Getting Your Product Into ChatGPT Isn’t The Hard Part, Getting It Through Checkout Is via @sejournal, @gregjarboe

Getting Your Product Into ChatGPT Isn’t The Hard Part, Getting It Through Checkout Is AI shopping orders are up 15X on Shopify, but QAwerk's testing shows checkout, product data, and refunds aren't ready for agents moving at machine speed.

I provide some pro bono consulting to a retailer located on the Upper East Side of New York City, and at our last video meeting, we covered some new ground. Structured data, catalog feeds, a connection to Google’s Universal Commerce Protocol or OpenAI’s Agentic Commerce Protocol.

But we didn’t talk about what determines whether a sale happens: Once an AI agent, not a person, is the one completing the transaction, does the checkout underneath still work? Shopify President Harley Finkelstein answered part of that question on the company’s February 2026 earnings call, and the number is not small.

Orders arriving through AI-powered search have grown 15 times since January 2025 and are already routing through three separate 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 rather than build a competing standard.

Etsy sellers went live inside ChatGPT first, with Shopify merchants including Glossier, Spanx, and Vuori following. OpenAI has since pulled back from native in-chat checkout, moving purchases into retailer apps instead, which makes the underlying question sharper rather than less relevant. I emailed Konstantin Klyagin to find out what happens after that.

He founded QAwerk in 2015 to give software a proper testing partner, and the agency has since tested more than 300 client projects across North America, Europe, and Africa.

His answer to the visibility question was getting a product surfaced in an AI platform’s results is the easy half, but most of the current friction sits downstream, in the part nobody is testing yet.

An Agent Shops Nothing Like A Person Klyagin’s framing is simple once you hear it: A human shopper browses at an inconsistent pace, gets distracted, abandons a cart, and comes back to it hours later. An AI agent fires rapid, structured API calls, evaluates a product against the criteria it was given, and executes a decision in seconds.

That speed is exactly what breaks systems tuned for humans. Rate limiting and bot detection exist to catch behavior that looks automated, which is precisely what a legitimate shopping agent looks like. Session logic built around one continuous human visit chokes on an agent that queries a product, closes the session, and returns later to finish the purchase.

Klyagin’s team has tested multi-agent systems in other regulated industries and keeps finding the same root cause: Most QA plans verify whether a system produces the correct output, and almost none verify whether the surrounding infrastructure tolerates a non-human actor moving through it at machine speed.

This is where a well-ranked, well-optimized product still fails to convert. The SEO, and AI-visibility work most retailers are focused on right now sits entirely upstream of it.

The Failure Pattern Isn’t What You’d Guess I asked Klyagin for a real example of a checkout, product-data, or refund failure caused specifically by an AI agent, expecting a dramatic story.

He hasn’t seen a verified production incident where an agent itself caused a client’s checkout to fail, and he was not willing to dress up an ordinary ecommerce bug as an agent failure and is exactly why his actual answer is worth more than a manufactured anecdote.

What his team has found, repeatedly, is a subtler problem that becomes serious the moment the buyer is software instead of a person. On one client project, a funnel called Pridefit, engineers found that two separate components had been maintaining their own copies of the same plan data, with small differences in pricing, and attributes between the two.

A human shopper might never notice, or might just refresh the page. An AI agent has no visual context and no judgment to fall back on. If it selects a plan based on one data source and checkout validates against the other, the mismatch in price, SKU, or availability can stall the transaction in a state the agent cannot resolve on its own.

Klyagin’s team removed the duplication and centralized the plan data, so every part of the funnel pulled from one source. But the pattern he expects to see most often across agentic commerce generally is not an agent picking the wrong product.

It’s systems disagreeing about the state of a purchase: An inventory feed says a variant is in stock while checkout says it’s sold out, a timed-out request gets retried against an endpoint that isn’t properly idempotent, or a refund clears on the merchant’s side before the updated order state ever reaches the agent that initiated it.

A person can often shrug off an inconsistenc