How to prepare your product catalogue for AI
An AI assistant answers about a product only as well as the data it reads. Prepare one reliable source: product sheets, care instructions, real customer questions and where to buy. An assistant that improvises fills the gap with videos, resellers and old PDFs.
Why assistants get products wrong
Because nobody gave them the approved answer. When the spec sheet does not cover it, the assistant may take the data from videos, resellers or old PDFs. Care instructions are the clear example: a cleaning video can contradict the manufacturer’s guidance and the AI cannot tell.
The assistant answers with the same confidence whether it pulled the data from your sheet or from a random video, and that uniformity misleads. A generally correct answer can be wrong for your specific product.
A wrong care tip spoils the product and the customer blames the manufacturer, not the assistant. An answer about compatibility that nobody reviewed ends in an order the factory cannot build.
The fix is not to tweak what the assistant says but to give it a place to read from: your names, your valid combinations and your care guidance.
One source of product data
Put names, references, sizes, materials and finishes in one place and let the website, sales channels and the assistant read from it. Two parallel lists drift apart. A 3D model with stable material names helps keep sizes and finishes consistent.
The source does not have to be a new tool. It can be the factory master sheet, tidied up, with the commercial name next to the internal code. If every channel keeps its own version, pick one as the origin and derive the rest.
3D fits for a practical reason: when the material in the model carries the real finish name, the same library serves the configurator, the showroom and the sheets. A material change is applied once, in one place.
What a product sheet needs
The sheet is the smallest unit the assistant can read and cite. Seven blocks cover most manufacturers: commercial name with reference, sizes and weight, materials and finishes, compatible combinations, care, publicly stated warranty and lead times, and where to buy. What is not in the sheet must not come out of the assistant.
Short, literal phrasing works better than prose: "dry clean only", "fabric not rated for outdoor use". A long sentence with caveats gives the assistant room to pick the part that does not apply. If a piece of data only holds for one channel, the sheet says so.
Warranty and lead times go in only when they are public. Where they vary by channel or volume, they stay out: the assistant does not promise what does not depend on you and leaves that figure to the sales network.
| Data | Fictional example | Why it matters |
|---|---|---|
| Commercial name and reference | Litoral armchair · Ref. LT-20 | The assistant cites the exact reference, not a description |
| Sizes and weight | 80 × 74 × 92 cm · 14 kg | Fit answers without working them out live with the client |
| Materials and finishes | Acrylic · 4 lacquers · 12 fabrics | Stable names that match the website and the factory |
| Compatible combinations | Chaise longue not available in fabric X | Stops it proposing configurations the factory cannot build |
| Care and maintenance | Upholstery, dry clean only | The data most often contradicted when it comes from outside |
| Warranty and lead times | 5 years on frame · 6 weeks | Only where public; otherwise it belongs to the sales team |
| Where to buy | Distributor by area · store | Handover needs a destination, not just an answer |
Collect real customer questions
The questions the sheet does not answer already sit in your support inbox, your reps, your distributors and your store staff. Gather them, phrase them as questions and write short answers. Someone who knows the product reviews them before publishing as FAQs.
The best source is the conversation that already happened: the email asking the same clarification, the doubt the distributor settles by phone, the misunderstanding repeated every season. Each one is an FAQ line and a line the assistant will no longer improvise.
Answers are written to be read, not to impress: two or three sentences, no internal jargon. If the question cannot be answered without commercial context, it does not go into the public FAQ; it becomes material for the sales team.
Care and limits: what the assistant may say
The assistant answers only from approved documentation. When something is not documented, it says so and hands over to a person. It never improvises on safety, care or compatibility: these are the three areas where an invented answer spoils a product, voids a warranty or creates an impossible order.
Saying "I do not know" is a feature, not an annoying limitation. An assistant that admits the gap and offers contact with the factory or the distributor keeps trust; one that fills the silence loses it the first time it is caught.
Approved documentation needs an owner: having the sheet is one thing and remembering to update it when care guidance changes or a new finish arrives is another. The review is tied to the catalogue, not to inspiration.
Send the customer to the store
The assistant can end the conversation with a commercial step: the nearest distributor or store, with the product and finish already identified. That keeps the sale in your network and makes the assistant work for it instead of competing against it.
Explain it to your network before turning it on. A distributor who discovers the site has been routing requests without warning sees a competitor; the one whose area is in the routing from day one sees a source of contacts. The context travels with the customer: product, finish and the question they asked.
The assistant does not close the sale and should not try to. Its job ends when someone in your network can pick it up with all the information on the table.
Make it readable for search and AI engines
Key data belongs in the page HTML, not only in PDFs or images. Add schema.org structured data, such as Product and FAQPage, and use the same product and finish names across pages. None of this guarantees being cited, but it helps engines read your product correctly.
The practical rule: if a piece of data only exists inside an image or a downloadable PDF, an assistant that reads the web may not find it. Sizes, materials, care and questions go as text. The PDF is for printing and sending, not the single copy.
Structured data describes what the page already shows; it does not invent content. And repeated names do the rest: if the site calls it "Litoral", the distributor calls it "LT-20-A" and the sheet spells it differently, no machine can join them.
How to tell if it works
Four signals are enough: questions answered without a person, questions left with no answer, wrong answers caught in review and enquiries handed to stores. We do not promise conversion percentages: they depend on the product, the channel and the sales force.
- Product questions solved without a person involved.
- Questions left with no answer: those are incomplete sheets.
- Wrong answers found when sampling conversations in review.
- Requests that reach stores and distributors with full context.
Frequently asked questions
Does this guarantee appearing in ChatGPT or AI Overviews?
No. Preparing the data is the precondition: without a reliable source none of this works, but no formula guarantees a citation. Engines choose what to cite and that is not controlled from your website.
Do I need to train my own model?
Usually not. Most assistants retrieve the answer from your approved documents, and that is easier to keep current than a custom model: you change the sheet and the change becomes available.
Can I start with one product family?
Yes. It is more sensible than starting with the full catalogue: prepare one family properly, see which questions get no answer and expand with what you learn.
Do I need 3D models?
Not for the assistant. They help because a model with stable material names keeps sizes and finishes consistent across the site, the sheets and the configurator, but prepared data works without 3D.
Who should review the answers?
Someone who knows the product: technical support, the engineering office or the rep who resolves those doubts today. And they are reviewed again when the catalogue changes, because an old sheet produces old answers.
