We run an SEO agency, and a lot of our clients are clinics and law firms. When we started letting AI draft their Google review replies, the drafts were usually fine. The problem was the few that were not: a reply that thanks a patient by name, mentions their treatment, or tells a reviewer their case went well. For a clinic or a law firm, one of those is a real problem, not a typo.

So when we built OpenGMB, our open-source Google Business Profile manager, we put a rule check between every draft and Google. Here is how it works.

Every draft takes the same path

Every reply, post and profile edit passes a set of rules and waits for a person to approve it before Google sees it. It does not matter who wrote the draft: a person, a template, the built-in AI, or Claude over MCP. In the README's words, the MCP server's writes "go through the same guardrails and approvals as the web app", and a human-in-the-loop token "can draft and submit, but it cannot approve or publish."

The rules run twice: once when the draft is created, and again just before anything is sent to Google, so a late edit cannot skip them.

39 rules, five packs, no AI

OpenGMB ships 39 rules in five packs. They are plain pattern checks, with no AI. That was deliberate: we did not want one model checking another model's homework.

  • Universal (10 rules) runs on every reply. Links, email addresses and hashtags are blocked, and so are offers in replies. It also checks length, repeated replies and the writing style.
  • Healthcare (11 rules) blocks a reply that confirms a care relationship, names the reviewer, a provider or a staff member, or mentions a condition, treatment or date. A reply to a 1 to 3 star review must carry a fixed privacy line.
  • Legal (4 rules) blocks a reply that confirms a client relationship, discusses case facts or promises results, and warns about testimonial wording. To be clear: the legal pack's wording is still waiting for a lawyer's review.
  • Posts (7 rules) and Profile (7 rules) check things like phone numbers, length, hours and addresses on posts and profile edits.

Blocks and warnings

Each rule is either a block or a warning. A block stops the draft, and nobody can approve it until the text changes. A warning does not stop anything, but someone has to acknowledge it before the draft can go on. On a healthcare client, an owner or admin can turn the name, condition and date rules into warnings. Everything else stays as it is.

Who approves what

  • Posts and profile edits always need a person's approval.
  • Replies to 1 to 3 star reviews always need a person and are never published automatically.
  • Replies to 4 and 5 star reviews need a person, unless the owner turns on auto-approve for that client.

Why this matters if you are wiring AI into local business tools

If you connect an AI agent to anything that publishes in a business's name, the useful question is not "is the model good?" but "what happens when it is wrong?" Our answer is boring on purpose: fixed rules, a person who signs off, and the same checks again at the last moment.

The full list of packs and what each one checks is on the OpenGMB guardrails page. The community edition is MIT licensed and self-hosted with Docker Compose. OpenGMB is not affiliated with or endorsed by Google.

Written by the OpenGMB team at Synectus.