Integrating OpenAI’s ChatGPT with Shopify: the complete guide for eCommerce

It’s not a chatbot. It’s an intelligence layer on top of the store.

When a brand asks us to “put ChatGPT on the site”, the first thing we do is take the word chatbot off the table.
A widget that answers questions is the least interesting part.
The interesting part is connecting a language model to the store’s real data: catalogue, stock, orders, customers.

Integrating ChatGPT with Shopify means exactly this: giving the model governed access to your eCommerce data and deciding what it is allowed to do with it.

‍

The four use cases that make sense today

• Product content. Descriptions, attributes, translations and feed copy generated from the structured catalogue data, not from thin air. On catalogues with thousands of SKUs it is the difference between launching in days and launching in weeks.
• Pre- and post-sales support. An assistant that knows sizes, availability, delivery times and the brand’s return policy, and hands over to a person when needed.
• Conversational search. “A tailored blazer under €400, delivered by Friday”: the model reads the intent, the catalogue answers. This is the piece we are working on with our Aura prototype.
• Internal operations. Bulk tagging, sales analysis, campaign drafts: the eCommerce team asks in natural language, the model executes through the Admin API.

‍

How the integration actually works

The architecture we use is always the same; only the weights change.

• Data. Shopify’s Storefront API and Admin GraphQL API as the single source of truth for products, variants, stock and orders. No ageing copies of the catalogue.
• Model. OpenAI’s API with function calling: the model doesn’t “make up” a price, it calls a function that reads it from the store. Every answer about stock, price or delivery is anchored to real-time data.
• Knowledge. Policies, size guides, FAQs and the brand’s tone of voice indexed (embeddings) and passed to the model only when relevant. This is what makes the assistant “the brand’s” rather than generic.
• Guardrails. What it can say, what it cannot, when it hands over to a human. A log of every conversation. No personal data sent to the model unless strictly necessary.

‍

Where it goes wrong

We see the same three mistakes over and over.

• Letting the model answer from memory. Without function calling the model says a product is available when it isn’t. In fashion, with stock changing every hour, that is unacceptable.
• Starting from the widget. Clean data first (attributes, sizes, images, policies), interface second. A messy catalogue produces a messy assistant.
• Ignoring privacy and cost. Every call has a price and every piece of data you send is a processing activity. You need a DPA with the provider, data minimisation and a spending cap per conversation.

‍

The checklist before you start

• A catalogue with complete, consistent attributes (material, fit, sizes, care).
• Return, shipping and sizing policies written so that a model can quote them.
• One use case only, with a metric: response time, resolution rate, assisted conversion.
• Tests with real customers before launch, with a human reading the conversations.
• A monthly budget and a plan B if the provider changes pricing or APIs.

‍

This is the direction

Artificial intelligence doesn’t replace the store. It reads it, interprets it and makes it easier to use, for the customer and for the team.
Start from the data today and tomorrow you have an assistant that sells.
Start from the widget and you have a widget.