How to add an AI chatbot to your Shopify store

What you actually need before you start

Adding a chat widget is trivial. Adding one that answers correctly is a content problem, not a technical one. Before installing anything, gather the material the assistant will answer from — because an assistant with nothing to read will improvise, and improvisation is how bots end up promising next-day delivery you don't offer.

• Your shipping policy, with real timelines per region and the cost thresholds that change them.
• Your return and exchange window, and any categories excluded from it.
• Size guidance if you sell apparel — charts, fit notes, and the 'runs small' knowledge that lives in your team's heads.
• The ten questions your support inbox answers most often. Pull them from real tickets, not memory.
• Anything the assistant must never say: discount authority, medical or legal claims, stock promises.

Step 1 — Install from the Shopify App Store

Install from the Shopify App Store and approve the access scopes during OAuth. The scopes decide what the assistant can answer: products and collections let it recommend and answer catalogue questions, orders let it handle tracking, and customers let it look up a specific shopper's history. Approving fewer scopes is a legitimate choice — it just narrows what the assistant can do.

Installation through the App Store means Shopify handles authentication and billing. You never hand a password to a third party, and you can revoke access by uninstalling.

Step 2 — Let the catalogue sync, then spot-check it

The first sync pulls products, variants, prices, inventory, and collections. This is the step people skip verifying, and it is the one worth verifying. Ask the assistant three questions you already know the answer to:

• A price question about a variant that recently changed — does it quote the current price?
• A stock question about something you know is sold out — does it say so, or invent availability?
• A comparison between two similar products — does it use real attributes, or generic marketing language?

Step 3 — Load policies and the questions you actually get

Paste your shipping, returns, and size content in as plain text. Resist the urge to write it in marketing voice; write it the way you would answer a customer who asked directly. Grounded, specific answers come from grounded, specific source material.

Then add the ten common questions with the answers you want given. This is the highest-leverage hour of the whole setup — it converts your team's institutional knowledge into something that works at 2am.

Step 4 — Set the guardrails

Decide three things explicitly. What tone the assistant uses. What it is allowed to promise. And when it stops trying and hands over to a human. Escalation is a feature, not an admission of failure — an assistant that says 'let me get someone who can authorise that' preserves trust far better than one that guesses.

Step 5 — Test like a sceptical shopper, not like an owner

Owners test the questions they hope people ask. Shoppers ask the awkward ones. Try the objection you dread, a question about a product you discontinued, a request for a discount, and something in a language you sell in but don't speak. Then try being rude to it. The goal is finding the edges before your customers do.

Step 6 — Measure the only number that matters

Once live, ignore conversation counts. Watch whether assisted sessions convert better than unassisted ones, and whether attributed revenue exceeds what the app costs. If it doesn't after a fair trial, the setup content is usually the problem — not the model.