I present some professional bono consulting to a retailer positioned on the Higher East Aspect of New York Metropolis, and at our final video assembly, we lined some new floor. Structured information, catalog feeds, a connection to Google’s Common Commerce Protocol or OpenAI’s Agentic Commerce Protocol. However we didn’t speak about what determines whether or not a sale occurs: As soon as an AI agent, not an individual, is the one finishing the transaction, does the checkout beneath nonetheless work?
Shopify President Harley Finkelstein answered a part of that query on the corporate’s February 2026 earnings name, and the quantity isn’t small. Orders arriving by means of AI-powered search have grown 15 instances since January 2025 and are already routing by means of three separate protocols constructed within the final yr: Google’s Common Commerce Protocol, OpenAI’s Agentic Commerce Protocol, and Salesforce’s Agentforce Commerce, which selected to align with UCP relatively than construct a competing customary. Etsy sellers went stay inside ChatGPT first, with Shopify retailers together with Glossier, Spanx, and Vuori following. OpenAI has since pulled again from native in-chat checkout, shifting purchases into retailer apps as an alternative, which makes the underlying query sharper relatively than much less related.
I emailed Konstantin Klyagin to search out out what occurs after that. He based QAwerk in 2015 to present software program a correct testing accomplice, and the company has since examined greater than 300 shopper tasks throughout North America, Europe, and Africa. His reply to the visibility query was getting a product surfaced in an AI platform’s outcomes is the straightforward half, however many of the present friction sits downstream, within the half no one is testing but.
An Agent Outlets Nothing Like A Individual
Klyagin’s framing is straightforward when you hear it: A human shopper browses at an inconsistent tempo, will get distracted, abandons a cart, and comes again to it hours later. An AI agent fires speedy, structured API calls, evaluates a product towards the standards it was given, and executes a choice in seconds. That velocity is strictly what breaks methods tuned for people.
Price limiting and bot detection exist to catch habits that appears automated, which is exactly what a professional procuring agent seems to be like. Session logic constructed round one steady human go to chokes on an agent that queries a product, closes the session, and returns later to complete the acquisition. Klyagin’s crew has examined multi-agent methods in different regulated industries and retains discovering the identical root trigger: Most QA plans confirm whether or not a system produces the right output, and virtually none confirm whether or not the encompassing infrastructure tolerates a non-human actor shifting by means of it at machine velocity.
That is the place a well-ranked, well-optimized product nonetheless fails to transform. The web optimization, and AI-visibility work most retailers are centered on proper now sits fully upstream of it.
The Failure Sample Isn’t What You’d Guess
I requested Klyagin for an actual instance of a checkout, product-data, or refund failure induced particularly by an AI agent, anticipating a dramatic story. He hasn’t seen a verified manufacturing incident the place an agent itself induced a shopper’s checkout to fail, and he was not keen to decorate up an abnormal ecommerce bug as an agent failure and is strictly why his precise reply is price greater than a manufactured anecdote.
What his crew has discovered, repeatedly, is a subtler downside that turns into critical the second the customer is software program as an alternative of an individual. On one shopper venture, a funnel referred to as Pridefit, engineers discovered that two separate parts had been sustaining their very own copies of the identical plan information, with small variations in pricing, and attributes between the 2. A human shopper would possibly by no means discover, or would possibly simply refresh the web page. An AI agent has no visible context and no judgment to fall again on. If it selects a plan primarily based on one information supply and checkout validates towards the opposite, the mismatch in value, SKU, or availability can stall the transaction in a state the agent can not resolve by itself.
Klyagin’s crew eliminated the duplication and centralized the plan information, so each a part of the funnel pulled from one supply. However the sample he expects to see most frequently throughout agentic commerce typically isn’t an agent selecting the incorrect product. It’s methods disagreeing in regards to the state of a purchase order: A list feed says a variant is in inventory whereas checkout says it’s bought out, a timed-out request will get retried towards an endpoint that isn’t correctly idempotent, or a refund clears on the service provider’s facet earlier than the up to date order state ever reaches the agent that initiated it. An individual can typically shrug off an inconsistency like that and work out what really occurred. An agent wants each API, each product feed, and each order standing to already agree.
3 Checks Value Operating Earlier than You Chase A Fourth Protocol
Klyagin factors purchasers towards three particular checks, and I feel each retailer at the moment centered on catalog sync and structured markup ought to run all three earlier than including a fourth AI platform to the record.
- Load-test the checkout API the way in which an agent really hits it. Not one sluggish human session at a time, however many parallel calls fired in fast succession. A checkout that has dealt with thousands and thousands of human periods with out incident can nonetheless fail the primary time it meets that visitors sample, and most retailers connecting to UCP or ACP proper now genuinely don’t know whether or not theirs will.
- Test product information accuracy the way in which a machine reads it, not the way in which a browser renders it. A web page that appears completely constant to a human customer will be pulling from two disagreeing sources beneath, and an agent has no approach to discover the hole the way in which an individual scrolling the web page would possibly.
- Confirm refunds and returns clear accurately on the primary try. When a machine initiates the request, there’s no customer support rep within the loop to catch a partial failure or a standing that by no means syncs again.
None of those three requires ready for UCP, ACP, or Agentforce Commerce to mature additional. They check the inspiration all three protocols depend upon no matter which one, or which mixture, finally ends up main the market.
My Take
I feel the business has the sequencing backward. Everyone seems to be racing to get listed inside ChatGPT and Gemini earlier than checking whether or not their checkout can really full the sale as soon as an agent will get there, and that’s constructing visibility on prime of a basis no one has load-tested. Technical web optimization earns a product a spot in an agent’s outcomes. It has nothing to say about whether or not the agent can purchase it, and proper now that second, tougher downside is the one virtually nobody is engaged on.
Klyagin expects QA to separate into two coordinated tracks over the subsequent couple of years. One which retains validating the expertise a human has, and a second that validates whether or not an agent can parse the information, full the API calls, and get a predictable consequence when it strikes by means of the system at machine velocity. That’s roughly the identical shift ecommerce QA groups went by means of constructing mobile-specific check suites a decade in the past, and the retailers who deal with it as an engineering precedence now, forward of the quantity shift Shopify’s personal numbers present already underway, are going to have an actual head begin over everybody nonetheless centered solely on getting discovered.
In case your ecommerce technique for 2026 stops at getting surfaced in an AI platform’s outcomes, then you definitely’ve solved the a part of the issue that was by no means really the arduous half.
Extra Assets:
Featured Picture: tete_escape/Shutterstock
