AI sends them. Your store still has to close them.

For about eighteen months the assumed future of commerce was that the whole transaction would move inside the assistant. You'd ask ChatGPT for a running jacket, it would find one, and you'd buy it without ever seeing a storefront. Merchants would become suppliers to an interface they didn't control.

That's not what happened, and the correction is more interesting than the original prediction.

What actually happened

OpenAI shipped Instant Checkout in ChatGPT in September 2025 and quietly removed it around March 2026, with only a few dozen merchants live. The stated problems were mundane and structural: stale scraped pricing, stale inventory, an integration burden that didn't scale.

The more informative data point came from Walmart, which measured in-ChatGPT checkout converting roughly three times worse than simply clicking through to walmart.com, around 1.18% conversion with cart abandonment near 77%. Over the same period, ChatGPT was driving roughly twice the new-customer rate of search engines.

Those two findings belong together. The AI is very good at bringing you a shopper and bad at closing one. Walmart's response is the useful part: rather than retreat, they embedded their own conversational agent inside the assistant, and conversion recovered to around seventy percent of their site rate. A conversation layer, owned by the retailer, was what closed the gap.

Meanwhile the protocols formalised the division of labour. OpenAI and Stripe's Agentic Commerce Protocol, the Google and Shopify-led Universal Commerce Protocol, Shopify's Agentic Storefronts: all of it standardises how agents discover catalogs. Discovery is being commoditised by the platforms. Conversion on the merchant's own site is explicitly not.

The industry has settled on a sentence for it: discover in AI, buy on site. Which is the platforms saying, politely, that they'll send you the shopper and closing them is your problem.

The shopper who arrives mid-conversation

This is a genuinely new problem rather than the old problem at higher volume, and the difference is in what the shopper is carrying when they land.

A shopper from an AI referral has already had a conversation. They asked something specific, got a partial answer, formed a mental shortlist, and clicked through to resolve the remaining question. They arrive mid-thought, carrying context, expecting continuity. Then they land on a product grid, and everything they'd established (the use case, the constraint, the comparison they were weighing) has to be reconstructed by hand from a navigation menu and a filter sidebar. A static page cannot pick up a conversation it was never part of, and this shopper is more conversationally primed than any traffic source that came before.

Traffic from search behaved differently. Search sent you someone who had typed keywords and expected to browse. AI sends you someone who was in the middle of being helped and expects it to continue.

Gen-AI-driven traffic to US retail sites was up roughly 4,700% year over year as of Adobe's August 2025 report. Shopify reported AI-driven traffic to its stores growing around 8x year over year in Q1 2026, with orders from AI-powered searches up around 13x. Those are vendor-reported figures and worth verifying against your own analytics, which you can do today, because this segment is visible in everyone's referral data right now. That's the part that makes it urgent rather than speculative: you can go and look instead of taking a forecast on trust.

What this changes for a normal store

Walmart could build Sparky. You can't, and you shouldn't try.

The lesson generalises anyway, and it doesn't require Walmart's engineering organisation. What recovered their conversion wasn't a proprietary model. It was the presence of a conversation layer, owned by the retailer, positioned where the shopper actually lands. For a store doing a few thousand visitors a month, that's an install decision rather than a roadmap.

There's a second consequence that gets discussed less and may matter more. As more of your traffic arrives from opaque AI surfaces, your ability to explain your own numbers degrades. Referral data from assistants is thin. You'll see sessions appear with less context than you're used to, and the usual attribution scaffolding gets less reliable exactly as the segment grows.

So being able to say that this specific shopper asked this specific question and then bought this specific thing stops being a reporting nicety. For a growing share of your traffic it becomes the clearest signal you have, which makes conversation-level evidence a measurement layer as much as a proof that a tool works.

The honest uncertainty

Two things could falsify this.

If in-chat checkout un-stalls and agents start completing purchases off-site at scale, the on-site conversion layer matters less. The 2026 retreat was about integration quality, not about a law of nature, and those problems are solvable with enough effort. We think the Walmart conversion data argues it's harder than it looks, since buying is a higher-trust act than searching and people appear to want to see the store, but we could be wrong about that.

And if AI-referred traffic turns out to convert perfectly well on static pages, the whole argument weakens. Current evidence points the other way. The honest position is that this should be measured on real stores rather than asserted.

The direction of the traffic isn't uncertain. Every quarter, more of your visitors arrive having already had a conversation somewhere else, and land somewhere that can't have one. You can measure the size of that gap in your own referral data this quarter, and it will be a larger number next quarter.