On March 4, 2026, Walmart’s EVP of global technology, Daniel Danker, told a Morgan Stanley conference that buying inside ChatGPT converted at roughly one-third the rate of sending the same shopper to walmart.com. He called the in-chat checkout experience “a very temporary moment in time.” Five days later, reporting from The Information and Modern Retail confirmed what merchants had started to suspect: OpenAI was quietly walking back the feature it had launched with enormous fanfare six months earlier.
If you sell anything online, you were probably braced for the opposite story. The narrative all year was that AI agents would do the shopping for your customers, complete the purchase inside a chatbot, and leave you as an anonymous supplier feeding a platform that owned the buyer, the payment, and the margin. That is not what the 2026 data shows. In-chat checkout stalled. The buying is still happening on merchant sites, and the shoppers that AI sends to those sites are converting better than the ones coming from Google or Instagram.
That reversal changes what a small online seller should actually be doing right now. It is worth walking through the numbers, because most of the advice being sold about “preparing for agentic commerce” is preparing store owners for the wrong thing.
What actually happened to in-chat checkout
OpenAI shipped Instant Checkout inside ChatGPT on September 29, 2025, alongside an open Agentic Commerce Protocol co-developed with Stripe. The pitch was that a shopper could complete a purchase without ever leaving the conversation, and that “over a million” Shopify merchants would be reachable. By February 2026, Forrester counted roughly 30 Shopify merchants actually live. In March, OpenAI told Modern Retail that Instant Checkout was “moving to Apps, where purchases can happen more seamlessly.” The million-merchant surface never arrived.
The reason is not mysterious. In-chat checkout strips away the things that make people comfortable spending money: the product page, the reviews, the return policy in plain sight, the familiar cart. A Semrush survey of 1,030 US shoppers in December 2025 found that only 22 percent had ever completed a purchase directly inside an AI tool, while 50 percent had bought something after using AI to research it, usually somewhere else. People are happy to let a model help them decide. They are far less happy letting it hold their credit card.
Google read the same tea leaves and built the softer version. Its agentic checkout, which began rolling out on November 13, 2025 with Wayfair, Chewy, Quince, and select Shopify merchants, does not silently transact. It watches a price you set, notifies you when the item drops into your budget, and then asks you to confirm before it buys through Google Pay. The purchase still resolves against the merchant’s listing in a Shopping Graph of more than 50 billion products. Google’s newer Direct Offers pilot and its Universal Commerce Protocol point the same direction: discovery and matching happen in AI, the transaction lands on or through the seller.
The industry has a phrase for the pattern that survived: discover in AI, buy on your own site.
The number that reframes the whole panic
Here is the part almost nobody predicted. When AI sends a shopper to your actual storefront, that shopper is not a worse customer. They are a materially better one.
Adobe, which tracks over a trillion visits to US retail sites, reported that in March 2025, AI-referred traffic converted 38 percent worse than other sources. By March 2026, the same measurement had flipped to 42 percent better. That is an eighty-point swing in twelve months. Over the 2025 holiday season, AI-driven retail traffic was up 693 percent year over year (769 percent in November, 673 percent in December), AI referrals converted 31 percent higher than other channels, and revenue per visit from those referrals ran 254 percent above the prior year. In the first quarter of 2026, AI traffic to US retailers grew 393 percent.
Salesforce, in its Agentforce Commerce release on July 6, 2026, put harder edges on it: AI-referred traffic now converts at roughly eight times the rate of social, revenue per visit from AI referrals sits about 37 percent above non-AI traffic, and retailers running their own shopper agents grew sales 59 percent faster than those on the sidelines. Salesforce’s own framing is blunt about where the money lands: referral traffic will be shaped by AI platforms, but “commerce itself will happen predominantly on owned and operated properties: your website, your app, your messaging channels, your stores.”
The story is not that AI is coming to take the transaction. The story is that AI has become a high-intent referral engine, and the endpoint of that referral is your store.
| Channel | How AI-referred traffic compares (2026 data) |
|---|---|
| vs. all other sources (Adobe, Mar 2026) | Converts 42% better |
| vs. social traffic (Salesforce, Jul 2026) | Converts ~8x higher |
| Revenue per visit vs. non-AI (Salesforce) | ~37% higher |
| In-chat checkout vs. merchant site (Walmart, Mar 2026) | Converts ~3x worse in-chat |
Why owning the last click is a margin decision, not a branding one
I spent 20-plus years in IT operations before moving into fractional COO work, and the lens that has been most useful for reading agentic commerce is not a marketing one. It is a working-capital and customer-data one. Whoever owns the checkout owns three things that decide whether a small business survives: the customer’s contact information, the timing of when cash reaches you, and the cut taken off the top.
When a platform completes the sale inside its own interface, it keeps all three. OpenAI’s Instant Checkout terms took a 4 percent fee on completed purchases, and the merchant received an order, not a relationship. No email to remarket to. No SMS list. No way to earn the second and third purchase that is where a store’s actual profit lives. That is the same structural problem I wrote about with Amazon’s shift to deducting ad spend from seller proceeds: when a platform controls the plumbing of the transaction, it controls your margin and your cash timing whether you like it or not.
Buy-on-your-site inverts that. The AI hands you a pre-qualified visitor, and you keep the checkout, the email capture, the payment relationship, and the repeat-purchase economics. For a one-person store, that difference is not cosmetic. It is the difference between renting customers and owning them.
What this actually changes for a one-person store
If the endpoint is your site, your job splits cleanly into two problems: getting AI to recommend you, and converting the traffic it sends.
The first is a discoverability problem, and it is closer to SEO than to anything exotic. AI shopping engines pull from structured product data and clean feeds. That means accurate product titles, complete specs, real review content, and a merchant feed that actually validates. If you run a Shopify or WooCommerce store, supporting the emerging protocols is increasingly a configuration toggle rather than an engineering project, so there is little reason not to. This is the same muscle as generative engine optimization: being the answer an AI model reaches for. The stores that were already invisible to AI shopping, a problem I dug into when 50 million shopping queries a day started hitting ChatGPT, are the ones leaving this new, higher-converting traffic on the table.
The second is a conversion problem, and it is unglamorous. AI-referred shoppers arrive with intent already formed, which is exactly why they convert well, but they abandon a slow, cluttered, or untrustworthy storefront just as fast as anyone else. Page speed, an obvious return policy, visible reviews, and a checkout that does not ask for an account before it asks for a card. None of this is new advice. What is new is that it now compounds against a traffic source that outperforms your paid channels, so the payoff for fixing it is larger than it was a year ago.
The third thing to do is capture the relationship the moment the AI hands it to you. Email or SMS opt-in at checkout is the entire point of preferring buy-on-site over in-chat: you get to keep the customer. Skipping that step throws away the one durable advantage this model gives a small seller.
The part that should still worry you
None of this means agentic commerce is harmless to small sellers. It concentrates power at the discovery layer. If Google’s Shopping Graph, ChatGPT, and Gemini become the front door to product search, then the seller’s dependency simply moves upstream, from ranking in ten blue links to being surfaced by a handful of models whose selection logic is more opaque and less gameable than a search algorithm. The same disintermediation risk that gutted publisher traffic through AI Overviews applies here, just one step further down the funnel.
And the “own the checkout” advantage holds only as long as the AI platforms keep routing shoppers out to merchant sites. If in-chat checkout conversion improves, or if a platform decides the economics of keeping the transaction are worth the friction, the door that is open today can narrow. Danker’s phrase, “a very temporary moment in time,” cuts both ways.
For now, though, the data is unambiguous and it favors the small operator more than the panic suggested. The most valuable thing you can build in 2026 is not a slot inside someone’s chatbot. It is a fast, discoverable, trustworthy store that converts the intent AI is already sending your way, and an email list built from the customers it delivers. That is a strategy you own. The chatbot’s checkout button is not.
Ty Sutherland runs a portfolio of content and commerce sites and works as a fractional COO through Ops Harmony, after 20-plus years in IT operations. He writes about the operational and financial mechanics behind online income, not the hype. More at earninglivingonline.com/about.