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 inconsistency like that and figure out what actually happened. An agent needs every API, every product feed, and every order status to already agree.
3 Checks Worth Running Before You Chase A Fourth Protocol
Klyagin points clients toward three specific tests, and I think every retailer currently focused on catalog sync and structured markup should run all three before adding a fourth AI platform to the list.
- Load-test the checkout API the way an agent actually hits it. Not one slow human session at a time, but many parallel calls fired in quick succession. A checkout that has handled millions of human sessions without incident can still fail the first time it meets that traffic pattern, and most retailers connecting to UCP or ACP right now genuinely don’t know whether theirs will.
- Check product data accuracy the way a machine reads it, not the way a browser renders it. A page that looks perfectly consistent to a human visitor can be pulling from two disagreeing sources underneath, and an agent has no way to notice the gap the way a person scrolling the page might.
- Verify refunds and returns clear correctly on the first attempt. When a machine initiates the request, there’s no customer service rep in the loop to catch a partial failure or a status that never syncs back.
None of these three requires waiting for UCP, ACP, or Agentforce Commerce to mature further. They test the foundation all three protocols depend on regardless of which one, or which combination, ends up leading the market.
My Take
I think the industry has the sequencing backward. Everyone is racing to get listed inside ChatGPT and Gemini before checking whether their checkout can actually complete the sale once an agent gets there, and that’s building visibility on top of a foundation nobody has load-tested. Technical SEO earns a product a place in an agent’s results. It has nothing to say about whether the agent can buy it, and right now that second, harder problem is the one almost no one is working on.
Klyagin expects QA to split into two coordinated tracks over the next couple of years. One that keeps validating the experience a human has, and a second that validates whether an agent can parse the data, complete the API calls, and get a predictable result when it moves through the system at machine speed. That’s roughly the same shift ecommerce QA teams went through building mobile-specific test suites a decade ago, and the retailers who treat it as an engineering priority now, ahead of the volume shift Shopify’s own numbers show already underway, are going to have a real head start over everyone still focused solely on getting found.
If your ecommerce strategy for 2026 stops at getting surfaced in an AI platform’s results, then you’ve solved the part of the problem that was never actually the hard part.
- The New AI Marketplace: How ChatGPT’s Native Shopping Could Rewrite Digital Commerce
- Agentic Commerce Optimization: A Technical Guide To Prepare For Google’s UCP
- Google Expands UCP With Cart, Catalog, Onboarding
CategoryAI SearchEcommerceTechnical SEO
