An AI shopping assistant visits your store the way a very fast, very literal clerk would. It does not admire your photography. It reads your pages, pulls out facts and tries to match them against what its user asked for. If the facts are there and consistent, your product can be recommended. If they are missing, the assistant quietly picks someone else. This article is a working checklist for making a catalogue that such an assistant can read.
Start with the facts a buyer would ask about
For each product type you sell, list the questions a serious buyer asks before paying. For a pair of shoes: size range, width, upper material, sole material, colour, weight, price, stock by size, delivery time, return terms. For a bag of tea: grade, origin region, net weight, packaging type, best before, price per pack.
That list becomes your standard set of attributes for that category. Every product in the category should answer every question on the list, in text, in the same order. Consistency matters more than eloquence. An assistant comparing ten shoes wants to find “Upper: leather” in the same place each time.
A useful test is to ask whether a shop assistant at your counter could answer a customer’s question by reading only the page. If they would need to go and check the item, the page is incomplete.
Move information out of images
Many local stores put size charts, ingredient lists and offers inside images because it is quick to design in a phone app. Humans can read those images. Most assistants and search engines cannot read them reliably, and even when they can, they are less confident about what they found.
- Type the size chart as a simple list or table in the description.
- Write the ingredients or materials as text, even if the photo of the label stays.
- State any offer (“Buy two, get free delivery”) in the page text, with its end date.
- Give each image descriptive alternative text, such as “navy linen shirt, front view”, not “IMG_4471”.
Name variants properly
Variants cause a lot of confusion. A shirt in four colours and five sizes is twenty sellable items. If your store treats them as one product with a note saying “message us for other colours”, an assistant cannot tell whether the blue medium is available.
Set up real variants in your store software, each with its own stock count and, where needed, its own price. Give colours plain names that people search for. “Maroon” is better than “Royal Berry Dream”. If you must keep a fancy name, add the plain one next to it: “Royal Berry (maroon)”.
Use one unit system and stick to it. If some products say “500g” and others say “0.5 kg” and others “half kilo”, an assistant comparing them has to guess whether they match. Pick a format and apply it everywhere.
Keep price, stock and delivery honest
An assistant that finds LKR 4,500 on your product page and LKR 4,950 in the cart will distrust your whole store. The same goes for items shown as in stock that turn out to be sold out after an order is placed.
- Price: show the final price including any compulsory charge. If delivery is extra, say how much, by area.
- Stock: keep counts current. If you cannot, show “made to order, ships in 7 days” rather than a false “in stock”.
- Delivery: give real time ranges by area, and say whether cash on delivery is available.
- Returns: write the actual terms in a sentence or two on every product page, or link to them clearly.
Add the hidden labels
Product pages can carry structured data: a block of labelled information, invisible to the shopper, that says “this is a product, its name is this, its price is this, it is in stock”. Search engines have read this format for years, and it is a natural place for AI assistants to look because it removes guesswork. Most modern store platforms can add it automatically once the product fields are filled in correctly, which is another reason to fill them in properly rather than stuffing everything into one description box.
Check that the structured data matches what the page shows. A mismatch between hidden price and visible price is worse than having no structured data at all.
What this will not fix, and how to begin
A readable catalogue does not make a poor product sell, and it does not replace trust. If your delivery is unreliable or your reviews are weak, an assistant may well find you and still rank you low. It also cannot help with items that are genuinely one-off, such as handmade pieces where each is different, unless you take the time to describe each one.
It is also slow, unglamorous work. Say you have 300 products and each needs fifteen minutes of cleanup. That is 75 hours. Spread over a month, it is manageable. Done in a weekend, it produces errors.
Begin with one category, ideally your best seller. Write its attribute list, fix every product in it, and check a few pages by reading only the text. An AI tool can help draft the cleanup quickly, as long as a person checks each result against the real item; the product description tool is built for that kind of first draft. Then move to the next category. By the end, your catalogue will be easier to read for machines and for the people who skim on their phones between other tasks.