An AI agent for product selection: budgets, useful questions and relevant recommendations

“I need a backpack for work, with room for a laptop occasionally. Under €80.” Your AI agent can turn that purpose, requirement and budget into a useful selection from your store.

Breezaro can search a connected catalog, apply price limits and display selected products as cards in website chat. It can also use the content associated with the page a visitor is viewing. Good recommendations begin with reliable information and useful follow-up questions.

The backpacks, prices and conversations below are illustrative examples for planning your own setup, not results from a real customer deployment.

Check the catalog first

Open Sources → Products and check your connected catalog. Product search needs the product feature to be available, a qualifying plan and imported products. Telling the agent to recommend your products does not connect the catalog for you.

Pick three items you know well and compare their records with your store. Check the name, price, currency, availability and link. Does the card open the right product? Do your sources contain the details needed to judge suitability? For a backpack, the laptop compartment dimensions may matter more than the word “laptop” in its name.

Prices and availability come from the catalog stored in Breezaro. They are as current as its synchronization with the source. Each question does not trigger a new stock check against your store. After changing prices or your range, check the source update and a few individual items. For Shopify, see the catalog connection guide.

Use the budget when searching

Customers can give a maximum price, a minimum price or a range. Product search supports all three:

These numerical limits use the catalog's currency. The examples assume a catalog in euros. If your customer gives a budget in dollars and your catalog uses euros, do not rely on automatic currency conversion. Test whether the agent clarifies the difference. You can add an instruction to confirm an unclear currency before recommending products.

A product without a stored price cannot qualify as meeting a price limit. If the search finds no suitable option, a useful response may explain that and ask whether the customer wants to change a requirement. A more expensive alternative should be clearly identified as exceeding the original budget.

Ask one question that changes the selection

Ask for details that would otherwise make the recommendation a guess. For the backpack request, a useful follow-up is:

“What size laptop will you carry?”

That answer may rule out unsuitable products. A preferred color can wait unless the customer has made it essential. For a light fitting, the important question could concern installation; for a replacement part, it might be the exact device model.

You could add this short rule to AI instructions:

“Respect the customer's stated budget when recommending products. If a detail needed to judge suitability is missing, ask one specific follow-up question. Explain recommendations using properties documented in our sources. Do not guess missing specifications or compatibility.”

This is a suggested instruction to test, not a guarantee of every response. Your sources still need to contain the relevant specifications. The knowledge-base guide explains how to prepare useful content.

Understand “Will my laptop fit in it?”

In the website widget, Breezaro can give the agent context from the page currently open when its URL matches a processed source or catalog product. Customers do not have to repeat the complete product name in every question.

Someone viewing a backpack might simply ask, “What are its dimensions?” The agent can use the information associated with that product. If the dimensions are missing, knowing the URL does not supply them. If the page has no matching source, you cannot assume the agent automatically sees its contents. This feature also does not carry a view of your website into a conversation in another app.

Test two different product pages in sequence. Ask the same short question on each and check which product the answer describes. Include a follow-up about an earlier part of the conversation: page context should help interpretation without arbitrarily replacing the subject already under discussion.

Explain the recommendation alongside the cards

Your AI agent can display selected products as cards in website chat, using the names, images, prices, availability and links your catalog provides. A card gives the customer a route to the product page. It does not, on its own, explain the choice.

An illustrative explanation might be: “The first backpack fits your request: it is within budget, and its listed compartment dimensions accommodate your laptop.” Every part of that statement needs support from your sources. For a second option, the agent can identify a missing specification and explain what still needs checking.

Avoid instructions that demand several cards for every request. A requirement outside your range, or a missing compatibility detail, may call for a short explanation and a follow-up question instead.

Offer help while the visitor browses

Under Scenarios, you can configure an opener based on the product-page address and time spent there. For example, after 20 seconds: “Choosing a backpack for everyday use? Tell me what you need to carry.” Twenty seconds is a starting point for your test, not a proven best time for every store.

With custom text, you control the exact wording. If you let AI write the message, the relevant page content needs to be available in your sources. Scenarios referring to previously viewed pages also require Visitor context to be enabled. Select the intended mode, enable Reach out to visitors and set frequency limits. The scenarios documentation covers the controls.

Run six checks before launch

1. Set a budget just below a known product's price. The agent should not describe it as within budget.
2. Give a price range and check both boundaries and the currency.
3. Ask about a property your sources do not document. The answer should acknowledge the gap.
4. Ask “What are its dimensions?” on two product pages and check the subject of each answer.
5. Open every recommended card and compare its information with your store.
6. Request something outside your range. The agent should not invent a match.

Start with one category you know well. Repeated questions reveal which details customers need when choosing. Add those details to your sources, then repeat the checks to verify the improvement.