The first real data on AI-driven shopping has landed, and two of the findings sit uneasily next to each other. One says the shopper who arrives by way of an AI assistant is the most valuable visitor a store can get. The other says that same shopper will not stay with any single assistant for long. Both hold up. Read together, they redraw the question of who actually owns the sale.
Shopify reported that shoppers who reach a store through an AI platform, whether ChatGPT, Perplexity, or Google Gemini, convert at a rate about 50 percent higher than shoppers who arrive from organic search. They also spend about 14 percent more per order. That is a large gap, and it points at something merchants have wanted for a decade: traffic that shows up already knowing what it wants.
The AI assistant has become the most effective referral source on the web. It has not become a loyal one.
The conversion signal is real, and narrower than it looks
A 50 percent lift is the kind of number that ends up on a slide with no context attached, so it is worth saying what it does and does not prove.
It does show that AI-referred traffic is pre-qualified. By the time a shopper clicks through from an assistant, the comparison, the question about returns, and the does-this-fit step have mostly happened inside the chat. The merchant receives a buyer near the end of the journey rather than the start of it. Higher conversion is what you would expect from that.
It does not show that AI is driving a flood of new demand. The base is still small relative to search and direct traffic, and a shopper who would have bought anyway converts just as well whether the last click came from an assistant or a Google results page. The honest reading is that AI is concentrating intent, not manufacturing it. That is valuable, but it is a different claim from the one the headline number invites.
Consumers are using every assistant at once
The more interesting finding came from PYMNTS Intelligence, which looked at which tools people actually reach for when they shop. ChatGPT led, with 60 percent of AI users turning to it for shopping and purchasing. Gemini was not far behind at 57 percent.
Those two numbers add up to more than 100, which is the whole point. People are not choosing an assistant the way they once chose a bank. They are keeping several open and switching between them by task, by habit, or by whichever one is already in the browser tab. The assistant is behaving like a search engine, not like a wallet.
This is the part that should give pause to anyone who assumed the AI shopping race would resolve into a single winner. The demand is real and it converts, but it is spread across providers, and the shopper feels no cost in moving from one to the next.
This is a problem for everyone building a checkout rail
For the past year the contest in agentic commerce has been a contest over rails. Visa has its Trusted Agent Protocol, Mastercard has Agent Pay, OpenAI has its commerce protocol, and the open camp has x402. We tracked all of it in our State of the Stack report, and the running assumption was that whoever owned the moment of payment would own the relationship that came with it.
The multi-assistant data complicates that assumption. If the buyer's loyalty sits with whichever assistant they happened to open this morning, then owning the rail underneath the transaction does not give you the customer. It gives you a toll booth on a road the customer can leave at any junction.
Owning the payment rail is not the same as owning the shopper. The shopper now belongs to whichever assistant answered the question, and that changes by the day.
The rails still matter, because money has to move and someone has to carry the risk. But the durable relationship is forming one layer up, at the assistant, and none of the network credentials settle who controls that layer. Our live x402 tracker shows the plumbing maturing quickly. It says nothing about who the buyer trusts, because the plumbing was never designed to.
What merchants should actually do
The practical takeaway is unglamorous and immediate: stop optimizing for one assistant and start showing up well in all of them.
That means treating answer engines the way you once treated search. Clean product data, structured feeds, clear specifications, and content an assistant can read and cite without guessing. The shopper never sees your homepage in these journeys. They see whatever the assistant assembled about your product, and your job is to make that assembly accurate and complete across every assistant your customers use.
It also means dropping the idea that one integration is enough. A merchant who builds only for ChatGPT is invisible to the 57 percent who are asking Gemini, and the split will keep moving as new models arrive.
The layer no one owns yet
Underneath all of this sits a gap that none of the current players have filled. When a shopper moves between assistants, nothing carries with them. The preferences they stated in one chat, the budget they set in another, the brands they ruled out last week, all of it stays trapped in whichever assistant heard it.
There is no portable buyer relationship, no shared memory the shopper controls and the merchant can recognize, the kind of portable mandate the agentic economy will eventually need. Attribution is guesswork, and continuity does not exist. That is the layer worth watching, because the company that lets a shopper carry their intent from one assistant to the next, on their own terms, will own something the rails and the assistants cannot reach on their own.
For now the picture is settled enough to act on. AI shoppers are worth more, they are here, and they will not be pinned to a single front door. Build for that, and the conversion data is a gift. Bet on one assistant winning, and it is a warning.
Sources
If the shopper belongs to no single assistant, what is the one thing a merchant can own that an assistant cannot take away?
Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.