Get your catalogue ready for AI
- Version
- 3.x
- Updated
Audience: store owners and shop managers · Scope: preparing your catalogue so AI answers well, and choosing which AI surfaces to run · Last reviewed: 2026-08-28
Start here if you are new to the AI features. CBX has three of them and they all read the same thing: your catalogue. Time spent on the catalogue improves all three at once — and no amount of configuring the assistants makes up for product data that does not say much.
The three surfaces, in one table
| Who it talks to | Whose model | You should read | |
|---|---|---|---|
| AI product assistant | visitors on your configurator pages | yours, your bill | What the assistant is |
| Agentic storefront | your customers' own assistants, headless | theirs | Let customers use their own AI assistant |
| Your own assistant | you, in the back office | yours | Connect your own AI assistant |
They are independent. You can run one, two or all three, and turning one off does not affect the others.
Step 1 — write the product data (this is the whole game)
An assistant cannot invent facts about your product. Everything it says comes from what you have written, so the quality of your answers is decided here, before you switch anything on.
In rough order of how much difference each one makes:
- Fill in the detail panes. This is the single biggest win. Specifications, materials, delivery terms, warranty, what is in the box — everything a customer would ask a salesperson and cannot read off the configurator. See Product detail panes.
- Describe your questions. A question called "Lm/ft" means nothing to a customer. Its description is where you say what it is and how to choose.
- Describe your answers, especially where the difference between two options is not obvious from their names.
- Say what a product is for in its own description. "Who buys this and why" is what lets an assistant tell one of your products from another.
How to tell whether you have done enough: ask the assistant something a customer would ask, and see whether it answers or says it does not know. It is built to admit the gap rather than invent — and to file it on the wishlist, which turns into your to-do list for this step.
Step 2 — add advice where the data cannot say it
Some things are true about a product but are not in any field. A sensible order to ask questions in. Which answer usually decides the rest. A constraint your rules cannot express. The words your customers use that your catalogue does not.
Each product has an Advice for AI assistants field for exactly that (Add a product). It is read by assistants and never shown to customers.
Keep it to what is specific to this product. Assistants already know how a configurator behaves in general, and repeating that crowds out the part only you know.
Most products need nothing here. A product whose questions and descriptions are well written is already fully explained. Reach for this field when you catch an assistant getting something reasonable-but-wrong, and write the sentence that would have prevented it.
Step 3 — choose your surfaces
The AI product assistant is the one most stores want first: visitors get help without leaving the page, and you find out what they ask. It needs an API key and is off until you set one — Set up the assistant.
The agentic storefront is already on. There is nothing to install and nothing to switch on; a customer points their own assistant at your store's address and approves it in their browser. You may want to narrow what it can do — see Let customers use their own AI assistant, which also covers what customers see and how they turn one off.
Your own back-office assistant is for building and running the store rather than for selling — Connect your own AI assistant.
Step 4 — set the house rules
Both customer-facing surfaces take house rules: your standing instructions, in your words. Shop facts, what to emphasise, what never to promise. They are separate settings because the two surfaces are separate audiences, and a rule that makes sense on your own panel ("offer the extended warranty") may not be one you want given to a customer's own assistant.
Write them as plain sentences, one per line. Two that most stores want:
- Something you never want promised — a delivery date, a discount, a compatibility claim.
- How to handle "can you do X?" when the answer is no.
Step 5 — watch what actually happens
The assistant keeps a record of every conversation, and it is worth reading in the first weeks:
- The conversation journal — what was asked, and what it answered.
- The wishlist — the questions your catalogue could not answer. This is step 1's to-do list, written for you by your visitors.
- Conversation Insights — the patterns across many conversations.
Customers with a connected assistant of their own get their own version of this on their account page: what it has done, and a button to turn it off.
A note on languages
If you run more than one language, the things a person reads are per language — the assistant's greeting, and your extra note on the consent screen. You will see one input per language on those.
The things a model reads are not: house rules, and the per-product advice. An assistant applies instructions written in one language perfectly well while talking to a customer in another. If you want it to answer in the customer's language, write that as one of your house rules.