Agentic Commerce 2026: Discover in AI, Buy on Site

A year ago the story was that AI agents would buy things for you. You would tell ChatGPT to find a waterproof jacket under 120 euros, and it would compare stores, pick one, and pay, all without you visiting a single website. OpenAI and Stripe shipped Instant Checkout inside ChatGPT in late September 2025 to prove the point, starting with Etsy and a wave of Shopify merchants. It looked like the checkout button was about to move off your site for good.

That is not how 2026 played out. OpenAI retired Instant Checkout in March, roughly six months after launch, and repositioned ChatGPT around product discovery instead of payment. The reason was not broken technology. It was shopper behavior: people happily used AI to research and shortlist, then finished the purchase where they already had an account, a saved card, and their order history. Walmart measured in-chat checkout converting about three times worse than sending the shopper through to Walmart's own site, even though ChatGPT brought roughly twice the rate of new customers that search does. The lesson landed fast. The money is still made on your site. AI just decides who gets sent there.

What actually changed in 2026

The headline is a split. Discovery moved into AI assistants; the transaction moved back to the merchant. ChatGPT now behaves more like a recommendation layer that hands off to your storefront. Google leaned into its Universal Commerce Protocol and a shared cart across Search and Gemini. Perplexity kept an "instant buy" path but, like everyone else, is fighting the same conversion gap once real money is on the line.

For a store owner the practical takeaway is cleaner than the hype of a year ago. You are no longer being asked to bolt a new payment rail onto your checkout to survive. You are being asked a simpler question: when an AI assistant is deciding which three products to name, is yours one of them, and does the click it sends you actually convert? This is closer to a new referral channel than a new checkout. It behaves like the early days of Google Shopping, not like a replacement for your website.

Why "buy on your own site" is good news for merchants

The version of agentic commerce where the agent pays inside ChatGPT quietly took a lot from the merchant. You lost the post-purchase relationship, the upsell, the email, the branded unboxing, and a chunk of margin to whoever owned the checkout. The 2026 correction hands most of that back. The shopper lands on your site, in your funnel, with your retargeting and your account system intact.

It also plays to a strength smaller studios and independent brands actually have. You cannot outspend a marketplace on ads. But you can make your product data clean, your delivery promises explicit, and your pages fast enough that an assistant trusts them, and a well-built custom storefront converts the traffic once it arrives. That is a fixable, on-site engineering problem, not an auction you are destined to lose. If you are weighing what that storefront should cost to build or rebuild, we broke the numbers down in what an ecommerce website costs in 2026.

How to get shortlisted by an AI assistant

Getting named by an assistant is mostly about being legible to a machine that never sees your hero image. Five things move the needle:

• Structured product data. Clean titles, prices, availability, shipping windows and return terms in Product and Offer schema. This is what an assistant reads to compare you, and it is the same groundwork that helps you get cited in AI answers. We covered the markup in structured data for AI search.
• Explicit delivery and stock. "Ships to Lisbon by Friday" is a filter an agent can act on. "Contact us for availability" is a reason to skip you.
• A site an agent can read and act on. AI browsers like ChatGPT's Atlas and Perplexity's Comet now navigate sites on a user's behalf. A storefront that only works after three JavaScript-heavy clicks loses to one an agent can parse in one. See making your website agent-ready.
• Reviews and real specifications. Assistants lean on corroborated detail. Thin product copy with no specs gives a model nothing to recommend on.
• Being present in the discovery surfaces. Feeds, marketplaces the assistant already reads, and content that answers the buying question. Generative engines cite sources they can find; generative engine optimization is how you show up in the first place.

Then convert the traffic you were sent

Winning the shortlist is only half of it. AI-referred shoppers arrive further down the funnel than a cold search visitor: they have already been told your product fits. If they land on a slow page, a broken mobile checkout, or a form that asks for an account before showing the price, you waste the one advantage the assistant just handed you. The Walmart number cuts both ways. In-chat checkout converted worse, but the click-through converted well precisely because the shopper finished in a place built to close.

So the on-site work is the work. Fast pages, a checkout that does not fight the buyer, honest delivery dates, and a follow-up that turns a one-time agent referral into a returning customer. None of that is new. Agentic commerce just raised the stakes on getting it right, because the traffic is more qualified and less forgiving.

Where to put your effort now

Ignore the noise about payment protocols unless you sell in a category where in-chat checkout genuinely returns. For almost everyone, the durable strategy for the next year is discovery in AI, conversion on your site. Make your catalog legible to machines, keep your storefront fast and readable to an agent, and treat AI assistants as the highest-intent referral channel you have. The stores that quietly did this groundwork are already collecting the clicks the panic of a year ago said would never come.

If you want a second read on whether your product data and storefront are ready to be recommended and to convert what gets sent your way, tell us what you sell and how and we will help you find the gaps before your competitors close theirs.