Discovery moved. The transaction came home.

For about eighteen months, agentic commerce came down to one question every founder asked in some version. Will people buy inside ChatGPT, and if they do, what happens to my store?

OpenAI’s March 2026 direction makes one thing clear: discovery in an assistant does not remove the need for a merchant’s checkout. That is a product decision, not a measurement of where most shoppers complete their purchases.

That is a better outcome than it sounds, and it changes what is worth building.

What actually happened, in order

On 29 September 2025, Stripe and OpenAI announced Instant Checkout and released the Agentic Commerce Protocol. A shopper could complete a purchase inside ChatGPT, with the merchant’s own systems holding the checkout state and processing the payment.

Google had announced its Agent Payments Protocol two weeks earlier, on 16 September, with more than sixty launch partners. A separate specification, for a separate layer of the same problem.

On March 24, OpenAI announced that Shopify Catalog already supplied product data to ChatGPT without individual merchants building a separate discovery integration.

The same announcement emphasized merchant-owned checkout experiences and product discovery. Deeper native integrations remained an option; this was not a declaration that every in-chat purchase had ended.

Source: Stripe: Stripe powers Instant Checkout in ChatGPT and releases the Agentic Commerce Protocol

Why the checkout came home

OpenAI said its initial Instant Checkout did not provide the flexibility it wanted. That is the primary-source explanation, and it is narrower than saying shoppers rejected agentic commerce.

For a merchant, the practical question is which checkout path supports the real basket: inventory validation, delivery options, promotions and account expectations. Test that path rather than assuming an embedded interface replaces those responsibilities.

Discovery and checkout can live on different surfaces. The handoff between them still needs to preserve an accurate product and a usable route to purchase.

Source: OpenAI: Powering product discovery in ChatGPT

The traffic is real, and it is different

Adobe’s measurement of US retail is the most useful public number here, because it is first-party, dated, and it reversed.

AI-sourced traffic to US retail sites grew 393% year over year in the first quarter of 2026. In March 2026, that traffic converted 42% better than non-AI traffic, spent 48% longer on site, viewed 13% more pages, and produced 37% more revenue per visit.

A year earlier, the same measurement had AI traffic converting 38% worse.

That reversal is the finding. Adobe compares AI referrals with aggregated non-AI traffic; it does not rank every channel or establish each visitor’s intent. A shopper may arrive after comparing options in an assistant. Treat that as a journey to accommodate, not a fact about every session.

The growth figure does not tell you AI’s share of your traffic. Report your own identifiable referral volume alongside conversion, and do not use an unrelated Shopify estimate as the denominator for Adobe’s US retail data.

Source: Adobe: AI traffic grows but retail sites lag in AI search visibility

Source: Adobe: Q2 2026 AI-sourced traffic insights (page 12)

What this means for your storefront

If the purchase happens on your site, then the assistant is a referrer, and the work is the work you already recognize.

The record has to be right so both people and connected systems can verify it. The landing should make format, quantity, price and supporting evidence easy to find. You cannot recover a shopper’s chat from a referrer alone, but you can make common purchase questions answerable without restarting at the catalog.

Adobe’s proprietary AI content visibility measure scored product pages at 66%, against 75% for homepages. That is a result from its audit, not a universal readability score or a measure of ranking and revenue.

Those page improvements can be made independently of a new protocol integration.

What we are not saying

We are not prescribing an adoption year. Choose agentic checkout when a defined journey, platform eligibility and operational capability justify it. Record accuracy and a usable landing page remain necessary whichever path you choose.

We are also not saying this gets you recommended. Appearing in an AI answer and being recommended by one are different problems, and the second is not something anyone can promise. You control the accuracy of the product information you publish. Check external answers separately and record any remaining errors.