AI-drafted replies, edited and sent by you.
One click drafts a professional, on-brand reply to any review — public or private. You read it, you edit it, you send it. The model never auto-publishes; the owner always stays in the loop.
How it works
Starvo uses Llama via Groq as the reply model. When you click Generate reply, Starvo sends the model the following structured prompt — and only the following:
- The customer's rating (1–5 stars).
- The customer's comment (if they left one).
- Your business name and type.
- Your active reply templates (if you have any).
- A short system prompt that instructs tone (empathetic, brief, on-brand, no fake promises).
The response comes back in 2–4 seconds. The model never sees customer email addresses, your other reviews, your other businesses, or any data outside the review at hand.
Generate reply lives on every review row in the Inbox — expanded or collapsed.
Using the feature
1. Open the review in your dashboard
Open your dashboard and click any review row in the feedback list to expand it.
2. Click Generate reply
A draft appears in seconds. The model uses rating, comment, your business name, type, and any active templates.
3. Edit if you want
The textarea is editable. Soften, sharpen, add a name, fix a fact.
4. Send
Email the customer (private feedback) or publish to Google (when GBP is connected).
The draft is always editable in a textarea with a character count. Send is one button. Tweak before publishing.
The shared reply quota
Starvo's monthly reply limit is a shared pool. It covers AI-drafted replies, manually-typed replies, template replies, and (on Starter and above) replies published to Google. The limit is per location, and it resets every month on your billing anniversary — the same day of the month your business started. A business created on the 16th resets on the 16th of each month, not on the 1st.
| Plan | Monthly replies / location |
|---|---|
| Free | 20 |
| Starter | 50 |
| Pro | 500 |
| Pro Max | 2,000 |
The pool design is intentional. AI replies as a separately metered add-on creates the “quota surprise” problem — owners hit the AI limit, drop back to writing by hand, and the team disengages. One pool, all reply types, predictable.
Your remaining quota is visible in the dashboard. When you hit the limit, the Send button disables until the next reset or upgrade.
Tone and templates
The model tunes tone based on three signals:
- Star rating. 5-star drafts are warm and brief; 3-star drafts are curious and apologetic; 1-star drafts are calm and offer offline resolution.
- Business type. A restaurant reply reads differently from a clinic reply — different vocabulary, different default formality.
- Active templates. Templates you save under Dashboard → Settings give the model a baseline voice to imitate. A salon with a casual, first-name-basis template will produce drafts that sound the same.
What the model knows (and what it doesn't)
The model only sees the inputs listed in the “How it works” section above. It does not see:
- The customer's email address.
- Any of your other reviews (it will not write “we've been hearing this a lot”).
- Your other businesses or other locations.
- Any operational data — your reservations, your booking system, your POS.
Drafts are never auto-sent. The model never sees the final edited reply you ship — only the input. Replies are also never used to train external models. Groq operates on a zero-retention basis for the requests Starvo sends through.
Limitations
- The model hallucinates specifics. If the review says “the pasta was cold,” the draft will not invent the chef's name or the table number. But it may guess a dish type that wasn't mentioned. Always read before sending.
- It will not write defensive replies. Even when the customer is wrong, the draft will be calm and offline-redirecting. That is a deliberate constraint.
- It does not detect fake reviews. If the comment looks like obvious spam (e.g. “asdfg”), the draft will still try to respond politely. Use the review's Delete / Flag actions for obvious spam.
- It is English-first. Drafts in other languages work but are less reliable. Manual replies are the safer path for non-English customers right now.
Troubleshooting
Generate reply is greyed out
You are out of monthly replies for that location. Wait for the quota to reset on your next billing-anniversary day, upgrade to a higher plan from Dashboard → Billing, or reply manually — manual replies count against the same pool but the button is still active until you cross the limit.
The draft sounds robotic
Most often, your templates are empty. Add 2–3 templates under Dashboard → Settings → Templates in the voice you want the AI to imitate. The next generation will lean on those.
The draft repeats the customer's words back at them
Symptom of a very short review (e.g. “good”). The model has nothing else to work with. Edit before send, or skip the AI on one-word reviews and just send a thumbs up reply manually.
Generation fails with an error
Rare. Usually a transient Groq outage. Retry once. If it persists, manual reply still works.