Key takeaways
- Ratings move revenue: a one-star Yelp increase was linked to a 5 to 9% revenue rise for independent restaurants[1][2].
- Let AI draft, never publish: a named person approves every reply, and anything about illness, injury or staff goes to a manager.
- The FTC review rule bans fake reviews and incentives tied to positive sentiment[3][4]; AI makes both easier to do by accident.
- Give the drafting tool facts, a voice guide and banned phrases, or every reply will sound the same.
Why review replies are worth getting right
Online ratings shape where people eat, and the effect is strongest for independents. Harvard Business School research found that a one-star increase in a restaurant's Yelp rating led to a 5 to 9% increase in revenue[1], and the effect was driven by independent restaurants rather than chains[2]. A UC Berkeley study of San Francisco restaurants found that moving from 3 to 3.5 stars raised the chance of selling out at prime time from 13% to 34%[5].
Source: University of California, Berkeley, 2012[5]
Show the data
| Item | Value |
|---|---|
| 3 stars | 13% |
| 3.5 stars | 34% |
Replies do not change the star rating directly, but they shape how future guests read the reviews, and a thoughtful reply to a fair complaint often reads better than a perfect score. The catch is time. Writing considered replies to every review on several platforms is a job nobody on a busy team wants, which is why AI drafting is appealing, and why it needs guardrails.
What AI drafting does well and badly
AI is good at the first draft: acknowledging specifics, matching tone, keeping the length right and avoiding the defensive reflex of a tired manager at midnight. It is bad at anything that needs judgment or facts it does not have.
- Good: thanking positive reviewers and mentioning the dish they praised.
- Good: turning a manager's bullet notes into a calm, short reply.
- Good: keeping a consistent voice across platforms and locations.
- Bad: inventing facts, such as promising a refund policy you do not have or claiming the chef was off that night.
- Bad: handling food-safety, allergy, injury or discrimination complaints, which need a manager and sometimes legal advice.
- Bad: repeating the same phrases until every reply looks automated.
Seven guardrails to set before the first draft
- A person approves every reply. Drafts never post automatically, whatever the star rating.
- Escalation topics are hard-coded: illness, allergic reaction, injury, harassment, discrimination, staff names and anything mentioning a lawyer go straight to the general manager, with no AI draft at all.
- Facts come only from a fact sheet: hours, policies, chef's name, what the restaurant will and will not do. If it is not in the sheet, the draft says nothing about it.
- No offers in public replies. Do not offer discounts or freebies in the reply; invite the guest to contact a named person instead.
- No staff blame. Drafts never name or criticize employees.
- A banned-phrase list: remove the stock openings and closings that make replies sound generated.
- Disclosure is honest: replies are signed by the person who approved them, and nobody pretends to be a customer.
If an AI tool can post on its own, switch that off. One wrong automated reply to a food poisoning review can do more harm than a year of unanswered praise.
Where the FTC rule comes in
The Federal Trade Commission's rule on consumer reviews makes it an unfair or deceptive practice for a business to write, create or sell fake consumer reviews that misrepresent the reviewer's experience[3]. It also bars offering compensation or incentives conditioned on a review expressing a particular sentiment, positive or negative[4]. AI does not change those rules, but it makes breaking them faster: generating a batch of glowing reviews takes seconds, and so does writing a follow-up email that offers dessert for a five-star review.
- Never use AI to write reviews of your own restaurant, or of competitors, under any name.
- Ask every guest for an honest review, never only the happy ones, and never tie a reward to the rating.
- Do not let staff post reviews from personal accounts without disclosing that they work there.
- Keep a record of how review requests are worded and sent.
Platforms police this too. Harvard research estimated that roughly 16% of restaurant reviews on Yelp are filtered as suspicious, and restaurants with weak reputations or recent bad reviews were more likely to commit review fraud[6]. The pressure to cheat is highest exactly when a restaurant can least afford to be caught.
Reply patterns by review type
| Review type | Goal of the reply | Draft must include | Never |
|---|---|---|---|
| Five stars, specific praise | Thank and reinforce | The dish or person they mentioned; a reason to return | A sales pitch |
| Five stars, no detail | Short thanks | One sentence, signed | Copy-paste wording used last week |
| Mixed, fair criticism | Acknowledge and show the fix | What you changed, if true and in the fact sheet | Arguing about details |
| One or two stars, service complaint | Take it offline | Apology for the experience and a named contact | Offers, excuses or staff names |
| Health, safety or allergy | Manager handles | No AI draft: manager writes after checking facts | Any statement about cause |
| Suspected fake or wrong restaurant | Report, then reply briefly | A neutral note that you cannot find the visit | Accusing the reviewer |
The approval workflow
Collect
Pull new reviews from each platform into one queue daily. Tag each by rating and topic.
Route
Escalation topics go to the general manager. Everything else gets an AI draft.
Draft with context
The tool gets the review, the fact sheet, the voice guide, the banned-phrase list and any manager notes about the visit.
Approve or edit
A named person reads, edits and approves. Aim for replies within two days.
Post and log
Post under the approver's first name and role. Log the reply, the approver and any follow-up promised.
Review monthly
Read a sample of last month's replies side by side. If they sound alike, update the voice guide and banned phrases.
A drafting brief you can reuse
Most bland AI replies come from bland instructions. Give the tool a brief like this, adapted to your restaurant, and keep it in the same document as your fact sheet.
Write a reply under 80 words in a warm, plain voice. Mention one specific detail from the review. Use only facts from the fact sheet. Do not offer discounts, name staff or discuss health issues. Do not start with Thank you for your feedback or end with We hope to see you again soon. Sign off as the approver's first name and role.
Operators already see customer feedback as one of the areas most affected by technology[7]. The restaurants that do it well treat AI as a fast writer working under a strict editor, not as the voice of the business.
Build a voice guide in an hour
The voice guide is what stops a chain of AI replies from sounding like one template. It does not need to be long. Gather ten replies your best manager has written and pull out what makes them work.
- Three words that describe how you sound, for example warm, direct and a bit playful.
- How you refer to the restaurant and the team: we, the kitchen, our crew, or first names.
- Words and phrases you never use, including stock openers and over-the-top apologies.
- How long replies should be by review type, and whether you use exclamation marks.
- How you sign off: first name and role, never a generic team signature.
- Two strong example replies, one positive and one negative, for the tool to follow.
Review the guide every quarter. Menus, managers and policies change, and an out-of-date guide produces replies that sound right but say the wrong thing.
Example: one complaint, two replies
Example (fictional): a two-star review says the burger took 45 minutes on a Tuesday and the server seemed overwhelmed. An unguided draft apologized at length, blamed short staffing and offered a free meal on the next visit. The guarded draft, using the brief and the manager's note that Tuesday staffing had been fixed, read: Sorry your burger took so long on Tuesday. That is not the wait we aim for, and we have added a second server on weeknights since your visit. If you are willing, email Dana, our general manager, and she will make it right. The second version is shorter, factual, avoids blaming the server and keeps any make-good offer private.
For the wider picture on review management, connect this workflow to your weekly restaurant dashboard so rating trends sit next to sales. Read AI sales and labor forecasting if slow-service reviews point to a staffing problem, state AI disclosure laws for customer calls for the honesty rules that apply to voice and chat, and connecting POS, reservations and ordering if you want reviews matched to real checks.
Frequently asked questions
Is it allowed to use AI to reply to reviews?
Yes, as long as replies are accurate and do not misrepresent who you are. The FTC rule targets fake reviews and incentives tied to sentiment[3][4], not businesses drafting their own replies. Have a person approve every reply.
Should we reply to every review?
Aim to reply to every negative or mixed review and to a good share of positive ones, especially those with detail. A short, specific reply beats a long generic one.
Can we ask happy guests to leave reviews?
You can ask guests for reviews, but ask everyone, not only those you expect to be positive, and never tie a reward to the rating.
What if a review is fake?
Report it through the platform's process with any evidence, such as no matching reservation or check. If you reply, keep it neutral and brief.
References
Harvard Business School (Michael Luca) (2016). Reviews, Reputation, and Revenue: The Case of Yelp.com (HBS Working Paper 12-016) (opens in a new tab) ↑
Harvard Business School (Michael Luca) (2016). Reviews, Reputation, and Revenue: The Case of Yelp.com (HBS Working Paper 12-016) (opens in a new tab) ↑
Legal Information Institute (Cornell Law School), eCFR text of FTC rule (2024). 16 CFR § 465.2 - Fake or false consumer reviews, consumer testimonials, or celebrity testimonials (opens in a new tab) ↑
Legal Information Institute (Cornell Law School), eCFR text of FTC rule (2024). 16 CFR § 465.4 - Buying positive or negative consumer reviews (opens in a new tab) ↑
University of California, Berkeley (2012). Yelp reviews boost restaurant business (Anderson and Magruder, Economic Journal) (opens in a new tab) ↑
Harvard Business School (2015). Fake It Till You Make It: Reputation, Competition, and Yelp Review Fraud (Luca and Zervas) (opens in a new tab) ↑
National Restaurant Association (2026). Report: Sales to hit $1.55T in 2026 despite challenging business environment (opens in a new tab) ↑