Only 23% of Merchants Can Tell AI Traffic From Human Traffic. That's a Problem for Your Fraud Rules.

For most of eCommerce's history, a simple rule held up reasonably well: if traffic behaved like a bot, it was probably fraud. That rule just broke. AI shopping agents now browse catalogs, fill carts, and check out on behalf of real customers, behaving exactly like the automated traffic that used to be an easy red flag. According to Visa's 2026 research, only 23% of merchants surveyed could actually distinguish AI-driven traffic from human traffic arriving on their sites. Everyone else is flying partially blind at exactly the moment fraud pressure is climbing.

That finding comes from Digital Commerce 360's AI Spotlight report on agentic commerce, drawing on data from Visa, PYMNTS Intelligence, J.P. Morgan, and fraud-prevention company Signifyd. It paints a specific, uncomfortable picture: the volume of both legitimate AI shopping activity and AI-assisted fraud are rising together, using the same signals, at the same time, and most stores don't yet have a reliable way to tell them apart.

For more on where AI agents are already operating inside your store, see our piece on AI agents touching your checkout.

'Bots Used to Be Bad.' Not Anymore.

Nicole Jass, senior vice president of enterprise strategy at Signifyd, put the shift plainly in the report: "The big shift is bots used to be bad. And so you could say, 'Hey, we got a bot over here.' And once you figured out it was a bot, then we could say: This is likely fraud. Because it was really just the fraudsters sending the bots. Now, the good guys are sending the bots."

That single sentence describes the whole problem. A detection model built on "automated behavior equals risk" now flags your best customers' AI shopping assistants right alongside actual attackers, or worse, waves both through because the model can't tell the difference anymore. Worldpay's head of partnerships for AI and agentic commerce, Harsh Mehta, gave a concrete version of the same shift: "Forty orders in a second once meant an attack. Now it might be an agent simply doing its job. Retailers need models that can tell those apart."

The Volume Problem Is Real, and It's Getting Worse Fast

This isn't a theoretical shift. According to Signifyd's own data cited in the report, fraud pressure increased 33% in the first third of 2026 compared to the same period in 2025. Card-testing attacks specifically, the automated process of validating stolen card numbers by running small transactions, rose 175% over the same window, with AI making it easier to launch more sophisticated attacks at greater scale.

J.P. Morgan Commercial & Investment Bank reported a similar trend from a different angle: attempted fraud volume has increased by more than 20% annually over each of the last five years. The bank is fighting that increase with AI of its own. "It's continuous adaptation," said Joel Kalamba, the bank's business analysis director, in the report. "Fraudsters evolve, but so do we, using AI to help detect subtle anomalies and sophisticated attacks." Meanwhile, legitimate AI shopping traffic is climbing at an even steeper rate: Adobe Analytics data cited in the report shows AI-referral traffic to U.S. retail sites was up 62% year over year in July 2026, and up 1,219% compared to October 2024. Both curves are climbing together, on top of each other, using overlapping signals.

Both lines point the same direction for a reason. The same AI capabilities that make agentic shopping cheap to build also make automated fraud cheap to run, and they're scaling at the same time.

New Fraud Vectors AI Specifically Enables

The report also documents fraud tactics that didn't really exist as a meaningful category before generative AI made them cheap. Jass described return-abuse cases where consumers use generative AI to create images depicting damaged packaging, then use those fabricated images to claim a chargeback while keeping the undamaged goods, a scheme easy to spot once you know to look for it, since the same generated image often gets reused across multiple claims.

She also described a lower-effort version of an older attack: fake storefronts. Where building a convincing scam site used to require real time and technical skill, generative AI has made it possible to "spin up a fake shop in an afternoon without having to know how to code," complete with fabricated sales, stolen branding, and phishing pages designed to harvest payment details. Both examples point to the same underlying shift: AI hasn't just changed how attackers automate transactions. It's lowered the cost of the fraud itself.

Why Rules-Based Fraud Detection Is Losing Ground

Jass was direct about where static, rules-based fraud systems are heading: "It's going to get more complex before it ever gets simple... You're going to see a lot of different modalities, different types of attacks, different protocols." She described the accumulation of manual rules over time as building "this Jenga puzzle," where removing any one rule risks unknown consequences, and where each new rule is, by definition, reactive, written only after something bad has already happened once.

The alternative the report points to is behavioral and device-level signal analysis that doesn't rely on "is this automated" as the deciding question. Jass described the kind of granular data available to fraud-prevention platforms, devices, IP addresses, emails, SKUs, whether a purchase involved 15 gift cards stacked on a $1,000 order, or two items a shopper spent 30 minutes deciding between, as the material that actually separates legitimate agentic activity from an attack, precisely because it doesn't depend on automation as the signal.

What This Means for How You Should Be Evaluating Fraud Tooling Right Now

If your store's fraud detection still leans on rules written for a pre-agentic internet, traffic that looks automated, rapid-fire requests, scripted-looking browsing patterns, no mouse movement, you're accumulating false positives against real customers using legitimate AI shopping tools, and probably still missing more sophisticated attacks that have adapted around your rule set entirely.

• Audit your current fraud rules for anything that treats automated behavior itself as a risk signal, rather than a neutral fact that needs more context
• Confirm whether your fraud or payments provider distinguishes agent-level identity, which agent, on whose behalf, within what limits, rather than just flagging non-human traffic patterns
• Watch for AI-enabled return abuse specifically, generative images used to fabricate damage claims, and check whether the same image is being reused across multiple disputes
• Treat fraud detection as an ongoing, adapting system rather than a fixed rule set, the report's own framing: this only gets more complex from here, not less

Tracking your own AI-referral traffic growth the way Adobe's data tracked the broader market, so a spike doesn't look anomalous when it's actually a legitimate new acquisition channel, is squarely an analytics and attribution problem.

The Takeaway

The old shorthand, automated equals suspicious, was never perfectly accurate, but it was good enough to lean on. It isn't anymore, and the data in this report makes the size of that gap concrete: rising legitimate AI traffic, rising fraud pressure, and most merchants still unable to reliably tell the two apart at the door. The businesses that adapt their fraud logic before that gap costs them a real incident, rather than after, are the ones who'll actually capture the AI shopping traffic instead of either blocking it by mistake or getting burned by the fraud hiding inside it. That's exactly the gap our AI agent implementation service is built to close.

Frequently Asked Questions

Not Sure if Your Store Can Tell a Legitimate AI Shopper From an Attack?

SnapTec helps eCommerce businesses implement AI agents and evaluate the fraud and identity infrastructure around them, so agentic traffic gets captured as revenue, not blocked as risk or missed as fraud. Talk to our AI team →