Train your AI chatbot on store policies and FAQs

Your catalog is the AI’s product encyclopedia. The knowledge base is everything else - shipping policy, returns, sizing guides, gift cards, store hours, sustainability claims, brand story. Documents are embedded the same way products are, and the AI retrieves them on every relevant question. It lives at Catalog → Knowledge in the sidebar.

Every document has one of three types - Policy (shipping, returns, warranty - the trust layer), FAQ (common questions with concise answers), and Page (about, contact, brand context) - and the page groups them into those sections.

Ways to get content in

Best-practice content

Start with authoritative answers to the questions your team actually receives. Examples to adapt to your policies:

• Returns - “What is the return window? How do I start a return? Who pays for return shipping?”
• Shipping - “Do you ship to PO boxes? International shipping countries? Cutoff times for same-day dispatch?”
• Sizing - “How does your sizing run? Is X tagged true-to-size? What if I’m between sizes?”
• Gift cards - “Do you sell gift cards? Are they combinable with codes? Do they expire?”

The self-learning loop

Knowledge suggestions start from real shopper questions and require review. When the live AI hits a question it can’t answer well, the prompt lands in Knowledge Gaps; one click drafts a grounded answer, and the draft appears back on the Knowledge page in the AI-suggested articles panel. Approve publishes the reviewed entry and queues its embedding job; it becomes searchable after that job succeeds. Discarddeletes the suggestion. Review the draft and replace placeholders before approval. Publication is not evidence that indexing or answer quality has already been verified.

Auditing what the AI actually uses

The Most referenced by the AI panel ranks documents by how often the AI cited them when answering shoppers, with a total citation count and each document’s last-retrieved date. Keep those top documents accurate and fresh - they do the most work. For a single conversation, the detail view under Inbox → Conversations shows the tool calls and sources behind each reply.

Documents untouched for 90+ days surface in the AI Improver inbox as stale-doc reminders, so quietly outdated policies don’t linger.

Editing and removing documents

Every document can be edited in place (title, type, content - edits re-embed automatically) or deleted with an inline confirmation. If a synced page shouldn’t be retrievable - a legacy returns policy, an internal memo - remove it here and fix the upstream source or narrow the next import. Deletion is not a persistent source exclusion: a later platform sync, URL import, or sitemap import can recreate it. Do not re-import an obsolete source while relying on its local deletion.

How to add your first FAQ

You need a team role with catalog access to create, edit, or delete knowledge documents. Use this for approved shopper-facing information; keep internal-only notes and unapproved policy drafts out of the agent’s sources. If the controls are unavailable, ask a workspace owner to check your role rather than sharing their account.

1. Open Catalog → Knowledge and click Add document.
2. Set the type to FAQ, give it a question-shaped title (“How do returns work?”), and write the answer the way you’d say it to a customer. Question-shaped titles retrieve better because they match how shoppers actually ask.
3. Save, then allow the queued document-embedding job to complete. A saved document is not necessarily searchable yet.
4. Verify by asking the widget the question yourself and inspecting the conversation source. Use Most referenced by the AI to monitor later usage; citation counts do not by themselves prove accuracy.

Refresh and removal responsibilities

Worked example and troubleshooting

Illustrative FAQ: “Can I return an opened item?” If the approved policy only allows unopened returns within 30 days, write that condition and the contact route. Test an unopened item inside the window, an opened item, and an item outside the window. The answer should preserve each condition, cite the approved document, and hand off an exception request instead of inventing an exception.

If the entry saves but cannot be found, check whether indexing completed and review indexing troubleshooting. If an old answer returns, inspect duplicate imported documents and the original source. Keep private operational notes and unapproved legal or medical claims out of shopper-facing knowledge.

Adjacent pages

• Training & knowledge gaps - turn missed questions into new KB entries.
• System prompt - define brand voice and rules around the KB content.
• Security & privacy - how KB content is stored and accessed.