The Maze: Microsoft Clarity, a free tool for analysing website visits, found Checkout events in 25.21% of visits involving AI agents, versus 14.65% of human visits. The October 29 study gives merchants a reason to examine automated browsing. It does not show that software completed those purchases.
Source: Microsoft Clarity. Original study table showing agent-involving session counts, percentages and the denominator definition.
The Details:
The visit can contain two operators. Clarity studied activity from August 13 to September 9 across 230,511 website projects. It detected browser providers in the OpenAI, Google and Anthropic ecosystems. An AI agent browses on someone’s behalf, but a person can take over within the same recorded visit. Clarity cannot separate their actions. Its new Agent Activity Dashboard is generally available to projects receiving agent traffic, helping businesses inspect these visits. That makes it a tool for investigating behaviour, rather than a receipt proving which participant bought something.
The percentage needs a clearer base. The study reports more than 5.2 million agent-involving visits overall. Yet its 306,621 Checkout visits divided by 25.21% imply a base of about 1.22 million. PPC Land identified this discrepancy; the form-submission and outbound-click rows produce similar results. Microsoft’s published methodology does not explain the gap or provide the human visit count. A filtered sample could explain it, but that remains unconfirmed. The figures therefore cannot establish a reliable sales advantage for agents. Nor can the tables reconstruct a customer journey: events can overlap across different sites and industries.
An event name is a measurement choice. Clarity’s Smart Events combine signals such as button clicks and page visits into named actions. Administrators can edit automatic events and create their own. A February technical guide explains that automatic detection uses button text and interaction patterns, with potential false positives. Its author, digital analyst Matteo Zambon, describes counts diverging from Google Analytics 4 in several ecommerce projects. That account provides no representative error rate, and it does not prove this study contains errors. It does explain why merchants should inspect what triggers a label before treating it as revenue.
MarketMaze Take: The useful distinction is between finding friction and counting sales. Clarity’s comparison can direct attention to visits worth inspecting. A higher Checkout-event rate does not justify shifting a sales forecast, particularly while the denominator and the human contribution remain unresolved. Better visibility earns an investigation, not an automatic budget increase.
Measure the finish separately. The February guide describes explicitly coded events that fire after server confirmation or a verified success state. Merchants anywhere can apply that distinction: use interaction labels to diagnose a journey, then a confirmed completion signal to count its commercial result. This is a measurement principle, not evidence of a worldwide conversion uplift.
Do not turn missing proof into failure. Human visits show Order success at 1.79%; the study omits the exact agent rate. Some agent-assisted visits may genuinely end in orders. The publication cannot establish how many were paid, who completed them, or whether agents caused a better outcome. Those are the numbers a business case still needs.
