenge was building a system that could match the quality of a skilled human response writer, at scale, without sacrificing brand voice or producing generic outputs that customers immediately recognise as automated.
I built a multi-agent pipeline on n8n where each agent owns a discrete responsibility: one classifies, three specialised agents write (positive, negative, neutral), and a revision agent quality-checks before anything is published. Separating concerns across agents rather than using a single monolithic prompt produces consistently higher-quality, more contextually appropriate responses.
A webhook triggers whenever a new review is posted to any connected Google My Business listing. The review text, star rating, reviewer name, and listing details are extracted and passed to the classifier agent.
The Classifier Agent uses GPT-4o to detect the review's sentiment (positive, negative, or neutral) and extract the key themes — what the customer praised, what they complained about, and whether there are any specific service or staff mentions that the response should acknowledge. This structured output routes the review to the correct specialist.
Three dedicated writing agents handle their respective sentiment categories. The Positive Agent crafts warm, specific thank-you responses that encourage return visits. The Negative Agent produces empathetic replies that acknowledge the issue, take responsibility where appropriate, and offer a resolution path. The Neutral Agent bridges both tones. Each agent is prompted with the business's brand voice guidelines.
Before anything is published, the Revision Agent reviews the draft for length (responses over 150 words are trimmed), off-brand language, hollow filler phrases, and any factual overreach. This layer prevents the LLM inconsistencies that make automated responses obvious — overly formal apologies, repeated sentence structures, or generic phrases that no real person would write.
The approved response is posted directly to Google My Business via the API. Every review — the original text, sentiment classification, draft response, revision notes, and post timestamp — is logged to Google Sheets for auditing and periodic quality review by the business owner.
