Accurate AI Answers on Your Store, Grounded in Your Catalog

Most shop owners who hesitate about AI are not worried that it will be rude to customers. They are worried that it will be wrong. Politely, confidently wrong: recommending a product that does not exist, quoting a price from last season, or promising free returns you never offered.

That worry is justified. This article explains why made-up answers are the real risk of AI on a store, and how AskMerra keeps products, prices, stock and policies tied to your data.

Why wrong answers hurt more on a store

On a help site, a slightly wrong AI answer is an annoyance. On a store, it turns into a cost.

• An invented product sends the shopper looking for something you do not sell. They leave disappointed.
• A wrong price creates an unpleasant surprise at checkout, or a dispute if the shopper kept a screenshot.
• An out-of-stock recommendation wastes the shopper's time at the moment they were ready to buy.
• A wrong policy ("yes, returns are free") lands on your support team later, with an unhappy customer attached.

Language models produce text that sounds plausible. When they lack the facts, they fill the gap with something that fits the pattern. So the fix is not a better-behaved model. It is a system where the facts never depend on the model's memory in the first place.

Rule 1: products come only from your catalog

When a shopper asks for a recommendation, AskMerra first searches your catalog and hands the model a short list of matching products. The model can search again or look up a product's details through tools, but those tools only read your catalog.

Then comes the check that matters most. Every product the AI wants to show is identified by its id, and those ids are validated against the products it actually received. Anything that does not match is dropped before the shopper sees it. There is simply no path for an invented product to reach the screen.

A few more filters apply:

• Out-of-stock products are never recommended, which makes accurate stock flags in your feed one of the most valuable things you can provide.
• Products that disappear from your latest full sync are deactivated (never deleted), so they stop being recommended.
• You can exclude individual products from the AI in the catalog view, for example items you sell but do not want suggested.

Rule 2: prices and stock come from the database, not the model

This part surprises people. When AskMerra shows a product card, the price, sale price and stock status are read live from your catalog data at that moment. The model writes the sentences around the products. It never writes the price.

That holds even for answers prepared in advance. Precomputed answers refer to products only by their id, and the names, prices and cards are filled in live when the answer is shown.

How fresh is "live"? It depends on how you connect your catalog:

• File upload (CSV, JSON or XML): updated whenever you upload.
• Feed URL: synced automatically, every 1 to 6 hours depending on your plan.
• Push API: real time, for shops that want stock changes reflected immediately.

Price and stock changes apply as soon as they sync and do not trigger any reprocessing of the product, so updating prices often costs nothing extra.

Rule 3: understand every product before anyone asks

Accurate answers need more than a name and a price. A shopper asking for "a moisturizer that won't clog pores" needs the assistant to know which products fit, even if the name only says "Daily Cream".

So after import, AskMerra enriches each new or changed product. AI reads the description and attributes and extracts what shoppers ask about: use cases, skin or hair types, concerns, key ingredients or specs, routine steps and safety notes. Enrichment works from your product data and is told not to add facts or medical claims of its own. The result is indexed for semantic search.

You stay in control of this layer:

• You can inspect each product's enrichment in the dashboard.
• You can edit the tags by hand, and your edits are kept when the product is enriched again later.
• When a product's name, description or attributes change, enrichment runs again for that product only.

The most useful thing you can do here is also the least exciting: write detailed descriptions and fill in attributes such as ingredients, materials, compatibility and sizes.

Rule 4: policies come from your knowledge base

Shipping, returns and payments are where generic AI tools are most likely to guess, because every shop is different. AskMerra uses the documents you write in its knowledge base (shipping, returns, payments, brand, FAQ, and custom pages like size guides) in every conversation, and it does not invent policies.

Two habits make policy answers sharper:

• For shipping, list countries, costs, free-shipping thresholds, delivery times and cut-off times.
• For returns, state the window, the conditions, who pays return shipping and when refunds arrive.

Rule 5: not every answer needs AI

One good way to avoid AI mistakes is to skip the AI when it is not needed. Every message in AskMerra goes through a router that tries the simplest correct answer first:

1. Rules. Greetings, thanks, goodbyes and "are you a bot?" are answered instantly in six languages, free of charge.
2. Your FAQ. If the question closely matches one of your FAQ entries, your answer is returned word for word. No AI is involved.
3. Precomputed and cached answers. Common generic questions ("what do you recommend for dry skin?") are answered from earlier good answers. A cached answer is skipped if any product it mentions has gone out of stock, and it is always shown with live names, prices and cards.
4. The AI model. Everything else, with the relevant products, your policies and the conversation in front of it.

This keeps answers to repeated questions consistent, and cheaper: greetings are free, and FAQ and cached answers cost almost nothing.

Rule 6: know when to stop

Accuracy also means knowing what not to answer alone.

• Health topics get careful, non-diagnostic answers with a recommendation to consult a professional. A question like "Is the retinol serum safe during pregnancy?" is not answered as if the assistant were a doctor.
• Off-topic requests are politely declined.
• Hidden instructions are ignored. Catalog and knowledge content is treated as data, so text like "ignore your rules" inside a product description has no effect.
• When it cannot help, the assistant says so and offers a contact form. Your team gets the message and the full transcript by email, and the conversation is marked as escalated in the dashboard.

How to check it yourself

You do not have to take any of this on trust. Before going live, use the playground to ask the questions your customers ask. The debug panel shows which products were retrieved, which route produced the answer (rule, FAQ, cache or AI) and what it cost.

After launch, the conversations inbox shows every transcript with answer-type badges, the products recommended and clicked, and shopper feedback. A thumbs-down also removes that answer from the cache.

A short checklist before you switch the widget on:

• Stock flags in your catalog are accurate.
• Descriptions and attributes are detailed enough to answer real questions.
• Shipping and returns documents are filled in, with real numbers.
• Your most common one-answer questions exist as FAQ entries.
• An escalation email is set in Settings, so handovers reach a person.

Try it on your own catalog

The quickest way to judge accuracy is to test with your own products. Create your account, import your catalog and ask it the hard questions, or book a demo if you prefer a guided tour. For the technical picture, read how the assistant answers. New to the topic? Begin with what an AI shopping assistant is.