An AI assistant on a Shopify store: what it can and cannot see

The useful version of this question is not "can AI help my store". It is "which of my questions are lookups". Lookups are where an assistant is unambiguously better than a person, because it never gets bored and the answer is already in a database.

What it can see, and what it must not

Start with the boundary, because it is the thing that should worry you and the thing most vendors gloss. An assistant reading your Shopify catalogue through a storefront token sees what your storefront sees: products published to that sales channel. Shopify's own docs are careful about this — the Storefront API returns products “published to your app”, and a product's onlineStoreUrl is null when it is not on the online store channel.

The phrasing matters more than it looks. It is not that the API hides unpublished products — there is no filter doing that work. It is that publication to a channel is what makes a product reachable by a token belonging to that channel. Drafts were never in scope.

A product description is a text field that a supplier feed can write to. Treat everything in your catalogue as untrusted input, because somebody else can edit it.

The dull questions are the valuable ones

Gorgias (2023) reports that "where is my order" is the single most common ecommerce question, averaging 18% of incoming requests. It is their own platform data and it is from 2023, so treat the precision loosely — but anyone who has run a store will recognise the shape.

This is the least exciting slide in any AI pitch and the most defensible one. There is no judgement in it, no tone to get wrong, and no risk of inventing a refund policy. It is a date, read out loud, at 2am.

Does it sell anything?

Here the evidence is thinner than the marketing, and it is worth being precise about which claim is which.

• The strong independent evidence is modest. A randomised field experiment published in Information Systems Research (2025) found an AI shopping assistant raised sales 3.00% and cut product returns 12.55%. Three percent is a real, causally identified number — and it is a long way from the lifts on most vendor landing pages. The returns figure may be the more interesting half.
• Shopify's own figures measure something else. Its Q1 2026 data found AI-referred sessions convert nearly 50% better than organic search (2026), with 14% higher order value. That is about shoppers arriving from ChatGPT and similar — external AI sending traffic in, not an assistant on your site. Do not let the two blur.
• Abandonment is mostly not a support problem. The average documented cart-abandonment rate is 70.22% (2025), but Baymard's own survey puts 42% of that down to people who were just browsing. An assistant does not fix "not ready to buy". It can fix "could not find whether this ships to Norway".

Setting it up without regretting it

1. Connect the catalogue, then test with a draft product. Ask the assistant about it by name. If it answers, stop and fix that before anything else.
2. Add the pages that are not products. Shipping zones, returns windows, sizing, care instructions. These answer more questions than the catalogue does, and the dull documents beat the marketing ones.
3. Decide what happens when it is stuck before you put it live, not after. Even a contact address is enough to rescue the conversation — the handoff post covers the three endings.
4. Read the unanswered list weekly. It is a content brief written by your customers.

And measure the thing you actually care about. Running an assistant alongside a storefront experiment is the cleanest way to know whether it moved revenue rather than sentiment — Zinx Signal exists for that half of the problem, and its note on how many visitors a Shopify test needs is worth reading before you draw any conclusion from a fortnight of data. For what the replies themselves cost, the cost-per-conversation calculator takes your volumes directly.