Home/News/AI Search Orders Up 15x at Shopify, But Checkout Is the New Bottleneck

SEO & GEO

AI Search Orders Up 15x at Shopify, But Checkout Is the New Bottleneck

2026-09-04 Search Engine Journal SEO & GEO
AI Search Orders Up 15x at Shopify, But Checkout Is the New Bottleneck
图:Search Engine Journal

Search Engine Journal reports that Shopify president Harley Finkelstein told the company's February 2026 earnings call that orders arriving through AI-powered search have grown 15 times since January 2025. Konstantin Klyagin, founder of testing agency QAwerk, argues that getting surfaced in an AI platform is the easy half — most of the friction sits downstream, at checkout.

Key facts

  • Effective: Already underway. The Shopify figure comes from the February 2026 earnings call; UCP, ACP and Agentforce Commerce all appeared within the last year
  • Who it hits: Merchants connected to or planning to connect to UCP or ACP — especially sites where product data lives in several systems and checkout endpoints are not idempotent
  • Do something?: Yes, but not by adding another protocol. Run three tests: load-test checkout the way an agent hits it, verify product data as a machine reads it, and confirm refunds clear on the first attempt

What happened

The facts in the piece: AI-search orders are routing through three protocols built in the past year — Google's Universal Commerce Protocol, OpenAI's Agentic Commerce Protocol, and Salesforce's Agentforce Commerce, which aligned with UCP rather than competing with it. Etsy sellers went live inside ChatGPT first, followed by Shopify merchants including Glossier, Spanx and Vuori. OpenAI has since pulled back from native in-chat checkout, moving purchases into retailer apps instead. QAwerk has tested more than 300 client projects since 2015, and Klyagin was explicit that he has not seen a verified production incident where an agent itself caused a client's checkout to fail — he declined to dress up an ordinary ecommerce bug as an agent failure. What his team does see is systems disagreeing about the state of a purchase: an inventory feed saying a variant is in stock while checkout says sold out; a timed-out request retried against an endpoint that is not properly idempotent; a refund clearing on the merchant side before the updated order state ever reaches the agent that initiated it. A person shrugs off that kind of inconsistency and works it out. An agent needs every API, every feed and every order status to already agree. His three tests: load-test the checkout API with many parallel calls rather than one slow human session; check product data accuracy the way a machine reads it, not the way a browser renders it; and verify refunds and returns clear correctly on the first attempt.

What it means for DTC sellers

The sites that should be nervous are the ones where product data exists in several places that can drift. The classic setup: price and stock live in WooCommerce, in a feed plugin, and in a cache layer. A human browsing sees nothing wrong. An agent picks a product from one source, checkout validates against another, and the transaction stalls in a state the agent cannot resolve — that is not a low conversion rate, that is an order that never completes, and you probably never see it in your admin. Second group: sites whose checkout endpoints are not idempotent. Agents retry on timeout as a matter of course, and a duplicate charge or duplicate order turns straight into a dispute. Third: refund flows that quietly rely on a human to catch partial failures, because no one is in the loop when a machine files the request. In priority order: spend half a day establishing a single source of truth for product data, so price and stock for a SKU are written in exactly one place and read everywhere else; then have an engineer run a concurrency load test against checkout, which is cheap and surfaces problems fastest; refund idempotency third. None of this depends on which protocol you end up adopting — it is the foundation all three sit on. Skip it and connecting a fourth AI platform just routes more failed transactions into the same hole. To be fair to the source: Klyagin says there is no verified agent-caused checkout failure in production yet, so this is forward-looking engineering advice rather than a live incident. The three tests improve things for human traffic too, so the downside is small.

Someone should be watching these changes for you

I run AI automation for independent stores — content, SEO/GEO, product listing, social, and keeping up with policy shifts like this one.

See how it works