How to train an AI chatbot on your product catalog

Say a kurti page in Mohammadpur sets up its first chatbot by pasting the page’s About text into it, then expects it to sell 300 kurtis. A customer asks "eta M size e ache? dam koto?" and the bot replies with a polite paragraph about quality fabric. No price, no size, no sale.

The bot is not stupid. It has simply never been given the products. This guide shows how to give an AI chatbot your catalog the right way, so it answers the questions that actually make money: price, size, colour, stock and "send pic".

What does ‘training’ a chatbot on a catalog actually mean?

Training a chatbot on a catalog means connecting it to your product data so it can look products up, not teaching the AI model new knowledge. The AI already understands language. What it lacks is your prices and stock, and those change every week.

Modern shop chatbots use a method called retrieval-augmented generation. When a message arrives, the chatbot searches your product list and other content, picks the few items that match, and writes a reply using only those. Some go further and use tools: a "search products" step that returns live price, stock and options before the AI is allowed to quote anything.

That has a practical meaning for you. You will not "upload once and forget". You keep the catalog current, and the chatbot stays current with it.

What product information does the chatbot need?

The chatbot needs the same information a good salesperson would check before answering: what the item is, what it costs, which options exist, whether it is in stock, and what it looks like. The table shows each field and the customer question it answers.

You do not need all of this on day one. Name, price, options and stock are the minimum. Photos come next, because in Bangladesh most buyers decide from the picture.

How do I clean my catalog before connecting it?

Clean your catalog by making each product one real item with clear options, a plain name and current numbers. An hour of cleaning saves weeks of wrong answers.

Work through this list:

1. Merge duplicates. If "Cotton Panjabi Black" and "Black Cotton Panjabi" are the same item, keep one.
2. Use options for sizes and colours. One product "Cotton Panjabi" with Size: M, L, XL and Colour: Black, Maroon is much easier for the AI than six separate products.
3. Write names customers would say. "Women's Lawn Three-Piece, Printed" beats "LP-23 New". Keep your code in the SKU field.
4. Fix the prices in one place. If your Facebook post says ৳1,450 and the catalog says ৳1,390, the chatbot will quote the catalog, and the customer will screenshot the post.
5. Set stock honestly. If you sell from a shared stock with a physical shop, update the online stock at least once a day during busy periods.
6. Remove dead products. Items you will never restock should be archived, not left at zero stock with an old price.

How do I connect the catalog to the chatbot?

Connect the catalog by syncing your online store or by adding products to the chatbot tool directly, whichever keeps the data most current with the least effort.

The trap for page-only shops is keeping two lists. If the team updates the Excel sheet but not the chatbot, the bot is wrong by Friday. Pick one place to edit prices and treat it as the truth.

How do I teach the chatbot how to sell, not just what I sell?

Teach the chatbot how to sell with a short written guide: a few rules about recommendations, promises and handover, in plain language. The catalog says what you have. The guide says how to behave.

A good guide for a clothing page fits in ten lines:

- Recommend only products from our catalog. If we don't have it, say so politely.
- When someone asks for a colour or size we don't have, suggest the closest item in stock.
- Never promise a delivery date. Say: inside Dhaka usually 1-2 days, outside Dhaka 2-4 days.
- Never offer a discount that is not written in our offers.
- Bulk orders (10+ pieces), custom stitching and complaints: hand over to the team.
- Keep replies short. Send a photo when the customer asks to see an item.

Add your delivery charges, exchange policy and payment method as separate entries in the chatbot's knowledge, not inside product descriptions. Product text should describe the product. Policies change on their own schedule.

Will the chatbot understand Bangla and Banglish product questions?

Yes, a capable AI chatbot understands Bangla and Banglish questions and can match them to English product names. "Kalo panjabi XL ache?" can find "Black Cotton Panjabi" with option XL.

You can make matching more reliable:

• Use the plain English or Bangla word your customers use, not an internal code, in the product name.
• Put common alternative words in the description: "three-piece, 3 piece, salwar kameez".
• For local terms with no English equal, such as "jamdani" or "fatua", keep the local term in the name.

If many of your customers write in Bangla script, test with Bangla questions as well. A bot that is perfect in English and confused in Bangla is half a bot.

How do I test the chatbot before customers see it?

Test the chatbot with the real questions your page gets, in the way customers actually write them. Open last week's inbox and copy 20 questions word for word, typos included.

Include these on purpose:

• A product you sell, by its exact name.
• The same product described vaguely: "oi je blue ta, last week post korechilen".
• A size or colour that is sold out.
• A product you do not sell at all.
• A price trap: "last time apni 900 bolechilen".
• A request that needs a person: "50 piece lagbe office er jonno".

For each one, check three things: the product is right, the price matches the catalog exactly, and the bot hands over when it should. Fix the data first and the rules second. Most wrong answers trace back to a messy product entry, not the AI. If prices are your biggest worry, our guide on answering ‘price koto?’ automatically covers that question in depth.

How does SaleMag use your product catalog?

SaleMag's AI chatbot can answer from your whole catalog: products synced from a connected Shopify, WooCommerce or BigCommerce store, or products you add by hand. The knowledge base re-indexes when products change, so answers follow your current list.

With AI order taking switched on, the AI must search the catalog by name, SKU, category or brand before it states a price, option or stock level. It translates the customer's words to your catalog's language when needed, sends up to three product photos per reply, and never offers an out-of-stock item as available. Any money amount in its reply that is not a real catalog price is caught and rewritten. A test playground shows the sources behind each reply before you go live. The AI runs on your own OpenAI, Anthropic (Claude) or Google Gemini key.

Next steps

Pick your 20 best-selling products and clean them first: one item, clear name, real options, current stock and at least one photo. Connect them, write your ten-line selling guide, and test with last week's questions. When the answers are right, add the rest of the catalog. To try it on your own products, start free with SaleMag.