AI shopping in 2026: discovery happens in the chat, the sale still happens on your site

For about eighteen months the pitch about AI and e-commerce was that the storefront was going away. Customers would ask an assistant for running shoes, the assistant would buy the running shoes, and your website would become a warehouse API that nobody looked at. Plenty of people sold consulting on that premise.

It did not happen, and in March 2026 the company most committed to it said so out loud. That is worth walking through carefully, because the thing that replaced it is real, it is growing fast, and what it asks of a small store is different from what everyone spent last year preparing for.

What actually happened

In September 2025, OpenAI launched Instant Checkout in ChatGPT. You could find a product in a conversation and buy it without ever leaving the chat window. Starting that November, Walmart put roughly 200,000 products through it.

Walmart then measured it. Daniel Danker, Walmart’s EVP of product and design, said purchases completed inside ChatGPT converted at about a third the rate of the same customers clicking through to walmart.com. He called the experience unsatisfying. When your biggest partner runs the experiment at that scale and reports a 3x conversion penalty, the experiment is over.

On March 24, 2026, OpenAI announced the change. Its own framing was that “the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide,” so it would let merchants use their own checkout experiences and focus its efforts on product discovery instead. There were practical reasons underneath the diplomatic language: real inventory levels, shipping costs, delivery estimates, and US state sales tax are all things a merchant knows and a chat window does not, and OpenAI had not built the merchant-of-record tax plumbing that being the checkout requires.

Walmart’s response tells you where this landed. Rather than sell inside ChatGPT, it put its own assistant, Sparky, inside ChatGPT as an app, with account linking, cart syncing, and loyalty intact. Discovery in the chat. Transaction, customer relationship, and data on Walmart’s side.

The model that won: discover there, buy here

What ChatGPT does now is product discovery. Browse visually, upload a photo to find something similar, compare options side by side on price and reviews and features, refine conversationally, then click through to the merchant to buy. Target, Sephora, Nordstrom, Lowe’s, Best Buy, The Home Depot, and Wayfair are all in.

Google arrived at the same architecture from the other direction. At the NRF Big Show on January 11, 2026, Sundar Pichai announced the Universal Commerce Protocol, an open standard co-developed with Shopify and backed at launch by more than twenty retailers and payment companies including Etsy, Wayfair, Target, Best Buy, PayPal, Visa, Stripe, and American Express. Google’s stated position is that the merchant stays merchant of record even when an agent is doing the shopping. Same idea as OpenAI’s Agentic Commerce Protocol, which Stripe and OpenAI built: the assistant handles finding, the merchant handles selling.

Two competing standards from two competing giants, both landing on “the store keeps the transaction.” That convergence is the actual news.

The traffic is not hypothetical

Adobe Analytics measured the 2025 holiday season, November 1 through December 31, across more than a trillion visits to US retail sites. Referral traffic from generative AI sources was up 693 percent year over year, 769 percent in November and 673 percent in December.

The quality numbers are the more interesting half. Visitors arriving from an AI assistant converted 31 percent better than traffic from other sources, roughly double the advantage they held the previous holiday season. They were 33 percent less likely to bounce, spent 45 percent more time on site, and revenue per visit from that channel was up 254 percent year over year. Salesforce separately estimated that AI tools influenced more than 20 percent of all online retail sales globally over the same period.

That combination makes sense once you think about what the assistant did before the click. It already narrowed the category, compared the options, and matched the product to a stated need. The person landing on your product page is not browsing. They are close to done. That is why they convert, and it is also why the page they land on matters more than it used to, not less.

What a small store should actually do

None of this requires rebuilding anything. Most of it is catalog hygiene you should be doing anyway, which is the usual shape of these things.

1. If you are on Shopify, go look at your agentic storefront settings. Shopify’s Agentic Storefronts feature makes your products automatically accessible to AI channels including ChatGPT, Shop, and Copilot through Shopify Catalog, and since May 11, 2026 it has had its own page in the admin at admin.shopify.com/agentic. It is on by default, which means the useful action is to confirm what is exposed and that it looks right, not to switch it on. There is no separate app and no extra transaction fee beyond your normal Shopify rates. This is the single highest-value five minutes available to a Shopify merchant right now, and it is one of the reasons Shopify is still our default recommendation for most stores.

2. Fix your product feed before you do anything clever. Both protocols run on structured product data: title, description, images, price, availability, brand, identifiers. Assistants that cannot tell whether an item is in stock will not recommend it, and one that recommends an out-of-stock item once learns not to trust the source. If you already maintain a clean Google Merchant Center feed, you are most of the way there, and Google’s simplified UCP onboarding runs through Merchant Center.

3. Put Product structured data on your own product pages, with price and availability. This is item one on our answer-engine optimization checklist applied to a catalog. It is how a machine reads your page without guessing, and unlike a feed it works for whoever shows up next.

4. Make sure your content survives without JavaScript. If your product name, price, and stock status only appear after a spinner finishes, a crawler frequently gets an empty page. We have written this at length and it keeps being the thing that quietly disqualifies otherwise good stores.

5. Treat the landing page as the conversion point, because it is. The whole model now hinges on a high-intent visitor arriving on your site and buying. A slow product page, a checkout with five steps, a shipping cost that appears at the last screen, these were always leaks. Now they are leaks on your best-converting traffic. Interaction to Next Paint is the metric to watch here.

6. Do not pay anyone to get you ranked in ChatGPT. OpenAI has said its results are not influenced by commercial partnerships. Anyone selling guaranteed placement in an AI assistant is selling something they do not have.

The honest summary

The scary version of this story, where AI assistants disintermediate your store and you become an anonymous supplier, got tested at Walmart scale and lost on the numbers. Customers do not want to buy in a chat window. They want to decide in a chat window and buy where they already have an account, a saved card, and a returns policy they trust.

What that leaves is a new referral channel that sends fewer visitors than search but sends them further along and converts them better. The work it asks for is accurate product data, pages a machine can read, and a fast checkout. Which is, unhelpfully for anyone selling a new methodology, exactly the work that has always paid off.

If you want someone to look at your product feed, your structured data, and how your store actually performs for a visitor who lands ready to buy, send us a note.