Can Shopify chatbots answer product questions on product pages? We tested 2,652 stores.

Shopify chatbots mostly cannot answer a question about the store's own best-selling product. Pier39 AI tested 2,652 Shopify stores running Gorgias in October 2026 by asking one product question in each store's chat. 18% gave a product-specific answer. The rest showed no chat, routed to offline staff, or replied with articles and forms.

How did we test 2,652 Shopify stores?

We opened the chat on 2,652 Shopify stores and asked one question about each store's best-selling product, the kind a shopper asks before deciding whether to buy. All 2,652 stores run Gorgias for support chat, so every store was tested on the same platform under the same conditions.

Two independent AI agents scored each reply on whether it answered the product question, and we recorded how long the first reply took. A store passed only if the reply was specific to the product asked about. Articles, forms, ticket confirmations and generic reassurance counted as failures.

Most failures in the results below are how the store set up and staffed its chat, not a fault in the platform. A chat that is hidden, offline or gated behind an email form cannot answer anything, whatever it runs on.

What happened when we asked?

82% of the 2,652 Shopify stores did not answer the product question. Only 18% gave a real, product-specific reply.

The stores that did answer were slow. The median time to a reply was 16 seconds, and no store answered in under 5 seconds. A shopper who has just typed a question is deciding between buying and leaving during that gap.

The support bots were doing the job they were set up for: collecting emails, opening tickets and routing to people. A shopper who walks into a store and asks a salesperson about a product does not expect to be told to file a ticket. Online, that is the standard experience.

What do unanswered product questions cost?

Unanswered product questions could be costing the 2,173 failing stores around $700 million a year in sales, based on a model with conservative assumptions. This is an estimate, not measured lost revenue.

The model works as follows:

1. The 2,173 stores that failed the test receive an estimated 4.8 billion visits a year (StoreLeads traffic estimates).
2. Assume 2% of abandoned sessions involved a product question the shopper could not get answered.
3. Assume 5% of those shoppers would have bought if someone had answered.
4. At the stores' average order values, that is roughly $700 million in annual sales.

Change the assumptions and the number moves, so treat the $700 million as the scale of the problem rather than a precise figure. What does not change is where the money sits. These shoppers were already on the product page, considering a purchase. A store does not need to pay to acquire them again. It needs to answer the question they asked.

Why did 81% of the bots that replied sound the same?

747 of the 2,652 stores had a bot that replied instantly. We read all 747 replies. 611 of them, 81%, opened with one of a handful of stock lines.

218 brands, selling everything from skincare to power tools, started with the same eleven words. Only 136 of the 747 had an agent with a voice of its own, either introducing itself by name or going straight to the product.

Brands spend years on voice: packaging, emails, the way the founder writes. Then a shopper asks the one question that decides the sale and gets the same sentence as 218 other stores.

What should a store's chat do instead?

A store's chat should answer the product question in seconds, in the brand's own voice, before the shopper leaves the page. Support chat is built to open tickets and route to people, which is why it fails the shopper standing on a product page with one question.

Pier39 AI builds Nash, which sits on the store, catches a shopper who is hesitating, and answers product and policy questions from a knowledge graph of the whole store. Nash is trained on each brand's voice before it goes live, and the brand controls how it sounds and what it can and cannot say.