Automate the watching and the first drafts, not the answering: review alerts, a log, a monthly themes summary and AI-drafted thank-yous that you approve. Answer every complaint yourself, meaning any one-, two- or three-star review and anything about allergies, illness, injury, staff or money. A generic reply to an upset diner is itself a public mistake.
Two developments make full automation tempting. Google has been testing AI-suggested replies inside Business Profile for some accounts since March 2026, and the reports I found disagree about whether the owner approves each one. Birdeye and Podium, two review platforms aimed at local businesses, sell on quotes rather than published prices. Neither changes who should write the reply to a bad review.
Sort reviews by what a bad reply could cost
Star ratings are a rough sorter. A four-star review can hide an allergy story, and a one-star review can be about something you never controlled, such as a delivery app's lukewarm bag. A better question is what a poor reply would cost, because the people who read it are mostly the next diners, not the reviewer. That gives seven piles, and only the first three let AI draft anything.
| Review type | Who writes the reply | What AI does |
|---|---|---|
| Five stars, no text or one line | AI drafts, you approve in a weekly batch | Writes a one-line thank-you that claims nothing |
| Four or five stars with detail (a dish, a team member, an occasion) | AI drafts, you edit | Echoes the detail the reviewer gave; you check names and facts |
| Three stars, or "lovely, but..." | You, starting from a draft | Drafts the shell; you rewrite the middle |
| One or two stars with a specific complaint | You, from scratch | Nothing public |
| Allergy, illness, injury, discrimination or staff conduct | You, after speaking to the guest and writing an incident note | Nothing public |
| Looks fake, extortion, competitor, wrong venue | Report it first; a short factual reply at most | Helps assemble the evidence for the report |
| Written in another language | You decide; AI translates both ways | Translates the review and your reply |
Three-star reviews are where AI drafts go wrong most quietly. The review holds praise and a grievance, and a draft will happily thank the guest for both. Treat the draft as a shell and rewrite the middle two sentences yourself.
On the fake pile, check the booking system and the till for that evening before you decide a review is fake, since a guest may have booked under another name. The reporting route, and the single appeal you get per reported review, are covered in how to handle a fake or unfair review.
One test for any draft before it goes live: read it aloud as though the reviewer's table were listening. A line that makes you wince was written by the tool, not by you.
What is safe to automate: alerts, a log, themes and drafts
Alerts that reach someone who will act
Most review sites have a notification setting for new reviews. Switch it on for Google Business Profile, Tripadvisor, Yelp and Facebook, whichever you use, and point them all at one shared mailbox that the owner and the manager can both see rather than one person's private inbox. Menus and labels move around, so look for notifications in each profile's settings instead of following a screenshot. Then add a single rule: a one-, two- or three-star alert is read the same day. The tool adds nothing clever here. What you gain is that nobody finds the one-star review three weeks late.
A review log that takes under a minute a row
A spreadsheet with nine columns does most of the work: date, site, stars, reviewer's display name, theme tags, reply status, who replied, private follow-up (yes or no) and notes. It shows at a glance what has not been answered, and it is the dataset you hand to an AI tool each month. Fix the tags early. Choose six to eight, such as service speed, food quality, noise, booking, value, staff, cleanliness and allergens, and keep them for a year, because a monthly summary is only comparable if the labels do not drift.
Themes once a month, counted by you
Paste the month's reviews into ChatGPT, Claude or Gemini and ask for themes, but make the tool show its working. Asking for an exact quote and the log row number against every count lets you check each claim in seconds. This shape of prompt works:
Here are 38 restaurant reviews from the last month, one per row, each with a row number.
Group them into themes. For each theme give the name, how many rows mention it, whether the tone is positive, negative or mixed, and the row numbers with a short exact quote from each.
Count a review once per theme. Do not infer problems that are not written down.
Finish with two things a manager could check next week.An output of the kind to expect follows. The figures are made up for illustration.
- Slow service between roughly 7.30 and 8.30pm: 6 rows (4, 9, 12, 19, 27, 33), negative or mixed.
- Praise for the lamb shoulder and the sticky toffee pudding: 9 rows, positive.
- Noise on Friday and Saturday nights: 5 rows, mixed.
- Named team members praised: 5 rows, positive.
- Booking confusion by phone: 2 rows, negative.
What I would fix: the model counted noise five times, but two of those reviews called the room lively as praise, so the honest count is three. Its suggestion to hire more servers rests on six reviews and nothing about the rota, so I would rewrite it as "check who was on the floor between 7.30 and 8.30pm on the evenings named". For the fuller method, including what to do with a hundred reviews at once, see analysing restaurant reviews with AI.
A 45-seat restaurant's month of reviews, with the minutes
Picture an independent 45-seat restaurant that collects 38 public reviews in a month, 30 on Google and 8 on Tripadvisor. Sorted by the table above, that is 21 short five-star reviews, 8 detailed positive ones, 4 mixed ones and 5 complaints, one of which describes an allergic reaction. The minutes below are my assumptions for an owner who writes carefully. Time yourself for a fortnight before trusting them.
| Pile | Reviews | Minutes each, all by hand | Minutes each, with the triage |
|---|---|---|---|
| Short positives | 21 | 4 | 1 (approve a draft) |
| Detailed positives | 8 | 6 | 3 (check the detail, tweak) |
| Mixed | 4 | 10 | 7 (rewrite the middle) |
| Complaints | 4 | 20 | 20 (no saving; you write it) |
| Allergy complaint | 1 | 40 | 40 (no saving) |
| Replies, total minutes | 38 | 292 | 193 |
The triage also adds a five-minute weekly check of the log (20 minutes a month) and a 25-minute themes read, 45 minutes in all. So the fair comparison is 238 minutes against 292, a saving of 54. Here is how that sits beside the other ways an owner might handle the same month.
| Approach | Replies posted | Owner minutes a month | Complaints written by the owner |
|---|---|---|---|
| Every reply written by hand | 38 | 292 | 5 of 5 |
| Complaints, mixed reviews and five detailed positives only (a common habit) | 14 | 190 | 5 of 5 |
| Triage: automate the watching and drafts, owner writes complaints | 38 | 238 | 5 of 5 |
| Auto-post AI replies to everything | 38 | close to 0 | 0 of 5 |
Do not buy this for the hours. The triage saves 54 minutes against writing everything by hand and costs 48 more than the habit of answering a selection. What you buy is coverage, 38 of 38 answered, and a same-day view of every complaint. The last row is the cheapest in minutes and the only one where the allergy review is answered with "Thank you for your feedback!"
A mixed review drafted by AI: what comes back and what to fix
Take a four-star review that praises the food and grumbles about timing (invented for illustration):
Lovely slow-cooked lamb and the sticky toffee pudding was the best part. Service was a bit slow around 8pm but the staff were kind.
A prompt that says only "reply to this review" returns something like "Thank you so much for your wonderful review! We're thrilled you enjoyed your visit", which could sit under any restaurant's page. A boxed-in prompt does better:
You draft public replies to Google reviews for the owner of a 45-seat [cuisine] restaurant.
Rules:
- 40 to 70 words, warm and plain, no exclamation marks.
- Use one specific detail the reviewer gave (a dish, a team member, an occasion). Never add details they did not give.
- Never explain why something went wrong unless I tell you the reason. If I have not, say I will look into it.
- Never promise changes, refunds or compensation.
- Do not use: thrilled, delighted, valued customer, we take all feedback seriously.
- Sign off with [owner first name] only.
Review (4 stars): "Lovely slow-cooked lamb and the sticky toffee pudding was the best part. Service was a bit slow around 8pm but the staff were kind."
What I know: nothing yet about the service that evening.Here is the kind of output that comes back (illustrative):
Thank you for the kind words about the lamb and the pudding; I have passed them on to the kitchen. I am sorry service slowed around 8pm. Weekend evenings are our busiest time and we are looking at how we staff them. I hope we see you again soon. [owner first name]
It reads well and it still needs three fixes. "Weekend evenings" is invented, because the review never said which night. "We are looking at how we staff them" is a commitment you have not made, and it quietly admits a staffing problem. "I have passed them on" is only true if you do. After the repair it becomes:
Thank you for the kind words about the lamb and the pudding; I'll pass them on to the kitchen. I'm sorry service slowed around 8pm, and I'll look at what was happening on the floor at that time. I hope we see you again soon. [owner first name]
That took about three minutes, which is where the "7 minutes for a mixed review" in the table came from once you add checking the till for that evening. The pattern to notice is that the tool filled a gap in the facts with a plausible guess. Every rule in the prompt exists because of a guess like that.
A two-star complaint: the tool's draft and the reply you'd send
Now a review that should never be left to a draft (invented for illustration):
Waited 50 minutes for mains and the chicken was lukewarm. Nobody apologised. Won't be back.
An unboxed draft of the kind many tools produce:
Thank you for your feedback. We take all feedback seriously and strive to provide excellent service to every guest. We are sorry your experience did not meet your expectations and have shared your comments with the team. Please contact us so we can make things right.
Three failures show up straight away. It never mentions the 50 minutes or the lukewarm plate, so every reader can see it is a form letter. "Shared with the team" may not be true. "Make things right" invites a demand with nobody assigned to answer it, and it gives no name or direct route. Here is what an owner might send after checking what happened that evening:
I'm sorry, [first name]. Fifty minutes for mains and a lukewarm plate is not how we want anyone's evening to go, and not being apologised to at the table made it worse. I've looked into that evening and spoken to the team. I'd like to hear more and put it right: please email [restaurant email address] or ring [number] and ask for [owner first name].
The reply only works if the sentences are true, so the owner has to do the looking first: the kitchen timings, who was on the floor, what the team remembers. That is the part no tool can do, and it is why this pile stays with you. Once your own version exists, AI is useful as a second reader. Paste it in and ask whether it sounds defensive and what a stranger would take from it. Read the answer, then decide. The fuller playbook is in how to answer restaurant complaints with AI without escalating.
Allergy and illness reviews: the first hour
- Do not reply publicly yet. Read the review twice and write down exactly what it alleges.
- Pull the evidence: the booking, the order ticket and any allergy note, the dishes served, who was on the floor and in the kitchen.
- Write an incident note while memories are fresh: time, table, items, what was asked and what was said. Keep it factual and dated.
- Try to reach the guest through the booking details, offering a conversation rather than a defence.
- If anyone alleges harm, tell your insurer or ask a solicitor before posting anything.
- Only then post a short public reply: you have read it, you want to understand what happened, here is a direct contact. No account of your allergy procedures, no argument over who said what, and no booking details the reviewer did not give.
AI has no public role in any of that. It can help tidy an internal timeline from your own notes, provided guests' names and health details stay out of the chat. The same caution applies earlier in the chain, when a guest asks about allergens before booking, a question I would route to a person who knows the kitchen.
Asking for reviews without breaking Google's rules
Reputation tools love to automate the ask, so know the limits before you switch that on. As of October 2026, Google's review policy rules out incentives for reviews, review gating (asking only the guests you expect to be happy), staff quotas for reviews, and asking for reviews that mention specific things such as a team member's name. Google also added review-extortion detection and a report form in April 2026.
Two examples make the line clear. A QR code on the receipt that says "How did we do?" and opens your Google review page is ordinary. A tablet by the till that asks "Happy or unhappy?" and shows the Google page only to the happy is gating. And a staff league table of who collects the most reviews naming them runs into both the quota rule and the ban on asking for named mentions. If you text guests after a booking, send the same message to every guest, whatever they ordered. The longer explanation is in what review gating is and why it can get you penalised.
Google's AI replies, Birdeye and Podium: what to do with the offers
Google's test is labelled "Reply to reviews with AI". Search Engine Roundtable reported it on 24 March 2026, showing for some businesses only, and Search Engine Land also covered it. Google's own help page on replying to reviews did not mention it when I checked, and the reports contradict each other on whether every reply waits for you to submit it: some testers describe drafts they edit and post themselves, others say replies went out without edits. If the option appears in your profile, treat it as a drafting aid, check how your account handles approval, and leave nothing switched on that publishes without you. I explain why in should you let AI auto-post replies to Google reviews.
Birdeye and Podium are a different question. Both sell review requests and customer messaging, both market AI features, and neither publishes a price: Birdeye's pricing page asks for your details and number of locations before it shows anything, and Podium's tells you to talk to its sales team. Third-party comparison sites quote entry prices that conflict with each other, so I am quoting none. Ask each vendor for an itemised quote and compare these lines:
- Price per location per month, what it includes, and any setup or onboarding fee.
- Contract length, auto-renewal terms and the notice period.
- Which review sites it reads and replies on, and whether Tripadvisor and Yelp are included.
- Whether AI replies can post without your approval, and whether you can switch that off for every star rating.
- Any cap on review-request texts, and what extra volume costs.
- Who owns the review and customer data, how you export it, and what happens to your replies and automations if you cancel.
- How many staff logins come with the price.
My rule of thumb: for one restaurant, the free route above (alerts, the log, a monthly themes read and drafts you approve) handles most of what these platforms sell at that size. A quote starts to make sense when you run several sites, or when requesting reviews by text becomes a job you cannot keep up with. My side-by-side is in Birdeye vs Podium: reviews, texting and contracts.
A one-page reply policy for whoever covers for you
Put this where the manager can see it, and change the times to suit your opening days.
| Question | Rule |
|---|---|
| Who can reply | The owner and the general manager. Nobody else holds the logins. |
| Short and detailed positives | Within a week, in the weekly batch. An AI draft is allowed; read it aloud and vary the wording before posting. |
| Mixed reviews | Within two days. The owner or manager rewrites the middle of the draft. |
| One- and two-star reviews | Owner within 24 hours (48 if closed). No AI draft. |
| Allergy, illness, injury, discrimination, staff conduct | Same day. Guest contact and incident note first, then a short public reply. |
| Never in a reply | Guests' names, booking details, refund or compensation amounts, or anything calling the reviewer a liar. |
| Log | Every review gets a row within a week. |
Reply rate, reply speed and a spot check after two months
Pull the log after two months and work out four numbers: the reply rate (replies divided by reviews), the time to first reply on one- to three-star reviews, how many complaints got a private follow-up, and how many themes led to a named change. Suppose the log shows 76 reviews over two months, with 74 answered. That is a reply rate of 97%. Ten of the 76 were complaints, all answered within a day except two that landed over a closed weekend, and eight had a private follow-up. The slow-service theme appeared in 6 reviews in month one and 3 in month two, after the manager moved a server's start time. Those figures are invented to show the arithmetic, and your own will be messier.
Add a spot check. Ask the manager to read five posted replies at random and mark any that could sit under another restaurant's page. One generic reply in five means the style sheet needs another rule; two means someone is posting drafts untouched.
Be modest about the rating itself. Google ratings move for many reasons, so write down your average on Google with the date, for example "4.4 on Google at 1 December 2026", and do not credit the change to AI. What you can fairly claim is the process: complaints seen the same day, every review answered, and themes turned into rota or kitchen changes.
Questions owners ask before automating review replies
Do I have to reply to every review?
No rule says so, but a reply is public and future diners read it. Reply to every complaint and every detailed review, and give short positives a brief thank-you in a weekly batch. Google's review tips advise against sending the same message to everyone, so change a detail each time. Skipping a few five-star reviews with no text costs you very little.
Should I reply in the reviewer's language?
If you can do it well, yes: a short reply in the reviewer's language shows care. Use AI to translate the review so you understand it, and to draft the reply, then have a fluent speaker check anything longer than a sentence or two. If nobody can check, reply simply in your own language and avoid idioms.
Does replying to reviews help my Google ranking?
Google's local-ranking help page says more reviews and positive ratings can help your ranking, and that helpful replies help you stand out. It does not say the replies themselves move you up. The case for replying is what the next diner reads under the review. That is also a reason not to stuff replies with keywords: write each one for someone deciding where to book tonight.
How long does a reply take to appear on Google?
Usually within about ten minutes, though Google says it can occasionally take up to 30 days. If nothing shows after a day, check that you are signed in as an owner or manager of the right profile, and avoid posting the same reply twice.
Can I quote a review on my menu or website?
Ask first. Reviews belong to the reviewer, so get their consent before reproducing one, and quote it exactly. Describe your rating as being on Google with the date you checked it, because the figure moves. Skip any review that mentions an allergy, an illness or a named member of staff.
How do I handle a review that names a member of staff?
For praise, thank the team in general terms, or use a first name only if that person is happy to be named. For criticism, speak to the person and the manager before you reply, and never name or discuss them in public, because that is an HR matter. Say what you will look into, and keep the reply about the guest's experience.
Further reads
- How to Get More Google Reviews With Automated Review Requests — Request reviews by text without tripping Google's rules.
- Restaurant Google Business Profile: AI-Drafted Menu, Hours, Photos — Keep the profile that sits beside your reviews accurate.
- Franchise Review Management: Replies for Several Stores, On Brand — If you run several sites, keep replies consistent.
- How to Reply to Negative Reviews With AI: Examples and Pitfalls — More wording for negative reviews, with pitfalls.
- Should an AI Chatbot Answer Allergen Questions for Your Restaurant? — Whether a chatbot should answer allergen questions at all.
- 7 AI Mistakes Restaurant Owners Make With Bookings and Reviews — Seven slips restaurants make with AI on bookings and reviews.
- AI SEO for Restaurants: Can You Get Recommended by ChatGPT? — How a restaurant can test what ChatGPT and Google AI say about it, then fix listings, menu text, reviews and crawler settings, and what stays out of its hands.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Google Business Profile Help pages on reading and replying to reviews, tips to get more reviews and local ranking (checked October 2026); Google's review policy; Search Engine Roundtable (24 March 2026) and Search Engine Land reports on the Reply to reviews with AI test; Birdeye and Podium pricing pages, neither of which publishes a price.