The quick way to count AI orders is to search order records for the names of AI products. We ran that search and then checked what it caught. Of 32,828 orders that matched, 7,406 were not AI orders. Counting them would have overstated the channel by 29%.
The search looked for assistant names anywhere in an order’s landing link, referring site, browser string and sales channel. That is deliberately generous, and it is roughly what a first attempt at an AI report does.
| Group | Orders |
|---|---|
| Matched an AI name | 32,828 |
| Were AI orders | 25,422 |
| Were not | 7,406 |
The orders that fell out did so for a small number of reasons.
| Why it matched | Why it is not an AI order |
|---|---|
| A different tag on the landing link | The shopper came from an ad or an email, and happened to be using an AI app as their browser. The tagged source is the source. |
| An AI app browser with an ordinary referrer | The shopper followed a link from another site while inside the app. The assistant did not send them. |
| An assistant name inside an unrelated value | A campaign name, a product handle or a search query that contains the word. |
| A Google click inside the Gemini app | An ad click carries its own click identifier and belongs to the ad. |
Orders placed inside an assistant have a recognisable shape: created by a server, no shopper browser, no landing page. It is tempting to treat every order with that shape as an AI checkout. That would be wrong by a wide margin.
Over 31 days, three of the largest sources of such orders across the stores we measure produced 5,679 orders between them. They turned out to be an app for sending product samples, an app for replacement shipments and an app for gifting products to creators. Real in-assistant checkouts over the same days: 191.
| Source | Orders in 31 days |
|---|---|
| Three non-AI server-side apps | 5,679 |
| In-assistant checkouts | 191 |
The safe rule is the dull one. An order source counts as an AI assistant only after its name has been read off real orders. Growth, timing and shape are not evidence.
One referring domain accounted for 143 orders in 30 days and is easy to file under AI, because the company behind it also offers an assistant. The domain itself is a search engine. The assistant lives on a separate address, which sent a single order in the same period.
The general problem is that the line between search and assistant is moving, and product names do not follow it. Google’s AI answers arrive as ordinary Google traffic and cannot be separated at all, which pushes the error the other way. Any list of AI referrers needs to be checked against what each address actually serves, and rechecked.
A tag on the link wins. If the landing link names a source, that is the source, whether or not the browser belongs to an AI app.
A referrer is matched on the host, exactly. Not on any text in the address, and not on a name that appears somewhere in the path or query.
A browser string counts only without a competing source. An AI app’s browser identifies the assistant when there is no tag and no other referrer.
A sales channel counts only once confirmed. New server-side sources are reviewed by name before any of their orders are counted.
These rules under-count in places. That is the trade we chose: a smaller number that holds up is more useful than a larger one that includes sample requests.
Every figure comes from the order records of Shopify stores running Aimerce, read as one fleet. An order counts as AI-attributed when Shopify’s own record of it names an assistant in one of four ways: the sales channel that created it, a tag on the landing link, the referring site, or a browser that identifies itself as an assistant’s app or agent. Each order is counted once.
What this does not measure. Shoppers who asked an assistant and later came back another way, clicks that arrive with neither a tag nor a referrer, visits from Google’s AI Mode and AI Overviews (they arrive as ordinary Google traffic), and agents that run in the shopper’s own browser without identifying themselves. All four make the numbers here a floor.
Privacy. Only fleet-level aggregates are published. Store counts are rounded to the nearest ten, and no store, brand, product or order is identified anywhere on this page.