← All insights

News analysisLens: United States2 min read

Did an AI agent manage to book on your site? Microsoft Clarity now measures agents as a separate audience

Clarity's Agentic Analytics shows which agents visit, what they try and whether simulated agents finish the task. If customers send agents, your website is an interface to test.

Listen to this article · 4 min

AI-generated narration of the full article.

An old storefront with window displays, in La Madre duotone, beside the words Agents at the door
Photo: Hines Studio (Wikimedia Commons, Public domain)

Your analytics can tell you how many people abandoned the checkout. Can it tell you how many AI agents tried to check out on someone’s behalf, and why they gave up?

Microsoft’s free analytics tool, Clarity, now tries to. Its new Agentic Analytics treats agents as an audience of their own.

What Clarity measures

There are two parts. An Agent Activity dashboard shows agentic traffic over time, which agents visited, and what they did: searches, form submissions, logins, downloads, bookings and checkouts. Simulation assessments send agents to attempt tasks on the site and report whether they completed them, how long it took, how many interactions were needed and where they struggled.

The limits are stated plainly. Agentic Analytics recognizes agents operated by Google, OpenAI and Anthropic only. Simulations start with commerce and travel scenarios.

Microsoft also published a baseline. Between August 13 and September 9, Clarity detected 5.23 million agent sessions across 230,511 sites, and now sees almost 2 million agent sessions a week. That averages about 23 agent sessions per site over four weeks: small for most sites, but no longer zero.

The same post introduces Brand Agents, a shopping assistant for eligible WooCommerce and Shopify merchants that can surface products in Copilot, Bing and MSN, and Ask Clarity, a question-answering tool over site behavior in closed beta for retailers. Both are aimed at merchants. The agent measurement is the part with wider consequences.

Why enterprises should care, not only retailers

Most companies think of agents as something they deploy. The other half is the agents their customers, suppliers and candidates deploy, which arrive at the company’s website, portal or booking flow and try to complete a task.

Those agents fail in ways humans do not. A form that relies on a hover tooltip, a date picker that only responds to mouse events, a captcha on every step, or a login that times out during a multi-step flow will stop an agent that a patient human would get through. The customer never sees the failure. They just hear from their agent that the booking did not work, and may try a competitor.

We described one side of this when Radisson made ChatGPT a sales channel: a company can choose to build for agents. Clarity measures the other side, where agents show up whether or not you built for them.

What to do with it

Look before you optimize. Turn on the agent dashboard and see which flows agents attempt. Insurance quotes, appointment booking, order tracking and returns are the likely candidates outside retail.

Separate agent sessions in your funnel metrics. An agent that abandons a checkout is not a human losing interest. Mixing the two hides both signals.

Decide which agent actions you accept. Some flows, such as account creation or payments, may need stronger verification when an agent is acting. That is a policy decision for security and legal, not something analytics should settle by default.

Test the critical paths. Simulations show whether an agent can complete a task. Run them on the two or three flows that bring revenue, and treat failures as accessibility bugs, because often that is what they are.

Have an AI use case stuck between prototype and production?

Tell us what you’re trying to ship. We’ll reply with honest next steps.

Discuss a use case