Your listing shows a 4.6. One new one-star review that says "I called and got some robot that couldn't understand a simple order" pulls the average down a fraction, which is not the problem. The problem is that it will be the review people actually finish reading, because it names something specific while the four-star ones say "good food, will return."
So the reply matters more than the rating does. And the reply most operators write is the wrong one.
The reply that makes it worse
The instinct is to defend the system. "We recently upgraded our phone technology to serve guests faster and reduce wait times, and we're sorry you had a poor experience."
Read that as a stranger deciding where to order dinner. You've confirmed the complaint, explained nothing about what went wrong, and spent your two sentences on a benefit the reviewer just told everyone they didn't get. The word "upgraded" sitting next to a one-star review does more damage than silence would.
The second wrong reply is the denial. Any version of "our team is always available to take your call" when the caller demonstrably reached an automated line makes you look like you're managing perception rather than a restaurant. If you named the agent, this is where that choice gets tested in public, which is worth thinking about before you set it up rather than after. That's the argument in naming your AI host.
What a good reply contains
Four things, in about four sentences.
- Confirmation that the line is automated, said plainly, with no softening adjective attached to it.
- The specific failure, named. Not "your experience," but "it didn't get the no-onions on your second sandwich."
- What you changed, if you changed something. If you haven't yet, say you pulled the call and are fixing it rather than promising a review of your processes.
- A direct route to a person: a name, and a number or email that reaches them.
Something like: "You're right, our phone line is answered by an automated system. It missed the no-onion request on your order and that's on us, not you. We've corrected how it handles that modifier and I'd like to make the order right. I'm Dana, the GM, at 555-0134."
Sign it with a real name and post it from the owner or GM account rather than a marketing address. A reply signed by a person who works in the building carries weight that a house voice doesn't, and it makes the offer to call believable. Keep the whole thing under about eighty words. Long replies read as arguments no matter how carefully they're written, and the reviewer isn't the audience anyway.
That reply is useful to the reviewer and, more importantly, legible to the next two hundred people who read it. It says the restaurant knows what happened on a specific call and acted on it. Nothing about the technology needs defending, because a stranger reading it wasn't asking about the technology.
Pull the call before you reply
Do not write anything until you've read the transcript or heard the recording. This takes about five minutes and it changes what you say.
You're sorting the review into one of three buckets, and each one gets a different response.
The agent made a mistake the configuration can fix. A modifier it didn't recognize, an item name it heard as a different item, hours it quoted wrong. This is the common case and it's the good one, because it's specific and cheap to correct. Fix it that day and say so in the reply.
The agent refused to hand off. The caller asked for a person, more than once, and the system kept trying. This is the failure that produces the angriest reviews, and it's a policy problem rather than a recognition problem. Whatever your escalation rules are, they're wrong. Human handoff and failover covers what they should be.
The caller simply didn't want to talk to a machine. The order went through correctly, nothing broke, and they're annoyed on principle. This one is real and there's no configuration change that addresses it. Reply courteously, don't argue, and move on. Chasing this reviewer costs you time and gains nothing, and pretending they represent your whole customer base leads to expensive decisions.
The reviews you should treat as an alarm
Most single complaints are noise. Some are not.
- Two or more reviews naming the same failure inside a month means the failure is on a lot of calls and almost nobody reviewed it.
- Any review describing a caller who couldn't reach a person after asking. That's the pattern that ends up screenshotted and shared.
- A complaint about being deceived rather than about a mistake. That's a disclosure problem, not an accuracy problem, and it's fixed in the greeting rather than in the menu configuration. See telling customers you use AI.
The first of those is the one worth acting on hardest. A review is a very expensive way to find out about a broken modifier. Your order accuracy tracking should be surfacing that first, and if reviews are consistently ahead of your dashboard, the dashboard isn't earning its place. The metric set that catches this is in voice AI metrics and KPIs.
Answering the same complaint offline
Some of these arrive by phone or at the counter instead of on a review site, and those are easier. A guest who calls back angry about a botched order is doing you a favor by not writing it down. Handle it the way you'd handle any other kitchen mistake, which is the ground covered in handling refund and complaint calls.
The thing not to do is treat the automated line as an excuse. "The computer took it down wrong" is technically true and completely useless to a guest holding the wrong food. You chose the system, so it's your mistake. Staff who hear a manager blame the phone will start blaming it too, and then every wrong ticket in the building has a place to go.
The sixty-day read
One review tells you almost nothing. Two months of them tell you a lot.
There's also a quieter signal in the reviews that don't mention the phone at all. Complaints about a wrong order, with no explanation of how the order was placed, often started on a call. Before you conclude the kitchen is slipping, check whether the ticket came in by phone and whether the transcript matches what was made. A run of "they forgot half my order" reviews with no phone mention has sent more than one operator after a line cook who did nothing wrong.
Set a recurring reminder to read every review mentioning the phone, in one sitting, every sixty days. Count how many describe a fixable failure, how many describe a refused transfer, and how many are objections to automation itself. If the first category is shrinking and the third is flat, the system is working and you're looking at the small permanent share of people who won't like it. If the second category exists at all, stop and fix your escalation rules this week.
That count is a better signal than your star average, which moves too slowly to tell you anything while there's still time to act on it.