Visa Tested 22 AI Shopping Agents. Nearly Half the Runs Failed.
The next customer on your website might not be a person. AI agents that browse, compare and buy on someone's behalf are moving from demo to reality, and on 5 November 2026 Visa published the first industry benchmark of how well they actually shop. The short answer: not very reliably yet.
What Visa found
Visa put 22 AI models from 12 labs (opens in a new tab) through 15 realistic shopping challenges, giving each model three attempts at every one.
- Overall, 57.2% of runs passed. Put the other way, almost half failed.
- The best models passed 82.2% of attempts. The worst managed 28.9%.
- A model got the same task right on all three attempts only 34.2% of the time.
- In 54% of the failed runs, the agent still bought something. Just not the right thing.
Fortune reported (opens in a new tab) that agents from Google and Meta shared the top score.
Why the failures matter more than the pass rate
An agent that gives up is an inconvenience. An agent that confidently buys the wrong size, the wrong variant or the wrong delivery option is a return, a refund request and an unhappy customer, and that lands on the business, not the AI company.
The inconsistency is the other warning sign. A tool that gets the same task right one day and wrong the next is hard to trust with anyone's money, which is why every serious agent product still asks a human to approve the purchase.
What this means for your website
Agents read a website very differently from people. They don't admire your photography. They look for clear, consistent facts: what the product is, what it costs, whether it's in stock, how delivery works and what the returns policy says. When that information is buried in images, hidden behind pop-ups or worded differently on different pages, both people and agents get it wrong.
The good news is that the fixes are the same ones that already help real customers and Google:
- Prices, options and availability written as text, not baked into images.
- Structured data (schema markup) that spells out products, prices and opening hours in a format machines read directly.
- One clear version of every policy, so delivery and returns don't contradict each other across pages.
- A short, predictable checkout, without surprise steps at the end.
What to do now
You don't need an "AI agent strategy" this month. You need a website whose basic facts are clear, consistent and machine-readable, because that's what agents will lean on as they improve. If yours is built on an ageing template that makes this hard, our website rebuild package puts those foundations in from day one, and keeps the Google rankings you already have.
For the bigger picture on where agents are heading, read our piece on AI that works in the background.
Keep reading
We Sell AI. We Still Think Most AI-Written Marketing Is a Mistake.
Ad executives think customers like AI-generated marketing. Customers mostly don't. Why we keep AI out of your voice and put it to work behind the scenes.
AI That Works in the Background: October 2026's Big Shift
Anthropic, OpenAI and Meta have all launched AI agents that work on their own and check in with you. Here's what that means for a UK small business.