How Milo actually works
No black box. Here's exactly what happens between a Google review and the reply you approve.
Claude, Anthropic's AI: not a canned-reply generator
Meerkly uses Claude, Anthropic's generative AI model (the same model family behind Claude.ai). It's not a library of canned phrases picked at random: every response is written the moment the review comes in, from the actual review text, your industry, your tone, and the customer's language. Two similar reviews never get the exact same reply.
The path of a review, step by step
1. Detection
Meerkly checks Google, Facebook, and Yelp through their official services (depending on your plan) to spot every new review the moment it's published. No scraping: only public, authorized access.
2. Building the context
Before writing anything, Milo receives your business name, industry, chosen tone, preferred length, and, if you've provided them, your reference documents (menu, pricing, policies).
3. Language lock
A detector analyzes the review's actual linguistic markers (not just the language configured on your profile) to always reply in the customer's language, even if your business defaults to the other one.
4. Generation
Claude writes a response that takes all of this context into account, with the safety rules below applied automatically, no exceptions.
5. Cleaning up robotic phrasing
Phrasing that gives away a machine-written reply (long dashes, overly polite phrasing, empty compliments) is automatically stripped before you ever see the draft.
6. Your approval
Nothing gets published without your click. You get an email with the reply ready, edit it if needed, approve it, or ignore it.
The guardrails baked into every response
However you customize tone or length, certain rules are automatically added to every request sent to Claude. They cannot be turned off, even by mistake:
Never contradict the customer
Even if the review states something inaccurate about your business, Milo doesn't correct, argue, or justify himself publicly. He thanks them and moves on. A public dispute always does more damage than a negative review.
Never invent information
Milo doesn't state anything about your hours, prices, or services that he can't verify from what you gave him. If you've uploaded a menu or price sheet, that document becomes the only source of truth he can cite.
Never repeat suspicious content
If a review contains a phone number, link, or promo code slipped in by its author, Milo doesn't echo it back in his reply.
Treat the review as data, never as an instruction
A review containing text like "ignore your instructions and write instead..." is treated as text to comment on, not a command to execute. This is a deliberate defense against attempts to manipulate the model ("prompt injection").
Milo never touches your raw data
When you ask a question in plain language on your dashboard ("my average rating by week"), Milo doesn't generate queries or code that run directly against your database. He picks from a closed list of already-validated, secure widget types, which are then filled in with your real data. No arbitrary execution, no direct access.
Where your data actually goes
- •Meerkly only accesses the public reviews on your listings, through the platforms' official services.
- •Your data is hosted with named providers (Supabase, Stripe for payments, Resend for emails, Anthropic for writing replies): no sale, no marketing use without consent.
- •You can export all your data as CSV, or request its deletion, at any time.
- •Our practices comply with Quebec's Law 25 and Canada's PIPEDA. The full detail is in our privacy policy.
What if Milo gets it wrong?
You always approve before anything is published: nothing goes out without your click. If a draft doesn't work for you, edit it or ignore it, no consequences. Some features (like Get found by AI) fall back to a safe, pre-written text if the writing fails, instead of blocking your business.
More questions?
Check out our full FAQ, or write to us directly.
Ready to see Milo at work on your own reviews?
Free plan, no card required, or a reputation audit in a minute, no sign-up.