Best practices

Learn the habits that separate a good bot from a great one — before you invest time building the wrong thing.

Training data

Start narrow, go broad. Add your most frequently asked pages first: FAQ, pricing, product details, return policy. Once those are solid, add the rest of your site.

Structured content beats prose. A well-formatted FAQ with clear Q&A pairs produces significantly better answers than unstructured blog posts. The bot can pattern-match on explicit question-answer structure.

Keep sources fresh. When your pricing changes or you update a policy, retrain the relevant source. Stale content causes incorrect answers, which erode visitor trust.

Use plain language in your content. If your website uses industry jargon without explanation, the bot will too. Jargon-free source content produces jargon-free answers.

Bot configuration

Write a tight system prompt. The system prompt is your clearest lever. Tell the bot: what it is, what topics it can and cannot discuss, how to handle questions it cannot answer, and what tone to use. A two-paragraph system prompt beats a vague one-liner.

Set the tone deliberately. "Friendly" sounds warmer but can feel unprofessional in B2B contexts. "Professional" is safe but can feel cold in consumer contexts. Test both with real visitor questions before deciding.

Keep the welcome message short. Visitors see it first. A one-sentence welcome that tells them what the bot can help with outperforms a long paragraph every time.

Use suggested replies strategically. Add 3–6 suggested replies that represent the questions you most want the bot to handle. This guides visitors toward topics where the bot is strongest.

Monitoring & improvement

Review conversations weekly. The Conversations log shows exactly what visitors asked and how the bot responded. Look for patterns: repeated questions the bot couldn't answer, incorrect answers, or off-topic responses.

Act on self-learn suggestions. After every human handoff, DGbot extracts potential FAQ improvements. Review these weekly in the Self-learn section. Approving good suggestions is the fastest way to improve answer quality.

Watch handoff rate. A high handoff rate (>20%) usually means there are important topics the bot cannot handle. Look at which conversations get handed off and add content that would have prevented the escalation.

Track credit usage. Advanced LLM models (GPT-4o, Claude Sonnet) cost more credits per message. If you're on a credit-limited plan, consider using Economy tier (GPT-4o mini, Claude Haiku) for most conversations and reserving the advanced model for complex queries.

Common mistakes to avoid

Adding too much content too fast. More content doesn't always mean better answers. Low-quality or contradictory content confuses the retrieval system. Add good content gradually and verify each batch.

Skipping the preview. Always test in the preview panel before going live. What looks good on paper can behave unexpectedly in conversation.

Not setting escalation rules. If the bot has no handoff configured, frustrated visitors have no escape route. Even a simple "I'll connect you to our team" link is better than nothing.

Ignoring the analytics. The bot's most common unanswered questions are the clearest signal of what content to add next.

Making the welcome message a wall of text. Visitors close long welcome messages. Keep it to one sentence. Save the detail for the system prompt.