Saturday, 23:41. A physio clinic on a high street picks up a one-star review. Eleven words, none of them kind.
By 23:42 there is already a reply on the profile. A tool connected to the Business Profile wrote it: "So sorry your Thursday session wasn't what you hoped for, Marta. As we explained, some soreness in the first few days after the injection is normal."
It is polite, it is fast, and it is a mistake. In two lines it has confirmed publicly that this person is a patient, which day she came in and what was done to her. The owner reads it on Monday morning.
That is the whole tension in one scene. AI is good at exactly the thing a small business struggles with — being there, and being there quickly — and bad at exactly the thing that is expensive to get wrong here.
What people actually search, and what is already written
The queries with real volume are practical and slightly desperate: how to respond to a bad Google review, how to reply to Google reviews examples, can I get a Google review removed. What comes back is templates. Dozens of articles with the same three-step structure — thank, apologise, take it offline — and two or three near-identical example sentences.
They are not wrong. They are incomplete: none of them covers what can cost you money, which is precisely what changes when the writer is not a person but a model with access to your customer records.
It is also no longer only a third-party-tool question. Google is testing its own Reply to reviews with AI inside Business Profile: it surfaced on 20 March 2026, on a limited set of accounts, and has been spotted mainly in the US, Brazil and India rather than Europe (Search Engine Land). The question is going to land in front of everyone whether they go looking or not.
Where it genuinely helps
Two things, and both are logistics rather than talent.
Speed. BrightLocal's 2026 local consumer review survey — 1,002 adults, US sample, so read it as a direction of travel rather than a UK number — finds 19% now expect a reply the same day they post, up from 6% the year before; 32% want one by the next day, and 81% within a week. Nobody hits that by hand on a Saturday night.
Volume and languages. Forty reviews a month across three languages is pure admin. A decent draft in thirty seconds saves an afternoon a month.
And here is the trap. The same survey finds generic or templated replies put off 50% of consumers. AI is mostly being used to mass-produce precisely the thing half the audience spots and penalises. The stat doing the rounds on this topic — "89% read the responses" — is attributed to that survey, but we could not find it in BrightLocal's own published summary, so we are not using it.
The €4,000 reply
March 2026. Spain's data protection regulator fined a Barcelona restaurant €4,000 over its replies to two negative reviews (case EXP202408724). The public reply included a customer's full name, the university they attended, their sexual orientation and details about the person they were with. The business argued all of it was visible on public social media profiles. The regulator's answer was that being findable online does not license you to collect it and republish it, found breaches of Articles 6 and 9 of the GDPR, and ordered removal within ten days.
That was written by an angry human, not by AI. Which is the point. A model wired into your CRM never gets angry, but it also does not know the three things a public reply must not do:
- Confirm the relationship. "Thanks for coming in on Thursday" is itself published personal data: it says that person was there. In healthcare, legal services or anywhere the fact of being a client is sensitive, that is Article 9 territory. UK GDPR carries the same wording; we have not found a published ICO case on review replies, which is not the same as it being safe.
- Add detail in order to win the argument. The natural instinct — "your appointment was 17:00 and you arrived at 17:40" — is the breach.
- Drag in third parties. Staff included.
Google's own rules, which the AI also does not know
Your replies go through the same filter as reviews: Google screens them before they publish against the Maps user-generated content policy. Two clauses that a model trained to be agreeable walks straight past.
One: the policy bans content posted because of an incentive, and expressly includes "content posted following requests for revision or removal of a review or rating in exchange for an incentive". The single most tempting sentence in reputation management — "we'd love to offer you a complimentary session if you'd give us another go" — written in public, is rating manipulation under Google's own rule.
Two: "Do not distribute or post personal information without consent." Same problem as above, arriving down a different road.
This is not decorative. In its 2025 Maps trust and safety report, published on 16 April 2026, Google says it blocked or removed 292 million policy-violating reviews, restricted more than 782,000 accounts and took down over 13 million fake Business Profiles (Google blog). On 17 April 2026 it also tightened the rating-manipulation section to ban businesses setting staff review quotas or asking staff to solicit reviews that name an employee.
There is a UK-specific version of the same warning. Under the Digital Markets, Competition and Consumers Act 2024, whose consumer provisions took effect on 6 April 2025, the CMA can fine a business up to £300,000 or 10% of global turnover directly, without going to court. It wrote to 54 firms after reviewing more than 100 in July 2025, and on 26 March 2026 opened formal investigations into five: Autotrader, Dignity, Feefo, Just Eat and Pasta Evangelists (CMA case page). If the "AI reputation" package you are buying also sends review-request campaigns with a prize draw attached, that is the bit that carries the regulatory risk, not the replies.
What it does not solve
- It removes nothing. Neither you nor the AI can delete a review. You can report it and wait.
- It does not know whether the review is true. It has no notes, no job sheet, no diary. If it replies, it replies blind.
- It cannot read sarcasm. "Absolutely outstanding, as ever" attached to one star is a minefield.
- It does not know what you can promise. An automated "we'll refund you" is a public offer, with your name and a timestamp on it.
- It does not fix the underlying problem. If the complaint is that nobody answers the phone, a flawless reply at 23:42 changes nothing.
- Suing is not the backstop people think. Under section 1(2) of the Defamation Act 2013, a business trading for profit has to show the publication caused or is likely to cause serious financial loss. A single bad review rarely clears that bar — and an aggressive AI-drafted reply calling the reviewer a liar is itself a publication.
- Labelling it as AI will not save you, and probably is not required. Article 50 of the EU AI Act, which applies from 2 August 2026, covers AI-generated text published "for the purpose of informing the public on matters of public interest" (the article). A commercial reply on a Google listing does not sit there, and the UK has no equivalent duty. Check your own position — but the risk here is in the content, not the label. We wrote about Article 50 and chatbots here.
The short rule
Let the AI draft. Do not let it publish. No names, no dates, no services, no figures, no compensation. Two sentences and take it offline. And a mandatory human queue for anything touching health, money, staff or lawyers.
Zatio does not answer Google reviews. We live on WhatsApp and website chat, and we would rather say so than sell you something we do not do. What we do see daily is where a lot of one-star reviews come from: a weekend message nobody answered. That part we can help with — here is the product.
This is general information, not legal advice. For compliance decisions about personal data, speak to someone who can look at your specific case.