Shopify has put AI in front of merchants at every level: in the admin, in the theme editor, and through a few thousand apps. Some of it quietly removes work. Some of it produces output you will spend longer fixing than writing yourself.
Here is where it has actually paid off on stores we work on, and in what order we would try it.
The short version
Four areas are worth your attention: product descriptions, on-site search, customer support, and recommendations. Everything else is optional. Start with search, because it is the one where a bad experience costs you a sale immediately and you can measure the change.
Product descriptions
Generated descriptions are a good first draft and a poor final one. They are fluent, they are structured, and they say nothing a competitor could not also say, because the model only knows what is in the title and the specifications you gave it.
What works: give it the specifications, the materials, the sizing quirks and the two objections your support team hears most, then edit the result. What does not work: generating three hundred descriptions overnight and publishing them. Duplicate-sounding copy across a catalogue is not a ranking strategy, and shoppers reading two product pages in a row can tell.
Search and discovery
This is where we have seen the clearest return. Shopify’s default search matches words; shoppers use descriptions. Someone searching for “waterproof jacket for dog walking” is describing a use, not a product title, and a keyword search returns nothing.
Semantic search apps handle that well. Before installing one, look at your own search logs for a month. If a large share of searches return no results, you have found your highest-value fix. If they mostly return results and people buy, spend the money elsewhere.
Customer support
Support chat is the feature merchants try first and regret most often, usually because it was pointed at the whole internet rather than at a defined set of answers.
The version that works is narrow: it answers from your shipping, returns and sizing policies, it says “I will pass this to the team” for anything else, and it hands over cleanly. The version that causes problems is the one confidently inventing a returns window you do not offer, which is a promise a customer will reasonably expect you to honour.
- Restrict it to your own policy pages and FAQs.
- Give it an obvious route to a human on every reply.
- Read the transcripts weekly for the first month.
- Never let it quote prices, stock or delivery dates it cannot check.
Recommendations
Shopify’s built-in related products are free and reasonable. A dedicated recommendation app can do better on a large catalogue with enough traffic to learn from, and usually does nothing on a small one, because there is not enough behaviour to model.
If your store has a few hundred orders a month, spend the budget on photography and product pages instead. The algorithm is not the constraint.
What to measure
Every one of these should be judged against a number you already track, not against how modern it feels:
- Search: share of searches returning no results, and conversion rate of sessions that use search.
- Descriptions: add-to-cart rate on the products you rewrote, against the ones you did not.
- Support: tickets resolved without a human, and refund or complaint rate afterwards.
- Recommendations: revenue per session, not clicks on the widget.
Give each change a few weeks and compare it against the same period before. If the number does not move, remove the app. Unused apps still load scripts on every page.
The cost nobody mentions
Most of these apps inject JavaScript into your storefront. Three or four of them together will show up in your Core Web Vitals, and on mobile that is the same audience whose conversion rate you were trying to improve. Check your speed before and after installing, on a mid-range phone rather than a laptop.
A sensible order
- Fix search if your logs show failed searches.
- Rewrite the descriptions of your top twenty products, with AI drafting and a person editing.
- Add support chat only once your policies are written clearly enough for a model to quote them.
- Look at recommendations last, and only if you have the traffic to justify them.
If you would like a second opinion on which of these is worth doing on your store, send us the URL and what you are trying to improve. We will tell you honestly if the answer is none of them.
