On September 16, Google told retailers two things about AI shopping. One of them got the attention. The other is the one that will show up in your analytics.
The first is that AI performance insights in Merchant Center is now generally available in the US, Canada, Australia, India and New Zealand. It gives you a share-of-voice comparison — how your brand and products are being surfaced across AI Mode and AI Overviews, measured against other brands.
The second is buried in a paragraph about the Universal Commerce Protocol integration hub. Merchants using the hub for UCP-powered checkout can now also enable cart transfer to a merchant site. Google notes that analytics for the hub are “coming soon.”
Read those two together and the shape of the problem changes.
The funnel now ends on your storefront
For most of the agentic commerce conversation so far, the assumption has been that the transaction happens inside the AI surface. Google’s own framing supports that. UCP already enables direct checkout for hundreds of thousands of brands across Google, and Tapestry has Coach and Kate Spade selling inside Search, AI Mode and the Gemini app.
Cart transfer is the opposite motion. The shopper does the discovery and the selection in the AI surface, and then arrives on your site with a cart already assembled, to finish. Plenty of merchants will prefer that. It keeps the customer relationship, the post-purchase journey, the loyalty logic and the data on your side.
It also means the last step of an AI-mediated purchase runs on your infrastructure, on your slowest page, under your third-party stack.
Why this particular session is the hard one
Think about what that arriving session actually looks like. No homepage. No category browse. No product page. The shopper lands directly on cart or checkout, from an external surface, with an empty cache and no prior requests warmed.
That is the least forgiving entry pattern in eCommerce, and it lands on the page type where third-party code is densest. Cart and checkout are where payment providers, fraud tools, tax calculators, chat widgets, loyalty scripts and analytics beacons all stack up. Across eCommerce sites, third-party applications account for around 44% of total page load time, and each additional tool in the stack is associated with a 0.29% reduction in conversions.
It is also a session with unusually little patience behind it. Someone who has already chosen the product in a conversation with an assistant has done the deliberation. What is left is a transaction they expect to take seconds. A slow cart page at that moment is not a bounce risk in the ordinary sense. It is a transaction that was already won being handed back.
And almost nobody is measuring it as its own thing. In a standard analytics setup, that session is a referral from a Google property landing on a cart URL. It disappears into the average. Google’s hub analytics are not shipped. Your own reporting almost certainly does not segment it.
What Google is measuring, and what it isn’t
AI performance insights tells you how often you surface in AI results. The feed guidance, covering conversational attributes, loyalty data and the video link attribute, is about getting surfaced more often. Google reports that merchants adopting core Merchant Center feed best practices see a 5% conversion increase the following month, and that in testing with lululemon, conversational attributes were incorporated 50% of the time in relevant AI Mode product recommendations. Useful numbers, both about discovery.
Nothing in the announcement tells you what happens in the seconds after the handoff. That’s not a criticism of Google (it’s not their page). But it does mean that if you optimize only for what Google now measures, you will get better at being found, yet learn nothing about whether or not you can convert.
What to actually do about it
Three things, in order, and none of them require a decision about agentic commerce strategy:
- Segment cold entries into cart and checkout and look at what they cost. Not your sitewide average, not your product page numbers. The specific pattern of an external referral landing directly on a transactional page with no warmed cache. If you have never looked at that cohort separately, the number is unlikely to be what you expect.
- Inventory what loads on cart and checkout and rank it by measured cost rather than by importance. Most stores have never done this by page type, and the results usually include at least one script that has no business firing before the pay button renders.
- Decide about cart transfer. It is a good option for a lot of retailers. It’s a better one if you know what the landing page does under a cold load.
The broader point
Every previous round of the AI commerce conversation has been about someone else’s surface: whether the model would cite you, whether the agent could be trusted, whether the crawler would be blocked. Those were all arguments about infrastructure you do not own.
Cart transfer brings it home. The moment the agentic funnel hands off, it becomes an ordinary web performance problem on an ordinary storefront, and it is decided by the same things that decide every other high-intent session: what loads, in what order, and how quickly the shopper can act.
If AI is going to send you the most qualified traffic you will get this quarter, the least you can do is be ready to receive it.
Subscribe to our blog
Performant Experimentation – A Guide to A/B Testing
Every A/B test you run is a purchase. You’re buying information—an answer to a hypothesis—and you’re paying for it in degraded performance and lost conversion during the test window.
Download eBook
