Conversational Product Recommendations: How They Work and the Best Chatbots for Them

Three different things get called "product recommendations" on an online store, and they behave very differently.

The conversational kind is what a good shop assistant does: listen, bring three options, adjust when the shopper says "too expensive" or "in blue". Done with a language model alone it fails, because the model will happily invent a product. Done with retrieval-augmented generation, where the bot searches the store's own synced catalogue before writing anything, it works. What is a RAG chatbot explains that mechanism.

1. Read the intent. "Something warm for a toddler, not wool" contains a category (clothing), an audience (toddler), a requirement (warm) and an exclusion (wool). The model extracts these rather than matching a single word.
2. Retrieve matching catalogue items. The bot runs a search over the store's synced products, combining vector similarity (meaning) with exact-text matching (names, SKUs, sizes). Vatdi uses both over a per-store index that the platform plugin keeps in sync.
3. Filter by price, attributes and stock. "Under $100" removes items above that price; "in blue" narrows to the variant; where stock is enabled, the card shows it.
4. Present cards. Each suggestion is shown as a card with image, name, price, sale price while a sale is on, a link to the product page, and stock if the merchant turned it on. Two to four cards beat a list of ten.
5. Refine. The shopper says "cheaper" or "do you have it in a larger size", and the bot repeats steps 2 to 4 with the new constraint, keeping the earlier ones.

Two things make the difference between a demo and a tool that sells. The retrieval step must run on every product question, so the bot never falls back to guessing. And the bot must be willing to say "we do not have that" and offer the nearest match or a person, which is what a trustworthy assistant does. How to train a chatbot on your data covers how the catalogue and policies get into the index.

Plan caps decide how much catalogue the bot can see: on Vatdi, 200 synced products on Free, 2,000 on Starter and 10,000 on Grow. A bot can only recommend what is in its index, so a partial sync produces partial recommendations.

Realistic exchanges from a grounded product recommendation chatbot on a store that sells outdoor clothing, kitchenware and pet supplies. Product names are illustrative.

• Shopper: "I need a rain jacket for cycling to work, under 100." Bot: "Here are three waterproof jackets under $100 with reflective details for commuting." Three cards follow, each with image, price (one showing a sale price) and a link. Shopper: "Do any of those pack small?" Bot: "The Ridge Shell packs into its own pocket; the other two do not. It is $79." One card follows.
• Shopper: "Which of your kettles sells best?" Bot: "By units sold this month, the top three are the 1.7 L Glass Kettle, the Gooseneck Pour-Over Kettle and the Travel Kettle." Three cards follow, ranked from real sales data. Shopper: "Is the gooseneck good for tea or just coffee?" Bot: "Both; its description says the temperature presets include 80 °C for green tea and 100 °C for black tea."
• Shopper: "Grain-free food for a senior cat with kidney issues." Bot: "We stock two grain-free senior formulas. We do not carry a veterinary renal diet, so for kidney issues please check with your vet first." Two cards follow, then: "Would you like me to connect you with our team for advice on which one suits your cat?"

Note what the bot does in the third exchange: it recommends only what exists, names the gap, and offers a person. That is the behaviour to test for in any tool.

Measure the recommendation feature on three levels. Some numbers come from the chatbot, some from your store analytics.

Read ten transcripts a week for the first month. The failures that matter (a wrong price, a sold-out card, a made-up product) are visible in transcripts long before they show up in a conversion number.

• Hallucinated products. The cause is always the same: no retrieval step, so the model answers from its training rather than your catalogue. The fix is a tool that searches the synced catalogue on every product question and refuses when nothing matches.
• Stale prices. A card showing last week's sale price is a refund request waiting to happen. Use a plugin-based sync rather than a page crawl, and check that sale prices disappear when the sale ends.
• Recommending out-of-stock items. The bot can only be as accurate as the inventory field it reads. Many stores do not maintain stock precisely, which is why Vatdi keeps stock optional: with it off, cards simply omit stock instead of showing a wrong number; with it on, the sync supplies it.
• Too many cards. Ten options recreate the problem the shopper came to chat to avoid. Two to four, then refine.
• Ignoring the follow-up. "Cheaper" must keep the earlier constraints. Test this specifically.
• Partial catalogue. If the sync covers 200 of 1,500 products, shoppers will be told you do not sell things you sell. Match the plan cap to the catalogue.

One boundary to be clear about: conversational recommendations answer a shopper who asks. They are not an upsell or cross-sell engine that inserts add-ons into the cart, and Vatdi does not include one. If automated cart add-ons are the goal, that is a different category of tool.

Ranked by how well each tool grounds recommendations in real catalogue data, then by price. Prices are published list prices as of September 2026; prices change, so check each vendor's pricing page.

Vatdi is first because its cards come from a plugin sync with sale prices and real sales counts across seven platforms, on a free plan. Its limitations are real: on Shopify it trains on pages, collections and policies without a catalogue API sync or order lookup; it has no upsell engine; and it covers the website widget and WhatsApp only. Gorgias is the better pick for a Shopify store with a support desk, and Shopify Inbox is enough when a person is always available. The wider e-commerce chatbot ranking covers support and order handling as well as recommendations.

The quality of conversational recommendations follows the quality of the catalogue feed, so the platform matters.

• Plugin sync (best): WooCommerce, OpenCart 3 and 4, PrestaShop 8 and 9, Magento 2, Shopware 6, Joomla with VirtueMart or HikaShop, Drupal Commerce. Products, prices, sale prices, images, links, optional stock and sales counts flow from the store. Order status lookup is available on these too. See integrations.
• Shopify: the Vatdi app adds the widget and trains on pages, collections and policies plus the dashboard knowledge base. There is no catalogue API sync or order lookup; recommendations are as good as the collection and product pages.
• Script embed (BigCommerce, Wix, Webflow, Squarespace, custom, React or Next.js): the AI trains on crawled pages and uploads. Recommendations work from product page content and a CSV upload of the catalogue; keep the CSV current when prices change. Details on other websites.

If you are on a script-embed platform, upload the catalogue as CSV with name, description, price, URL and image URL columns, and re-upload after price changes. It is manual, but it gives the bot something to retrieve rather than leaving it to guess. Plans and caps are on pricing.