A shopper types "do you have this in blue, size medium, and will it get here before the weekend?" That's three separate facts packed into one sentence: color availability, size stock, and delivery time. A chatbot can only answer it if it genuinely knows your catalog. Train it well and questions like that get a real answer in seconds. Train it badly and it either guesses or punts, and the sale cools off.
Getting a bot to know your products isn't magic. It's mostly about feeding it the right data in a shape it can use, then keeping that data honest as things change.
Decide what the bot needs to know
You don't have to pour your entire back office into the bot. Start with the facts customers actually ask about before they buy.
- Product names and the plain-language terms people use for them
- Descriptions, materials, dimensions, and what's in the box
- Prices, including any variant pricing
- Variants: sizes, colors, styles, and which combinations exist
- Stock status, at least roughly
- Shipping options, costs, and delivery estimates
- Return and warranty policies
That list covers the vast majority of pre-purchase questions. You can always add more later, but this is the core a shopper needs to feel safe clicking buy.
Structured data beats a wall of text
There are two ways to teach a bot your catalog, and the difference matters.
One is to let it crawl your storefront and read the pages like a visitor would. That's fine for descriptions and general info. It struggles with precise, structured facts, because prices and stock buried in page layout are easy to misread.
The better way for catalog data is to feed it a clean, structured source: a product feed or an export with one row per product or variant, columns for price, SKU, stock, and attributes. Structured data leaves far less room for the bot to guess wrong. In SpideyChat you'd combine both, crawling the site for the descriptive content and connecting or uploading a product export for the hard facts like price and availability.
Be honest about variants and stock
Stock is where bots get people in trouble. Nothing sours a customer faster than being told an item is available, ordering it, and getting a "sorry, out of stock" email an hour later.
If your bot has live access to inventory, great. It can say "the medium in blue is in stock, two left." If it's working from a snapshot that updates once a day, it should hedge honestly: "That was in stock as of this morning. Want me to have someone confirm before you order?" A bot that admits the limits of what it knows keeps trust intact. One that states day-old stock as gospel generates returns and refunds.
The same goes for discontinued items and pre-orders. Tell the bot about them so it doesn't cheerfully sell something you can't ship. It's worth deciding this policy before launch, not after the first complaint, because the returns and refunds from overpromised stock cost far more than the honesty ever will.
A practical training order
Here's a sequence that works, rather than trying to do everything at once.
- Export your catalog to a clean file: one row per sellable variant, clear columns for price, stock, and attributes.
- Add descriptive content by crawling your product pages or uploading your descriptions.
- Write out your shipping, return, and warranty policies as short, plain answers.
- Add a handful of question-and-answer pairs for the quirky things customers ask that aren't written anywhere, like "are these true to size" or "is the packaging plastic-free."
- Set a handoff rule so anything about a specific order, a defect, or a refund goes to a human.
- Test, fix the gaps, and only then widen the scope.
Doing it in that order means the bot is accurate on the basics before you ask it to handle edge cases.
Keep it in sync, or it'll lie for you
A catalog isn't static. Prices change, styles sell out, new items land, seasonal products retire. A bot trained once and forgotten slowly becomes a liar, confidently quoting last month's price on a product you've since raised.
Decide how fresh the bot's data needs to be and set up a refresh to match. A store with stable pricing might re-sync weekly. One running frequent promotions needs its price and stock data connected as close to live as possible. Whatever the cadence, make it a real process, not a thing you remember to do when a customer complains.
Test the way shoppers actually type
Once it's trained, don't test it with tidy questions. Test it the way real people write.
A quick before-and-after makes the point. Take Riverbend Outfitters, a fictional gear shop. Early on, their bot could answer "What is the price of the Summit 40 backpack?" perfectly but choked on "how much for that 40l pack, and does it fit a laptop?", the same question in shopper language. The fix wasn't more product data. It was adding the casual phrasing and the laptop-fit detail to its training. After that, the messy version worked too.
Run your bot through the real questions your team sees in chat and email. Typos, slang, two questions mashed together, vague references like "that blue one." Every place it stumbles is a small, fixable gap.
This points to a bigger truth about catalog bots: your catalog speaks in product names and specs, but customers speak in problems and use-cases. Someone shopping for a gift types "something for a coffee lover under $40," not "pour-over kit, SKU 2231." A parent types "a jacket warm enough for a New England winter," not "650-fill down parka." If your bot only knows the catalog's language, it stumbles on these even though you clearly sell what they want.
Close that gap on purpose. Alongside the product data, feed the bot the categories, use-cases, and plain-language descriptions that map casual requests to real products. A handful of question-and-answer pairs does a lot of work here: "What do you recommend for a beginner?" paired with your honest answer, "Do you have anything waterproof for hiking?" paired with the right line. You're teaching it how people actually shop, not just what's on the shelf. And every conversation where a shopper describes what they want in words the bot didn't recognize is a phrase worth adding. A few weeks of that light tending, and the bot gets noticeably better at turning "I need X for Y" into "here's the exact product," which is the moment discovery becomes a sale.
A chatbot that knows your catalog cold turns product questions from a reason to leave into a reason to buy. Feed it structured data for the hard facts, keep that data fresh, and test it against the way people actually shop. Do that, and "do you have this in blue, medium, by Friday?" gets the confident, correct answer that keeps the sale alive.