Replying to Google reviews with AI

How the bot works

Eight steps. Picking up reviews, classifying them and the hard stops are ordinary code; the model is asked only to read the review and draft the words.

1. It watches the profile, or all of them. New reviews are picked up as they land, across every location you run, into one queue. A review that arrives on a Saturday is in the queue on Saturday — which is the difference between answering a customer and answering an archive.

2. It reads the review, not just the rating. What the person is actually talking about: a named member of staff, a wait, a price, a product that arrived broken, a room, a refund they are still waiting for. A five-star rating with no text is a different job from a careful paragraph, and the bot treats them differently.

3. It classifies what is at stake. Rating, topic, which location, whether the review names an employee, whether it alleges something specific, whether it contains a detail that must never be repeated in public, and whether it reads like a review of a business that isn't yours. The classification decides the route, and the rules behind it are yours.

4. It drafts the reply in your voice. Length, warmth, how you sign off, the phrases you always use and the ones you never use — learned from replies you wrote and approved, not from a generic template pool. The draft names the specific thing the review mentioned, because a reply that could have been pasted under any review is worse than no reply.

5. A complaint goes to a person, with the draft already written. Anything at or below the rating you set, anything that alleges a fact, anything naming an employee and anything privacy-sensitive is routed to a human. They get the review, the draft, and whatever account history you have given it access to — so the job is a decision, not a blank box.

6. It flags what shouldn't be answered in public at all. A review that needs a refund decision, a phone call, a safety look or a report to the platform leaves as a task with the reasoning attached, not as a reply. Some reviews are a conversation to have offline, and the bot's job there is to say so quickly.

7. A person approves, and then it posts. One queue: approve, edit, or reject with a reason. The reply is posted after that. Auto-posting is off by default; if you ever switch it on it is for the top rating with no text, and every hard stop below still holds.

8. It writes the record. One row per review: the review as it arrived, its rating and classification, the draft, who approved or edited it and what they changed, the text that was actually posted, the model and prompt version behind the draft, UTC timestamps, and what the run cost.

What you configure

It can be configured to run in any business. This is what that actually means — the list of things that are yours, not ours.

Your voice

Length, formality, whether you use the reviewer's first name, how you sign off, and the sample replies you want it to sound like. Approved by you before anything is drafted from them.

Your rating rules

Which ratings may be drafted for approval, which always go straight to a named person, and whether a rating with no text is handled differently from one with a paragraph.

Your never-say list

The phrases, offers and admissions that may not appear in a public reply — discounts, apologies that concede liability, anything about an individual employee, anything that confirms a customer's private details.

Your escalation routes

Who sees a one-star review, who sees an allegation, who sees a review naming staff, and where it lands: email, Slack, Microsoft Teams, or a task in the tool you already use.

Your locations

One profile or fifty, each with its own manager, its own approver and its own local detail. A reply about a store can mention that store's actual opening hours.

Your approval ceiling

Whether every reply is approved by a person, or only everything below a rating you choose. Off by default means everything is approved.

Your sources

Which review platforms are in scope. Google Business Profile is the common one; the same design reads the others we have access to.

Your response window

How quickly a draft should be ready, how long an unapproved draft waits before it nudges someone, and whether anything is drafted outside working hours.

The words it uses

The versioned prompts behind the drafting and the classification — yours to read, yours to tune, and replayed against reviews you have already answered before a change goes live.

Transparent, and configurable by you

The dark band below is what the system will not do. This is what you can see and change — and it is how we build everything, not something invented for this page.

You see what it did, and why

Every draft carries the rating, the classification, the reason it was routed the way it was, and the sentence in the review it is answering. You can see why it decided a review was a complaint before you agree with it.

Versioned prompts you can tune yourself

The prompts behind the drafting and the classification are yours to read and yours to tune. A change is a reviewable diff, replayed against reviews you have already answered so you can see what it would have written differently, and rolled back the same way.

A human in the loop wherever judgment actually matters

Automation handles the drafting. The decision to publish is a person's, and anything at or below the rating you set, anything alleging a fact and anything naming an employee is routed to a human rather than through one. That boundary is a property of the system, not a promise about how carefully we'll use it.

Every run logged, along with what it cost

One row per review: what arrived, how it was classified, what was drafted, who approved or edited it, what was posted and when — plus token usage and cost for that run.

It can be configured to run in any business. The voice and the rules are yours: your phrasing, your rating thresholds, your never-say list, your approvers, your locations. Review replies on this page are one configuration of it, not the limit of it.

Works with your tools

This runs on top of the accounts you already have. We're not replacing your profile or asking your customers to go anywhere else.

• Google Business Profile
• Yelp
• Facebook Pages
• Trustpilot
• Birdeye or Podium
• HubSpot
• Zendesk or Freshdesk
• Email, Slack or Microsoft Teams
• Asana, Jira or Trello
• One location or fifty
• A spreadsheet, if that's where your notes live

Works well for

• Dental practices — Patients name the hygienist and the procedure, and a reply that doesn't is worse than silence.
• Restaurants — Reviews land nightly about a dish or a server, and only a specific answer is worth posting.
• Auto repair shops — A review arguing about a price or a diagnosis needs care, so that one goes to a person first.
• Med spas — Treatment detail in a public review is a privacy problem, so some reviews get flagged rather than answered.
• Hotels — Every property has its own profile and its own tone, and one voice has to hold across all of them.

How a rollout works

Nothing is posted on day one. A dry run over reviews you have already answered comes first, and it is where you find out whether it sounds like you.

1. 01Discovery

   The quote call, then a working session. We read your profiles, the replies you have already written, the phrases you want and the ones you ban, who approves what, and what access each platform allows. You get the voice brief, the rating rules, the escalation routes and a written scope.

2. 02Dry run

   The bot drafts against reviews you have already answered. Nothing is posted. You compare its draft with the reply you actually wrote, review by review, and every difference is either a voice note, a rule to fix, or a case worth knowing about. This is where you find out whether it sounds like you, on your own reviews.

3. 03Assisted mode

   Live reviews. The bot drafts and routes; a person approves, edits or rejects every reply before it posts. The work gets much faster. The control doesn't change.

4. 04Optional top-tier auto-post

   Optional, off by default. The highest rating with no written text can post on approval-by-exception once you have watched it work for a while, with every hard stop in force and the kill switch. Plenty of teams approve everything forever, and that's a good place to stop.

What we don't claim

• No numbers we haven't earned. No response-rate figure, no rating improvement, no time saved, no “replies in seconds”. We haven't run this on your reviews. The dry run is where those numbers come from, and they'll be yours.
• No ranking claims. We don't claim that replying to reviews moves you up the map pack or improves your local visibility. Other pages do. We're not going to pretend to know that.
• No customer claims on this page. No names, no case studies, no logos.
• It won't fix a bad review. A good reply answers it honestly and takes the rest offline. Nothing here removes a review, disputes one on your behalf, or asks a customer to take one down, and no part of this is review-gating — the bot never decides who gets asked for a review.
• Fully autonomous replying isn't on offer. The design assumes a person owns what gets published. That's deliberate, and it's why the boundary is written down.

Questions people ask

Can AI respond to Google reviews?

Yes - and the honest version is that it drafts the reply and a person approves it. An AI agent can pick up every new review across your locations, read what the customer actually wrote about rather than only the star rating, classify what is at stake, and draft a reply in the voice you approved, naming the specific thing the review mentioned. What it does not do is publish a reply to an unhappy customer that nobody read. That is the one that costs you the relationship, so it goes to a person every time.

Will it post replies automatically?

Not by default, and never for a complaint. The bot drafts and a person approves, edits or rejects. If you later decide to let the highest rating with no written text post on its own, that is a choice you make per profile after you have watched it work - and every hard stop still holds: a review that alleges something, names an employee, contains a private detail or threatens legal action can never be posted on confidence alone. One flag puts the whole system back to draft-only with no deploy.

What happens with a one-star review?

It is routed to a person, with the draft already written and the review quoted. They get whatever account history you have given the bot access to, so the job in front of them is a decision rather than a blank box. The draft will not dispute the reviewer's account, admit liability, or offer a refund or a discount unless that exact offer is on a list you wrote. Where the right answer is a phone call rather than a public reply, it says so and raises a task instead.

Will the replies sound like us, or like a bot?

That is what the dry run is for. The voice comes from replies you have already written and approved - length, warmth, whether you use first names, how you sign off, the phrases you always use and the ones you never use - not from a generic template pool. Then the bot drafts against reviews you have already answered and you compare them one by one. Every difference is a voice note or a rule to fix before anything goes live.

Does Google allow AI-written replies?

Google's policies are about the content of a reply, not about what typed it: it has to be accurate, relevant and not misleading, and it must not impersonate anyone or carry prohibited content. That is exactly why this design keeps a person on anything that makes a claim, and why the bot will not state a fact you have not given it. We follow each platform's rules as they stand, you keep the account, and you can read every draft before it posts.

Can it handle more than one location?

Yes, and that is usually the reason people come looking. Every location can have its own approver, its own local detail and its own voice notes, while the hard stops and the never-say list stay the same across the group. One queue shows you everything waiting, by location, with the oldest unanswered review first.

What stops it from repeating something private a reviewer wrote?

A standing rule that sits outside the model: order values, health information, addresses, what somebody bought and who they came in with never appear in a public reply, even when the reviewer put them in theirs. The draft acknowledges the review and moves that part of the conversation to email or a phone call. A review that contains a privacy-sensitive detail is also routed to a person regardless of its rating.

Is there a record of what it did?

Yes, and it is designed for the person who has to explain a reply six months later. One immutable row per review records the review as it arrived, its rating and classification, the draft, the person who approved or edited it and what they changed, the text that was actually posted, the model and prompt version behind the draft and UTC timestamps. The intent is that someone who was not there can reconstruct any single reply from that one row.

How long does it take to get running?

The first milestone is a dry run, and it needs three things: access to your profiles, a set of replies you have already written that you are happy with, and a decision on who approves what. We will not quote you a timeline before we know which platforms are involved and what access they allow - you get a real one in your quote, once Ozzie's call has gathered the details.