AI & Automation

AI Product Recommendations in Chat: How They Work and How to Use Them Well

Written by The StoreChart team7 min read

In short

StoreChart's AI agent recommends products inside a conversation, not in site-wide widgets. On your website chat and on WhatsApp, it answers from a product knowledge base built from your own site. When a shopper needs help choosing, it asks one clarifying question, then suggests the best fit and explains why. It does not read live stock levels.

When store owners hear "AI product recommendations," most picture a grid of "you might also like" items that follows a visitor around the site. That is one kind of recommendation engine, and it depends on tracking behaviour at scale. StoreChart takes a different approach: its AI chatbot on your website and its AI agent on WhatsApp recommend products inside the conversation, based on what the shopper just told them. This guide explains how that works in practice, what product information the agent needs from you, how to improve its suggestions and where its limits are, so you can decide where it fits in your store.

Conversational recommendations vs. recommendation widgets

A recommendation widget guesses: the shopper hasn't said what they need, so the system infers it from browsing and purchase patterns. A conversational recommendation responds: the shopper has just written "I need a desk for a small room" or "which one is better for a tall person," and the agent answers that specific request.

  • Starts from a stated need, not from inferred behaviour.
  • Explains the choice: the agent says why a product fits this shopper.
  • Happens where the shopper already is: in your website chat or in WhatsApp.
  • Needs no widget on the storefront: the conversation is the interface.

The trade-off is reach. A widget shows items to every visitor, while a chat recommendation reaches only shoppers who start a conversation. The two are complements, not rivals.

How the agent chooses a product

The agent works from a product knowledge base built from your website. This is how it is built and kept current:

  • First scan: when you set it up, StoreChart scans your site and turns product pages into structured entries.
  • Product feed: for the WhatsApp agent, the scan also reads your public Shopify or WooCommerce product feed when available, so prices and variants arrive as data rather than as text read off a page.
  • Re-scans: the WhatsApp agent's knowledge is re-scanned on a schedule, so changed prices and discontinued models are picked up.
  • Editing: you can review and edit every entry.

Each product entry can hold:

  • price and currency, or a link to the page that shows the price;
  • options such as sizes, colours or variants;
  • what is included and what is not;
  • specifications such as dimensions, materials and capacity;
  • warranty and availability as stated on your site;
  • who the product suits, why someone would choose it over alternatives, and common concerns.

In conversation, the agent follows a simple pattern:

  1. A clarifying question: if the shopper names a category but hasn't chosen, it asks one short question about use, size, budget or preference, rather than listing every product.
  2. A recommendation: once it knows enough, it recommends the best one or two fits and explains why they suit this shopper.
  3. A link: it includes the product link from the knowledge base.

On WhatsApp, if several products match, the agent names all of them rather than presenting a partial list as the full range.

Example: a standing desk conversation

Here is a WhatsApp exchange for a store that sells standing desks:

  • Shopper: Hi, I'm looking for a standing desk.
  • Agent: Happy to help. What will you mainly use it for, and how much space do you have?
  • Shopper: Home office, the room is small, and I'm 1.90 m tall.
  • Agent: Then the compact model is a good fit: its top comes in 120×60, and its height range goes high enough for someone your height to stand comfortably. Here is the product page.

Notice what the agent did not do, and what it relied on:

  • No long list: it didn't list every desk in the catalog after the first message.
  • No guessing: it didn't guess a height range that wasn't in the product entry.
  • Written facts: the recommendation depended on two facts the merchant had written down, the available sizes and the height range.

If the height range had been missing, the agent would have said it didn't have that detail instead of inventing one. That is the difference between an agent that helps sell and one that creates returns and disappointed customers.

The figures in this article's examples are illustrative only and do not reflect actual customer data.

What the agent doesn't do

Being clear about limits prevents disappointment:

  • No live inventory. The agent is not connected to your stock levels. "Is the black one in stock right now?" gets a pointer to your WhatsApp or phone in website chat, and is passed to your team on WhatsApp, instead of an answer from memory.
  • No order lookups. "Where is my order?" goes the same way. On WhatsApp, someone from your team can reply in the same chat.
  • No browsing-based personalization. It doesn't track visitors or build profiles.
  • No invented details. It answers from the knowledge base. If a spec isn't there, it won't make one up.

On WhatsApp you also choose how independent the agent is:

  • Supervised mode: it replies on its own and stops in a conversation as soon as someone from your team replies there.
  • Auto mode: it keeps replying even after a person joins.

Improving the agent's recommendations

The quality of recommendations follows the quality of your product entries. Four habits help most:

  1. Fill in "suits" and "why choose." These fields are what let the agent match a product to a person. "Best for rooms under 10 square metres" is more useful than "compact."
  2. Write down common concerns. If shoppers keep asking whether a desk wobbles at full height, put the answer in the entry.
  3. Keep options and prices complete. A missing size or price forces the agent to refer the shopper elsewhere.
  4. Read real conversations. Where shoppers correct the agent ("no, I meant the other one"), the product data is usually ambiguous. Fix the entry, not the prompt.

The knowledge base page covers how entries are reviewed and edited. A practical launch order:

  1. Scan your site.
  2. Review the product entries for your ten best sellers first, and fill in the "suits" and "concerns" fields for those.
  3. Ask the agent the questions your customers really send you.
  4. On WhatsApp, start in supervised mode so your team can step into any conversation, and switch to auto once you trust the answers.
  5. After the first week, review the conversations the agent couldn't answer and check which ones came from information missing in your entries.

Measuring AI product recommendations

StoreChart doesn't tag an order as "recommended by the agent," so measure with simple, honest signals:

  • Tagged links. Use product links with UTM parameters in the knowledge base, so your web analytics can separate visits that came from chat.
  • Conversation outcomes. Review a sample of conversations each week: did the shopper get a clear recommendation, and did they click through?
  • Correction rate. Count how often shoppers reject or correct a suggestion. A falling rate means your product data is improving.
  • Order value. If the goal is to raise average order value, watch average order value in your reports before and after you launch the agent, keeping other changes to a minimum.

Site chat or WhatsApp agent

It depends on where your shoppers come from:

  • Site chat meets people while they browse product pages and hesitate.
  • The WhatsApp agent meets people who send a message, for example after tapping an ad, a social media link or a WhatsApp button on your site.

Each is set up separately, so you can start with the channel that already receives the most questions and add the other later. If most of your questions already arrive on WhatsApp, that is the natural place to start.

Conversational commerce in your sales strategy

Chat recommendations help a hesitant shopper find the right item. They don't replace clear product pages, fair pricing or a simple checkout. Use them for the questions a static page can't anticipate:

  • Unusual sizes that your size chart doesn't cover.
  • Mixed needs, such as a product that has to suit two uses.
  • "Which one for me", when several models look alike.

Let your storefront handle predictable pairings you already know sell together. For more on the product side, see the product recommendations feature page.

Frequently asked questions

Does the agent know what is in stock right now?

No. The agent answers from the product information in its knowledge base, including any availability your site states. For live stock on a specific item, the website chatbot points the shopper to your WhatsApp or phone, and on WhatsApp someone from your team can reply in the same chat, instead of the agent guessing.

Where does the agent get its product information?

From your website. StoreChart scans your pages, and for WhatsApp it also reads the public Shopify or WooCommerce product feed when your site has one. The result is a list of product entries you can review and edit.

Does it personalize recommendations from browsing history?

No. There is no behavioural profile or "you might also like" widget. The agent recommends from what the shopper says in the conversation: what it is for, size, budget and preferences.

Can the agent invent a product or a link?

It is instructed to answer only from its knowledge base, and product links come from the knowledge base rather than being made up. If the information isn't there, it says so and points the shopper to your WhatsApp or phone.

See StoreChart with your own store data

Open an account, connect your store and see your own numbers.