Skip to content
3 min read

Why Claude says it can't see your Shopify store

A real story about a connector, and what "connecting AI to your systems" actually means, including why prompting harder never fixes it.

Builds on What is an AI agent, actually?.

A while ago, someone close to me, a marketer who is very good at her job, tried to get Claude to help with her company's Shopify store. She asked it to look at the products. Claude said it couldn't. She tried telling it to look at her browser. It couldn't do that either, not without the right setup. After twenty minutes she gave up and went back to doing it by hand.

When she told me about it later, the fix took about two minutes: she didn't need a smarter AI; she needed to plug the one she had into Shopify. There's a thing called a connector that does exactly that. She'd never heard of it. Almost nobody outside the tech bubble has.

The AI can think about your store. It just can't see it.

By default, an AI assistant is a brilliant employee at your competitor. They know Shopify inside out: how it works, what a product listing should say, what good store hygiene looks like. What they have no view of is your Shopify: your products, your orders, your customers, the reason half of it is set up the way it is.

No connector

  • Knows Shopify in general
  • Cannot see your products, orders or customers
  • Answers from general knowledge, or refuses

A brilliant employee at your competitor.

Connector in place

  • Knows Shopify in general
  • Fetches your actual product list on request
  • Comments on your descriptions, not typical ones

The same employee, working at your company.

The same question, asked of the same model, on either side of a connector

So when you ask "look at my store and tell me which product descriptions are weak", the model genuinely cannot do it, not because the task is hard, but because there is no connection between the AI and your store. Asking again, more firmly, doesn't create one.

A connector is that connection

A connector does two things: it proves to the other system that the AI is allowed in (you log in and grant access, the same way you'd approve any app), and it gives the AI a specific set of actions it may perform there.

Once my marketer connected Shopify to Claude, "look at my products" stopped being an impossible request and became a task: Claude fetched the product list, read the descriptions, and made concrete suggestions about her actual products. Same model, same intelligence. The only difference was the connection.

A connector only allows what it allows

This is the part that trips people up even after they've connected things. A connector is not a magic door into the whole system. It's a menu of specific actions someone chose to expose.

On the menu

  • Read the product list
  • Update a description
  • Check recent orders
  • Create a discount code

The AI can do these.

Not on the menu

  • Delete a customer
  • Change your payout account
  • Anything nobody wired up

No phrasing makes these possible.

A connector is a menu of actions, and the limit is the safety feature

That limit is also the safety feature. You want the menu to be short. A connector that only reads your orders can never wreck your orders.

It also explains a whole category of AI frustration. The AI does A and B, then refuses C, and it feels random. It isn't. When an AI "can't do" something, the question to ask is "what is it connected to, and what does that connection allow?"

Four questions that solve most "it's not working" moments

  1. 01Is the tool connected at all? Most "it can’t see my stuff" problems end here.
  2. 02Did it actually use the connector, or answer from general knowledge? Say "use the Shopify connector" and compare.
  3. 03Did you grant access to the right account or workspace? Wrong-place looks identical to not-connected.
  4. 04Does the connector allow this action? Reading and writing are separate permissions.
  5. 05Still failing? Now it is a real question, and a much more precise one.
Run these in order before concluding the AI cannot do it

Why this matters more than it seems

Individually, a connector is a small convenience. Collectively, they change what an AI assistant is. An AI with your inbox, your documents, your project tracker, and your store connected isn't a chatbot any more; it's a colleague who can reach the work.

One connector gets you better answers. Several get you steps: read the orders, find what sold out this week, draft the restock email, put it in front of a human to send. Each step hands its output to the next, and the person still decides what goes out. That chain is a workflow, and a good few of your repeatable tasks already have its shape.

Most teams are one or two connectors away from AI being genuinely useful, and they don't know it. The gap isn't intelligence. The gap is the connection, and that's fixable in an afternoon.

If you suspect your team is stuck there, that's exactly the kind of thing I help with.

Read next

Put it to work

Practise the ideas on a marketing task.

The £49 AI for Marketing book explains the ideas through familiar problems and complete examples. Read the free sample to see the approach.

Explore the book and free sample →

For help with your team’s systems, discuss a workflow →

Email notes

Practical AI, in your inbox.

Occasional notes from Abai on marketing workflows, useful examples, and ways to work together.