How to Use AI to Respond to Guest Reviews
How to use AI to respond to guest reviews: sort the week, hold your voice in an agent, draft the five stars fast, and give the one star your own time.

Forty reviews are waiting, and you are about to answer thirty of them with "Thank you for your kind words, we hope to see you again soon."
Don't. Here is how to use AI to respond to guest reviews instead: write down how your property talks once, put it in an agent, let the agent draft, and read every draft before it posts.
A guest can spot a form letter. So can the stranger reading down the page, deciding whether to book your last room.
Let the model draft the easy replies, and spend your own time on the one star.
A five star review is a thank you with one specific detail in it, and a model writes that in seconds. A one star review is a small repair job done in public. A model does it badly until you hand it the facts and the rules first, and most of this post is about that second job.
Replying is worth the hour. Davide Proserpio and Giorgos Zervas studied what happens when businesses start answering reviews, and their 2018 Harvard Business Review write-up found that replying to customer reviews results in better ratings.
Everything below follows one Monday at the Harbor Street Inn, a made up 32 room inn with a restaurant. The guests and their reviews were written for this post. You are the general manager, the week's reviews have stacked up, and you have about an hour.
If you run a restaurant or a venue instead, the steps do not change. Only the facts in your instructions do, and Magai for hospitality shows how other properties set it up.
Sort the week before you answer any of it
Open a new chat in Magai and paste in the whole week: every review, with the platform and the star rating above it. Reviews are public, so nothing private goes into this step. Booking records are different, and they stay out.

Here is the pile, cut down to the five reviews this post answers.
| # | Platform | Stars | What the guest wrote |
|---|---|---|---|
| 1 | 5 | "Stayed two nights for my sister's wedding. The room was spotless, and the front desk found us a late checkout without us even asking. Breakfast was the best part." | |
| 2 | 5 | "Great breakfast!!" | |
| 3 | Tripadvisor | 4 | "Lovely inn and a great location. Dinner in the restaurant was slow, nearly an hour between courses, and the bathroom fan was loud." |
| 4 | 1 | "Worst stay of my life. We arrived at 4 pm and the room was not ready until nearly 6. Nobody apologized. Then the heat didn't work and the man at the desk said it was 'not his problem.' Paid $289 for this. Never again." | |
| 5 | Tripadvisor | 1 | "Do not stay here. Awful." |
Type this under the pile:
Here are this week's guest reviews for the Harbor Street Inn. Each one starts with the platform and the star rating.
Sort them into five groups:
1. Five star with a specific detail
2. Five star with no detail
3. Mixed (three or four stars, or praise and a complaint together)
4. One or two star with a specific complaint I can check
5. Anything involving injury, illness, safety, theft, discrimination, or a legal threat
For each review give me the group, a one line summary, and what I would need to look up before replying. Do not draft any replies yet.
The sort comes back as a table with a group, a summary and a column called Look up first. That last column is the useful one. For review 4 it said to check when the room was marked ready, when the heat was reported and fixed, and what the desk clerk remembers saying. It told you what to confirm before a single word is drafted.
It also made one mistake. It put review 5 in group four, as a complaint you can check. There is nothing to check in "Awful." That is a mood with a star on it, and it gets the shortest reply of the week.
Group five is the one the model never answers for you. If a review mentions an injury, an illness, a theft, discrimination or a lawyer, stop. Call your insurer or your counsel before anything goes public. The model can help you write down what happened for your own file. The reply the public sees is yours.
That sort decides who touches each review.
| Group | Who drafts | Who sends | Before it goes out |
|---|---|---|---|
| Five star with detail | The agent | You, after a skim | Check that every detail came from the review |
| Five star, no detail | The agent, kept short | You, after a skim | Cut anything padded |
| Mixed | The agent, with your manager note | You | Check the note was used and nothing was added |
| One or two star with something to check | The agent, after you do the homework | You | A second model checks it, then you read it aloud |
| Injury, illness, safety, theft, discrimination, legal | You | You, after calling the right person | Not an AI draft |
Write your house voice down once, as an agent
Now the part that saves the hour. An agent is the place you write those instructions once. In Magai it keeps your property's voice, rules and reply format, runs on whichever model suits the job, and can be shared with your whole team. If you want the longer walk through building one, how to build an AI agent without writing code covers it. For this job you need exactly one.

Create an agent, name it Harbor Street replies, and paste in the instructions below. They work as a system prompt that sits in front of every message you send. Swap in your own property, your own facts and the name of whoever signs.
You write public replies to guest reviews for the Harbor Street Inn, a 32 room inn with a restaurant on the waterfront. You write as the general manager, Maria Alvarez. Every reply is signed "Maria, General Manager".
Voice: warm, plain and specific. Short sentences. No exclamation marks. No "valued guest", no "we strive", no "your satisfaction is our priority".
Facts you may use: the inn has 32 rooms, a breakfast room, a restaurant open for dinner Wednesday through Sunday, free parking, and a front desk staffed around the clock. Standard check in is 3 pm. Use nothing else about the inn unless I give it to you in the message.
Rules:
1. Mention one specific detail from the review, in your own words. Never quote the review back.
2. Never invent a fact, a cause, a fix or an action by staff. If I give you a manager note, you may repeat what it says and nothing more.
3. Never offer a refund, a discount, a free night or any gift in a public reply. If a guest needs to be made whole, invite them to email guestcare@harborstreetinn.example or call the front desk, and stop there.
4. Never name staff other than the person signing.
5. On Google, greet the reviewer by the name shown. On Tripadvisor, never use a name, an address or any detail that could identify the guest.
6. Plain text only. No bullets, no bold, no links, no emoji. One language per reply, the language of the review.
7. Length: five star replies run 40 to 70 words. Mixed and one star replies run 80 to 130 words.
8. Never argue, never explain away, and never tell a guest they are wrong.
After every draft, add one line starting with CHECK: that lists each sentence you wrote that makes a claim or a promise, and where in my message it came from.
Some of those rules come from the platforms and not from taste. Rule 3 is the expensive one. Google's guidance on replies says to avoid using your response to offer deals or promotions, and Tripadvisor's management response guidelines say responses offering personal gain, such as gifts, services or money, will be removed. A model trying to be kind reaches for a free night. The rule stops it.
Rule 5 exists because the two platforms disagree. Google tells you to address the reviewer by name. Tripadvisor prohibits personal information that could identify a guest, and it lists names first. An agent without the rule picks one platform and breaks the other.
Rule 2 matters most, and the next section shows why.
The CHECK line is an editor's habit. Every draft arrives with the model's own list of the claims it made and where each one came from, so you verify five lines and not forty replies.
The agent keeps those instructions in front of whichever model you pick, so the voice holds when you change models partway through a job. Switching models in a Magai chat does not reset the thread's memory. You will lean on that in the one star step.
Run the five star replies and read what the model invented
Start a thread with the agent selected, and ask for Claude by name with a slash pill at the front of the message. Paste both five star reviews, one platform line each.
/Claude
Platform: Google. Rating: 5.
Reviewer name shown: Priya S.
Review: "Stayed two nights for my sister's wedding. The room was spotless, and the front desk found us a late checkout without us even asking. Breakfast was the best part."
Platform: Google. Rating: 5.
Reviewer name shown: Marcus T.
Review: "Great breakfast!!"
The first version of the agent had rules 1, 4, 5 and 7 and nothing else. Its reply to Priya opened with "We're so glad you enjoyed the harbor view from your room." Priya never mentioned a view. Half the rooms at the Harbor Street Inn face the parking lot.
Rule 2 came out of that sentence.
A model fills silence with pleasant details because it is trying to sound like it was there. A guest in a parking lot room reads that line and knows nobody read the review. So does everyone else scrolling past.

With the full rules in place, the replies came back like this.
Hi Priya, thank you for choosing us for your sister's wedding weekend. I'm glad the room was spotless, and that a late checkout was one less thing to think about on the way out. I'm also glad breakfast was the best part. We would love to host you again.
Maria, General Manager
CHECK: the wedding, the spotless room, the late checkout and breakfast all come from the review. No other claims.
Hi Marcus, thank you for the note about breakfast. I'm glad it started your morning well. We hope to see you again.
Maria, General Manager
Priya's reply uses three details, all hers, in different words than she used. That is what rule 1 asks for.
Marcus's reply runs 25 words, under the 40 word floor in rule 7. Leave it. Nothing in "Great breakfast!!" deserves more, and a padded reply to a three word review is the most obvious form letter on the page. Change rule 7 to read "40 to 70 words, or shorter when the review is short," and the agent stops padding.
Read the CHECK lines first, then read each reply once out loud. Two replies take about two minutes, and thirty take about the same time per reply. This is the part of the job that should feel boring.
Pick the model by what a bad reply costs you. A wrong five star reply costs a wince. A wrong one star reply costs bookings, so that is where your most careful model and a second opinion go.
Give the model one true sentence per complaint
Review 3 is where an agent starts needing you. The guest liked the inn and disliked two things, and the model has no idea why dinner was slow or whether the fan was ever fixed. Left alone, it will apologize for the entire building or invent a reason.
So give it a manager note: only things you know to be true, and what to keep out of public.
Platform: Tripadvisor. Rating: 4.
Review: "Lovely inn and a great location. Dinner in the restaurant was slow, nearly an hour between courses, and the bathroom fan was loud."
Manager note:
* That Saturday we had a 24 person wedding party in the dining room and one cook out sick. Do not mention either in the reply.
* The bathroom fan in that room was replaced on October 30.
* I want to hear more about dinner service from this guest.
The reply:
Thank you for staying with us and for the kind words about the location. An hour between courses is longer than any table should wait, and I am sorry dinner ran that slowly. The bathroom fan has since been replaced. If you are willing to tell me more about the dinner, please email guestcare@harborstreetinn.example.
Maria, General Manager
Four things to check in it. There is no name, because this is Tripadvisor. It repeats the fan fact from the note and adds nothing. It leaves out the wedding party and the sick cook, because you said to. And it apologizes for the wait without a speech about how unusual a bad Saturday is.
That last one is a choice. Explaining your staffing turns an apology into an excuse. Google's reply guidance says to acknowledge mistakes without taking responsibility for things outside your control, and the note told the model where you think that line sits.
The model can only be as honest as the facts you give it. A manager note takes forty seconds to type. It separates a reply that knows something from a reply that is guessing.
Keep the note to what you would say out loud to the guest. Dates, fixes and one sentence about what you want next. The guest's booking record, card details and reservation number do not belong in it, because the note is for your facts and not for their file.
Use AI to respond to guest reviews that sting, in three passes
Review 4 is the reason you are doing any of this. The first pass is yours, the second is the agent's, the third belongs to a different model. The send is yours again.

Pass one: do the homework before you open the agent. The model cannot know when a room was ready. Your property management system does, and so does the maintenance log, and so does the clerk who was on shift. Ten minutes of lookups are the whole reason the reply can be specific. You need the times and the clerk's account. You do not need the guest's card, address or reservation number, so keep those out of the prompt.
Pass two: draft it. Start a fresh message in the same thread, with the agent still selected.
Platform: Google. Rating: 1.
Reviewer name shown: Tom R.
Review: "Worst stay of my life. We arrived at 4 pm and the room was not ready until nearly 6. Nobody apologized. Then the heat didn't work and the man at the desk said it was 'not his problem.' Paid $289 for this. Never again."
Manager note, confirmed by me from our own records:
* The room was marked ready at 5:52 pm. Check in time is 3 pm.
* The heat in the room was reported at 6:40 pm and fixed at 7:15 pm.
* The clerk on shift says he told the guest maintenance was on the way. I found no record that anyone apologized for the wait.
* I am the person to contact.
Draft the reply. Do not defend the clerk. Do not offer anything to make it up to the guest in public.
The earlier version of the agent, the one without rule 3, ended its first one star draft with "We would love to welcome you back with a complimentary night on your next visit." It also opened with "Our standard check in time is 3 pm." Both lines are the model being helpful in the worst possible direction.
The first gets your reply removed on Tripadvisor and breaks Google's guidance. It also teaches every reader that a one star review earns a free night. The second is an argument with a guest, in public, over the thing they are angry about.
With the full rules and your note, the draft came back like this:
Tom, I'm sorry. You arrived at 4 pm for a room you had paid for, and our own records show it was not ready until 5:52. You should have heard an apology for that wait, and I can't find that you were given one. Then the heat was out, and being cold in a room after waiting that long is the opposite of what a stay here should be. Nobody on our team should tell a guest a problem is not theirs to handle. I would like to hear the rest from you. Please email me at guestcare@harborstreetinn.example or call the front desk and ask for me.
Maria, General Manager
Good bones. It concedes what your own records prove, it sends the rest of the conversation to email, and it offers nothing in public. Google's reply guidance says to protect privacy and, if a situation is complex, politely request that the reviewer contact you by phone or email. The draft does exactly that.
One sentence is a problem. "Nobody on our team should tell a guest a problem is not theirs to handle" assumes the clerk said it. Your note says the clerk remembers telling the guest maintenance was on the way. You do not know yet what was said at the desk. The third pass exists to catch sentences like that one.
Pass three: switch models and ask for a skeptic. Staying in the same thread, send this:
/GPT
Read the draft above against the agent's rules and my manager note. List every sentence that states a fact, makes a promise, or could read as defensive or as an admission I have not confirmed. For each one tell me whether my note supports it.
Because the thread carries over when you switch models in one workflow, the second model sees the rules, the review, your note and the draft without a single re-paste.
It flagged two sentences. The one about the team assumed what was said at the desk, which the note does not confirm. "Is the opposite of what a stay here should be" is editorializing about the guest's experience, which is fine for a manager to say but is not a fact from the note. It cleared the times, the apology and the invitation to email.
You fix the first and keep the second. The sentence becomes: "If anything said at the desk sounded like the problem was not ours, that is not how we want any guest spoken to."
Then you read the whole reply out loud, change the opening to your own words, and post it from your own login. Nothing in this workflow posts for you, and nothing should.
Review 5 gets the shortest reply of the week.
I'm sorry your stay was not what you hoped for. I would like to understand what went wrong, and I can't tell from your review. If you are willing to tell me, please email guestcare@harborstreetinn.example.
Maria, General Manager
It asks and guesses nothing. Never let a model guess at what went wrong behind "Awful." That is an invitation to invent, and every invented detail is a public claim with your name under it.
Check every reply against the platform's own rules
Both platforms publish what they will and will not accept in a reply. The two sets disagree in places, which is why rule 5 exists.
Here is where the two sets of guidance differ and overlap.
| Topic | Google's reply guidance | Tripadvisor's management response guidelines |
|---|---|---|
| The guest's name | Address the reviewer by name | No names, addresses or phone numbers |
| Offers and gifts | Avoid deals or promotions | No gifts, services or money, and no promotional content |
| Links | Not addressed in its reply tips | No links unless relevant to the review |
| Quoting the guest | Not addressed in its reply tips | No snippets of the traveler's review |
| Formatting | Keep it short and simple | No lists, bullets or other formatting that will not appear |
The agent already carries those rules. The check is a second model reading the finished reply against them, because a model that did not write the reply is the one most likely to catch it.
/Gemini
Check the reply above against these platform rules.
Platform: Tripadvisor.
Rules: no names, addresses or phone numbers. No gifts, services, money or promotional content. No links unless relevant to the review. No quoted snippets from the review. No lists or bullets. One language only.
Quote any sentence that breaks a rule and name the rule. If nothing breaks a rule, say so in one line.
That prompt takes ten seconds to run, and on a week of replies it is the cheapest insurance you have.
Two practical notes. Google requires you to verify your business before you can reply to reviews, and reviews are available only in certain countries and business categories, so a missing reply box is a setup problem and not a bug. Tripadvisor says a response that follows its guidelines will typically be published within forty-eight hours. A reply that has not appeared the same afternoon is not a failed one, so note the date you sent it and look again in two days before you rewrite anything.
Booking.com, Expedia, Yelp and OpenTable have rules of their own, and this post checked only the two above. For any other platform, paste its reply guidelines into the thread and run the same prompt against them.
Answer reviews in the guest's language
A guest writes in Spanish, and your fourth language is "menu French." The model can answer in the guest's language, and the platforms care how.
Tripadvisor says machine translated responses that are unreadable or incomprehensible will not be accepted, and its rules for notification responses reject replies written in two languages. Stay in one language, the guest's.
The risk is sending words you cannot read. So ask for a back translation you can.
Platform: Tripadvisor. Rating: 2.
Review (Spanish): "La habitación estaba limpia, pero el ruido de la calle no nos dejó dormir. Esperábamos más tranquilidad."
Reply in Spanish, following the agent's rules. Then, on a separate line after the sign off, give me a literal English translation of the reply so I can check what I am about to publish. The English line is for me and is not part of the reply.
Gracias por su comentario. Lamento que el ruido de la calle no les dejara descansar, y me alegra saber que la habitación estaba limpia. Si desean contarme más, pueden escribirme a guestcare@harborstreetinn.example.
Maria, General Manager
The English line under it reads: "Thank you for your comment. I am sorry the street noise did not let you rest, and I am glad to know the room was clean. If you would like to tell me more, you can write to me at guestcare@harborstreetinn.example." That matches the review and promises nothing.
The back translation is the check. If someone on your team speaks the language, show them before it goes out. A model's translation is a draft like any other draft.
Put Monday on a schedule
Everything so far takes about an hour, which is fine once. Every week, it is a chore you will start skipping in March.
A scheduled task in Magai runs an agent on a clock, with your connected apps, and every plan includes scheduled task runs. A weekly task uses a handful of them a month. The same pattern runs a social media content calendar, and the scheduled tasks page covers the options.

Review sites send you an email when a new review arrives, or they let you turn that on. Point the task at that mail through the Gmail integration, and it can read your inbox and leave a draft for you.
Every Monday at 7:30 am, open my Gmail and find the review alert emails from the last seven days from Google and Tripadvisor. Run the Harbor Street replies agent on each review. Sort the reviews into the five groups.
Draft replies for groups one, two and three. For group four, list what I need to look up and draft nothing. For group five, flag it at the top and draft nothing.
Save everything as one draft email to me titled "Reviews to answer this week".
By the time you sit down with coffee, the sort is done, the easy replies are drafted and the hard ones are listed with their homework. You are doing your actual job: reading the replies and calling the guest who deserves a call.
Check the first run against each platform's own dashboard. Any platform that sends no email alert never reaches the task, and its reviews have to be pasted in by hand. A task cannot answer a review it never saw.
The task drafts and never posts. Keep it that way.
Give the front desk the agent and keep one rule
Share the agent with your team, and the Saturday night supervisor gets the same voice you do. Anyone covering for you drafts in your tone with your rules, and nobody retypes "friendly but not gushing" differently each time.
One rule travels with it: the agent drafts and a manager sends. Write it on the first line of the shared instructions. The day a reply goes out unread is the day a model tells a guest about a rooftop terrace you do not have.
Three habits keep the agent honest.
- Update the instructions the day a fact changes. A new check in time or a closed restaurant night that the agent does not know about becomes a wrong public reply.
- Keep a short facts list beside it: hours, room count, parking, policies, anything a guest might ask. Paste from it into manager notes.
- Read last month's replies once a month, aloud, in a row. Repetition shows up in a row that never showed up one reply at a time.
Track what matters from the platforms' own dashboards: how many reviews you answered and how fast. The agent should move both numbers. If it does not, the drafts are sitting unread.
One last line of defense sits outside the agent. Use AI to answer reviews and never to write them. Google's policy says reviews and other contributions to Google Maps must reflect a genuine experience, and offering incentives for reviews, or for removing negative ones, is prohibited. A fake five star review written by a model is a worse problem than forty unanswered ones.
Your guests' privacy gets the same care as your own. In Magai, your prompts, files and outputs are never used to train any model, from any provider on the platform. That does not change what belongs in a public reply, which is nothing that identifies a guest.
Next Monday
You have one agent, one sort prompt, one manager note format and one second opinion.
Next Monday, open the thread, paste the week, and read the Look up first column before anything else. Draft the thirty. Do the homework on the one star, give the agent your facts, and let a second model argue with the draft.
The first Monday will take the full hour. The goal is that the sort and the five stars shrink to a few minutes a week, and the one star is the only job that still asks for you. Watch for the Monday when three reviews complain about the same thing. That is not a reply problem. It is a repair job, and the sort prompt just handed you the list.
Thirty thank-yous that sound like you. One reply that earns the next booking.
Start with the one star. Then do the rest in Magai.
More in AI Assistants

AI Assistants
Best AI Chatbot: Run Five Tasks, Keep the Winner
The best AI chatbot depends on the task. Five jobs, where to start for each, and a scored test you run in one thread.

AI Assistants
ChatGPT Alternatives: Pick by the Job, Not the Hype
ChatGPT alternatives sorted by the job that made you leave: Claude, Gemini, Perplexity, Grok, Copilot and DeepSeek, with prices and where each one fails.

AI Assistants
How to Use AI for Coaching Session Notes
How to use AI for coaching session notes: turn rough notes into a summary, the client's action steps and next focus, and catch what the model invents.