As of November 8, 2026, the evidence on AI shopping agents points at an uncomfortable conclusion for online retailers: the models are not the bottleneck, the catalogue is. Visa tested AI shopping agents against a benchmark released on 5 November and found that 57.2 percent of runs ended in a successful purchase attempt, while only 34.2 percent of episodes passed all three runs; a separate Blend Commerce audit sent an agent to 200 UK and US retailers and scored how well each could be read and bought from. For an online store, the lesson is that product data quality has become a transaction prerequisite rather than a search optimisation detail.
The failure sits in the catalogue
The headline number that matters is not the 57.2 percent pass rate but the 54 percent of failed runs that still ended in a purchase. An agent that fails its task and buys anyway is an agent operating on partial information, which is exactly the condition in which a wrong variant, a stale price or a discount that does not apply gets through. That is a margin and returns problem before it is a technology problem.
The spread between models is wide, from 82.2 percent down to 28.9 percent, and Visa itself cautions that no model clearly leads because the confidence intervals overlap across the next eleven models. The same data contains a sharper signal for merchants: models that surfaced the most hidden product factors scored 70.2 percent against 44.4 percent for those that surfaced the fewest. Better questions produced better outcomes, but a question can only be answered if the listing holds the answer.
| Stage the agent tests | What it needs from the store | How it typically breaks |
|---|---|---|
| Access | A crawlable site and a permitted agent | Blocked crawler or an outright refusal to serve |
| Product discovery | Findable category and search paths | Products reachable only through filters an agent cannot operate |
| Product comprehension | Explicit attributes and variant identity | Size and colour buried in a free-text title |
| Basket to checkout | Stable SKU, live stock, one price | Availability cached at product level, price differs in cart |
Variant data is the weak point
Among the failure modes documented in agentic checkout work, the dominant one is unglamorous: the agent cannot bind a shopper's request to a single purchasable variant. Colour and size live inside the title, SKUs are reused or regenerated, and stock is held at product level rather than variant level. The agent reasons correctly and then hands intent to a data layer that cannot resolve it deterministically, which fails silently after the shopper has been told the item is on the way.
The retailer audit adds a second layer. Twenty-nine of the 200 retailers refused the agent outright, and none of them blocked AI in robots.txt, which means the refusal is happening somewhere the merchant cannot see in a log. Seven of the ten furniture retailers tested were missing barcode numbers, a field no shopper looks at and every machine needs. Readiness is uneven by sector, and the leaders show it is achievable: Heal's scored 94 among UK furniture retailers and Slumberland 90 in the US. YouGov data cited alongside the audit puts planned UK Black Friday AI use at 19 percent in 2025 and 34 percent in 2026, so the traffic is arriving faster than the fixes.
Price consistency across page, feed and cart
An agent compares what it read with what the checkout charges. When the feed carries a price that the live page or the cart contradicts, the agent either abandons the basket or completes it at a price the merchant did not intend, and promotional terms that depend on a threshold make this worse. Keeping one canonical price per variant, with discount conditions stated explicitly, is the minimum. It is also where price monitoring earns its keep: a scheduled comparison of feed, page and cart catches the drift before a campaign does.
Checkout friction an agent cannot negotiate
Required account creation, a multi-step address flow, a verification pop-up and a bot challenge tuned for pre-agent traffic all terminate an agent-led purchase. None of them is visible in a conversion report as an agent failure, because the session simply ends. Guest checkout, wallet payment and a short required-field list are the practical fixes, and they help human shoppers in exactly the same way.
What to fix in four weeks
The order of work matters more than the volume. Access and variant identity come first, because a perfect feed behind a blocked crawler converts nothing. Then availability at variant level, then price parity across surfaces, then checkout friction. Each step is testable on its own with a scripted purchase attempt on an edge-case variant.
Two caveats should travel with this. The Visa benchmark is a synthetic test with its own task design, so the pass rates describe the test rather than the market, and the model cost spread, from about 0.54 to 1.55 dollars per run at the top of the table, matters more to a platform than to a merchant. The retailer audit covers 200 sites with ten per sector, which is large enough to show variation and too small to rank a sector.
FAQ
Why does variant data break agent checkout more often than model quality?
Because the model resolves the shopper's intent correctly and then needs one unambiguous, purchasable variant to bind to. When size and colour sit inside a free-text title or a SKU is reused, that binding fails silently after the customer believes the order is placed.
How large is the readiness gap between retailers?
The Blend Commerce audit scored 200 UK and US retailers out of 100 and found wide variation, with sector leaders near the top of the scale while seven of ten furniture retailers were missing barcode numbers entirely.
Should price monitoring be part of agent readiness?
Yes. An agent compares the price it read with the price in the cart, so any drift between feed, page and cart either cancels the basket or completes it at the wrong price, and discount thresholds make the mismatch harder to spot.
What do the test results still not tell us?
Whether synthetic benchmark tasks match real shopper requests, how much of the model spread comes from site data rather than model capability, and why the refusing retailers declined to serve an agent.
