Retailer checkouts break when an AI agent sends many quick requests

Search Engine Journal published a column on 4 September 2026 in which Greg Jarboe interviewed QAwerk founder Konstantin Klyagin about what happens after an AI agent selects a product. Klyagin's answer, tested across more than 300 client projects, is that getting a product shown in ChatGPT or Perplexity is the easy half, and that most of the current problems come at a later step, in checkout systems that have never been tested against agent traffic patterns.

The number the article started from

Shopify President Harley Finkelstein said on the company's February 2026 earnings call that orders arriving through AI-powered search have grown 15 times since January 2025. Those orders route through three protocols built in the last year: Google's Universal Commerce Protocol, OpenAI's Agentic Commerce Protocol, and Salesforce's Agentforce Commerce, which chose to align with UCP. Etsy sellers went live inside ChatGPT first, followed by Shopify merchants including Glossier, Spanx, and Vuori. OpenAI has since moved purchases out of native in-chat checkout into retailer apps, which makes the retailer's own checkout software more important.

The failure pattern

Klyagin has not seen a verified production incident where an agent itself caused a client checkout to fail, and refused to invent one. What his team keeps finding is systems disagreeing about the state of a purchase.

On a set of sales pages called Pridefit, two components each kept their own copy of the plan data, with small differences in pricing and attributes. A human shopper might refresh the page and continue. An AI agent has no visual context, so a mismatch between what it selected and what checkout validates against can stop the transaction in a state the agent cannot resolve. Klyagin's team fixed the case by centralising the plan data.

The wider pattern he expects across agentic commerce is inventory feeds saying a variant is in stock while checkout says it is sold out, a timed-out request retried against a non-idempotent endpoint, and a refund clearing on the merchant side before the updated order state reaches the agent that initiated it. Rate limiting and bot detection built to catch automated behaviour also catch legitimate shopping agents, because that is what a legitimate shopping agent looks like.

The three tests he tells retailers to run first

Load-test the checkout API the way an agent uses it: many parallel calls in quick succession, against a system tuned for one slow human session at a time. Check product data accuracy the way a machine reads it, so the two sources that supply data to a page cannot silently disagree. Verify that refunds and returns clear correctly on the first attempt, because no customer service person is involved when a machine initiates the request.

QA splits into two parts

For an engineering team supporting a commerce site, the useful point is where the QA work now belongs. Structured data, catalog feeds, and schema markup get a product included in an AI answer. None of that says whether an agent can complete the purchase once it arrives. Klyagin's view is that QA will split into two coordinated parts: one that validates the human experience, one that validates whether an agent can parse the data, complete the API calls, and get a predictable result at machine speed. The practical step for the current quarter is to run the three tests above before releasing the next AI-commerce integration.

Source

Reporting: Getting Your Product Into ChatGPT Isn't The Hard Part, Getting It Through Checkout Is by Greg Jarboe, Search Engine Journal, 4 September 2026.