Better AI Replies
A good reply to an ad comment answers the actual question, in a sentence or two, in your voice, with a next step. Three things decide whether Rypl’s replies get there: the knowledge base, the instructions and the rules.
1. The knowledge base decides accuracy
The agent can only state facts it has. Most weak replies trace back to a missing or outdated fact.
- Start with Q&A for the 20 to 50 questions you get most, written exactly as you want them answered.
- Add prices, variants and offers. “How much?” is the most common question under ads.
- Add shipping and returns, by country if they differ.
- Add your website for everything else; links in replies point there.
- Update the same day something changes. The knowledge base learns when you save.
2. Instructions decide the style
- Length: “Two sentences at most for simple questions.”
- Next step: “End with the product link when someone asks about a product.”
- Words: “Say ‘pieces’, never ‘items’.”
- Limits: “Never promise delivery dates.” “Never discuss pricing for wholesale; say a teammate will reply.”
- When unsure: “If the answer isn’t in the knowledge base, say a teammate will reply.”
Pick the tone that fits: Neutral, Seller, Positive or Direct. Most consumer brands do well with Neutral or Positive; Seller works for promotions; Direct suits brands with a clipped voice.
3. Rules decide what gets a reply (Pro and Ultra)
- Add My text replies for questions asked word for word again and again, so they’re answered the same way every time.
- Turn on thank-you replies to answer praise.
- Use Don’t reply for topics a person should handle in private.
- Turn on Read the post so “how much is this?” gets the price of the product in the ad.
- On Ultra, Remember the conversation lets the agent pick up where an earlier conversation with the same person left off.
4. Improve it every week
- Read a sample of replies in the inbox: are the facts right, is the tone right, is there a next step?
- Run Suggest improvements on the agent. It reads the questions customers really asked, answered and unanswered, and proposes Q&A and instruction edits, leaving gaps in [brackets] where only you know the answer.
- Check the suggestions above the agent’s settings: unanswered complaints, repeated questions and emoji reactions your rules don’t cover.
- Fix the source (knowledge base or instructions), never just the one reply.
Signs something needs fixing
| What you see | Likely cause | Fix |
|---|---|---|
| “Check our website” instead of an answer | Fact missing | Add it as Q&A |
| Wrong price | Outdated document | Update or remove it |
| Too long | No length instruction | Add one |
| Wrong product | The post wasn’t read | Turn on Read the post (Pro and Ultra) |
| Wrong formality in German or French | No instruction | Say du/Sie, tu/vous |
| Same reply on every emoji | Broad rule | Add several My text variants or narrow the rule |
More on rypl.is
- Why generic auto-replies hurt your brand: what context-aware replies change
- AI vs rule-based comment automation: what actually works on Facebook
- How to auto-reply to Facebook comments: Meta’s own options and where AI replies fit
- How to auto-reply to Instagram comments: Instagram’s own options and where AI replies fit
| Back to Best Practices | Back to Documentation | Visit Rypl |