Thursday, 10:40 p.m., Tampa Pasta House. The last guests have left and the chairs are up on the tables. Margaret Ellison opens the restaurant's Google profile: eight new reviews this week. Six kind ones. Two angry. In one of them a customer says he drove half an hour for lunch on Tuesday and found the door locked, because the restaurant's own Telegram had the days wrong.
A reply to a review isn't read only by the person who wrote it. Under the review, it reads as the restaurant's own word, and everyone deciding where to eat on Saturday will see it. We gave eight reviews to an AI and checked how it would answer.
Tampa Pasta House is our sample restaurant. The eight reviews are made up, written the way real ones look.
📬 What we actually tested
On October 22, we gave Gemini 3.8 Flash the eight reviews one at a time, with the same instruction every time: reply as the restaurant, in the reviewer's language, in two to four sentences. Name one detail from the review. Offer nothing for the review or for changing it. Don't argue. Admit a mistake only when the facts show it was ours. Make nothing up. If a review is about food safety, money or a legal threat, flag it for the owner.
0 of 8 replies offered a discount or a gift. 0 invented facts. On the Tuesday review it admitted our mistake plainly. All 8 were signed by Margaret, the owner flags were right in 8 of 8, and the Spanish review got a reply in Spanish.
- Mistake one: the angry review that was our fault. "That was our mistake… lunch is Thursday to Saturday only. Please contact the restaurant privately so we can speak with you directly." It owned the mistake, but the man who drove half an hour to a locked door got no "sorry" and no time to come back: 11 a.m. to 2 p.m.
- Mistake two: a hair in the tagliatelle. "We are sorry to hear about your experience with the tagliatelle." The guest's complaint is that nobody apologized, and the reply still doesn't apologize for what happened or even name it. The flag for the owner was right.
Where it overdid it: caution beat the apology. It stretched "don't take the blame for what you can't know" over everything. The replies to the 25-minute wait and to the locked door don't say "sorry" at all. And 5 of 8 replies end on the same line, "reach out to us privately," even the reply to a four-star review that opens by praising the osso buco.
An angry review isn't answered for its author. It's answered for everyone who reads it next.
🗂 One reply we would post
"Thank you for bringing this to our attention, Kevin. That was our mistake: our Telegram channel wrote that the lunch lasagna is available on […], but lunch is Thursday to Saturday only. Please contact the restaurant privately so we can speak with you directly. — Margaret, Tampa Pasta House"
And the reply we wanted: own the mistake in our post, give the right lunch hours (Thursday to Saturday, 11 a.m. to 2 p.m.), and invite him back. No discount, and no "dessert is on us next time." The model did the first half: it owned the mistake, named the days and gave nothing away. The hours and the apology are missing, and "come back" became "contact us privately." The next guest reading that reply needs the hours.
🛠 How to repeat without code
- Write down the facts the assistant may use in replies: hours, address, your main items and their prices. Everything else is off limits.
- If you've already had a mistake in a post or on your site, put it in the instruction in plain words, "that was our mistake." Otherwise the assistant can't know where to own it.
- Forbid it in plain words: discounts, gifts or "a treat on us" for a review or for changing one.
- Give it one review at a time, and ask for a flag line at the end, "owner should read," for food safety, money or a legal threat.
- Read every reply as someone choosing a restaurant for Saturday would, and only then post it.
🧾 The bill and the limits
8 requests in 6 minutes, 37 to 55 seconds per reply. The tool doesn't print token counts. Eight made-up reviews, one restaurant, one model, one run. We didn't test replies to real reviews, reviews with photos, or long back-and-forths under a review. We didn't ask the model to post anything for the owner: every reply is a draft for a person.
✅ Before you trust an assistant with your review replies
- ☐ Give the assistant a list of facts and forbid everything that isn't on it.
- ☐ Forbid discounts and gifts for a review or for changing one.
- ☐ Own your mistake plainly when it's yours, and don't take the blame for what you can't know.
- ☐ End an angry review with no details on an invitation to talk privately, not an argument.
- ☐ A review about food, money or a lawsuit is read by you first, not by the assistant.
- ☐ Before posting, reread the reply as the customer who will see it.
Want to know what AI already says about your business? Check: smashone.us/check/
Next in "AI on staff": Mateo from Casa Lista and a customer who wants it cheaper.
Do you answer angry reviews, and who do you write the reply for?
