TL;DR
- Book appointment (8.42%) and Check availability (2.41%) appear in Clarity's agentic top 10 but not in the human top 10.
- Outbound clicks lead human sessions at 26.99% against 14.28% of agentic sessions; form submissions are close at 19.85% vs 18.56%.
- Microsoft says agentic sessions may include human handoffs it cannot separate, and that Smart Events are not proof an agent completed a task.
- PPC Land noted on 1 November 2026 that 306,621 checkout sessions at 25.21% imply about 1.22 million sessions behind the percentages, not 5.2 million.
Microsoft Clarity says a checkout event was recorded in 25.21% of sessions involving AI agents, against 14.65% of human-only sessions on the same sites. The figure comes from a Clarity blog study published on 29 October 2026 that analysed more than 5.2 million agentic-classified sessions across 230,511 Clarity projects between 13 August and 9 September 2026. In the same post, Microsoft said its Agent Activity Dashboard is now generally available for projects that receive agentic traffic.
What Microsoft Clarity published
The post, titled "What AI Agents Do on Websites: Insights From 5.2 Million Sessions" and credited to Clarity Staff, sets out to answer a narrow question: once a session involving an AI agent reaches a website, what does it do there? You can read it on the Microsoft Clarity blog.
Clarity identified the sessions by detecting agentic browser providers across three ecosystems: OpenAI, Google and Anthropic. It then compared which Smart Events were recorded in those sessions with the Smart Events recorded in human-only sessions from the same set of projects over the same period. Smart Events are Clarity's built-in tracking for business actions such as checkouts, form submissions, appointment bookings, searches, downloads and logins.
The post also says that in Clarity's own data, agentic-session volume rose 88% in three weeks. It does not name which three weeks. The methodology note adds that the traffic growth figures represent observed weekly counts across all analysed projects.
That growth figure needs care. A rise of 88% says nothing about size without a starting point, and the post gives no absolute weekly count for agentic sessions or any share of total traffic. It also does not say whether the growth came from more projects receiving agent traffic or from more agent sessions per project. Read it as a sign of direction, not as a forecast.
Top Smart Events: agentic sessions vs human sessions
Microsoft published two top-10 lists, one for sessions involving agents and one for human sessions. The table below lines up the events that appear in the agentic list with the human share where Microsoft gave one.
| Smart Event | Share of agentic sessions | Share of human sessions |
|---|---|---|
| Checkout | 25.21% | 14.65% |
| Submit form | 19.85% | 18.56% |
| Outbound click | 14.28% | 26.99% |
| Book appointment | 8.42% | Not in human top 10 |
| Search | 7.16% | 4.79% |
| Login | 6.83% | 8.84% |
| Show more | 4.31% | 6.37% |
| Check availability | 2.41% | Not in human top 10 |
| Add to cart | 2.33% | 3.31% |
| Download | 1.69% | 2.97% |
The human top 10 also includes two events that are missing from the agentic list: Order success at 1.79% and Begin checkout at 1.62%. Microsoft also gave raw counts for the agentic sessions: 306,621 recorded Checkout, 241,391 recorded Submit form, 102,401 recorded Book appointment, 29,274 recorded Check availability, 28,289 recorded Add to cart and 10,521 recorded Find location.
How Microsoft reads the numbers
Microsoft's own conclusion is that agentic and human sessions perform largely the same kinds of actions, with different weight. Form submissions are close (19.85% vs 18.56%). Checkout is much higher in agentic sessions. Outbound clicks are the top human action at 26.99%, nearly twice their 14.28% share among agentic sessions. Search also takes a bigger share of agentic activity, 7.16% against 4.79%.
The post says task-oriented actions such as checkouts and appointment bookings occur at a higher rate in agentic sessions, and that understanding the gap will matter as agentic commerce develops.
The caveats Microsoft attached
Clarity spelled out what the data cannot show, and those limits change how the headline numbers should be read.
- Human handoffs are mixed in. An agentic-classified session may include a human taking over partway. Clarity cannot currently split the agent part and the human part into separate sessions, so the findings describe activity recorded during sessions involving agents, not actions an agent necessarily completed by itself.
- Events are not a funnel. Microsoft says the Smart Events are independent, non-mutually-exclusive session counts across different websites and industries. They should not be read as steps in one conversion funnel or as proof that an agent completed a task.
- One session can count several times. Because the events are non-exclusive, a single session that searches, expands a product list and then reaches checkout would add to Search, Show more and Checkout at once. As an illustration only, that is why the percentages cannot be added up into a path or read as drop-off rates between steps.
- Checkout is not a completed order. Clarity's human list separates Checkout from Order success and Begin checkout. The agentic top 10 has no Order success entry, so the study does not show how many agentic sessions ended in a confirmed purchase.
There is also a wording slip worth noting. The table footnotes define each percentage as a share of sessions, but parts of the post's prose describe the same figures as a share of "actions". The session definition in the footnotes is the precise one.
Questions the study leaves open
PPC Land, the trade publication, pointed out an arithmetic gap in its 1 November coverage. If 306,621 sessions are 25.21% of the agentic total, the total behind the percentages is about 1.22 million sessions, not 5.2 million. Dividing the published numbers confirms the arithmetic (306,621 ÷ 0.2521 is roughly 1.216 million). Microsoft's post does not explain the difference.
The post also leaves out several things a site owner would want:
- the exact detection method for agentic browsers;
- the number of human sessions used for comparison;
- any split by agent provider, industry, country or device;
- precise definitions for each Smart Event;
- pricing or regional availability for the Agent Activity Dashboard. Clarity itself is described in the post as a free web analytics platform.
Because there is no country breakdown, nothing in the study says how much agent traffic Thai sites receive.
The Agent Activity Dashboard
Microsoft used the study to introduce the Agent Activity Dashboard in Clarity. The post says it is now generally available for projects with agentic traffic, and that it helps businesses identify agentic traffic and see how AI agents interact with their websites. For ecommerce and for travel and hospitality businesses, the post adds that simulations can help uncover ways to improve the experience for agents. It gives no further detail on how the simulations work.
The study looks at what happens during sessions where an agent is driving the browser, rather than at visitors who arrive from an AI answer.
What this means for Thai businesses
The points below are Relevant Audience analysis, not Microsoft guidance. They assume the patterns in a global sample are worth checking against your own site, without assuming they hold in Thailand.
Check whether you already have agent traffic
Clarity is free to install. Microsoft says the Agent Activity Dashboard is available for projects with agentic traffic, so if your project shows it, agentic sessions are reaching the site. The post does not say what a project without agent traffic sees, so the absence of the dashboard should not be read as proof of zero agents.
Look at the three flows that stand out
Checkout and Submit form are the two most common agentic actions, and Book appointment ranks fourth, behind Outbound click. For a Thai online shop, a clinic or a hotel, those map to the cart and payment page, the contact or quote form and the booking widget. Watching Clarity recordings of agentic sessions on those pages is the most direct way to see where a session stalls or hands back to a human.
Treat "Check availability" and "Book appointment" as signals for service sites
Two events in the agentic top 10, Book appointment (8.42%) and Check availability (2.41%), do not appear in the human top 10 at all. For clinics, salons, restaurants and hotels, booking pages that only work inside a chat app or through a phone call cannot be completed in a browser session. Whether that loses agent-driven bookings is not something the study measured, but it is a reasonable thing to check.
Do not rebuild reporting on these percentages
Given the handoff problem, the non-funnel counts and the unexplained 1.22 million versus 5.2 million gap, the numbers are useful as a direction, not as a benchmark for your own conversion rates.
For help reading agent and AI-assistant traffic alongside AI search visibility, see Relevant Audience's generative engine optimization service. If your tracking needs to be rebuilt first, the GA4 and analytics setup service covers events and conversions, and online stores can look at ecommerce marketing.
FAQ: Microsoft Clarity's AI agent session study
Do AI agents really complete more checkouts than humans?
The study does not show that. It shows a Checkout Smart Event was recorded in 25.21% of sessions involving agents against 14.65% of human sessions, and Microsoft says these counts are not proof that an agent completed a task by itself.
Which AI agents did Clarity measure?
Clarity detected agentic browser providers in the OpenAI, Google and Anthropic ecosystems. The post does not break results down by provider or name specific products.
Is the Agent Activity Dashboard available to all Clarity users?
Microsoft says it is generally available for projects with agentic traffic. The post does not mention pricing or regional limits; Clarity itself is described as free.
Why do some reports say the totals don't add up?
The published counts imply about 1.22 million sessions behind the percentages, not 5.2 million. PPC Land raised this, and Microsoft's post does not explain it.
Does the study include Thai websites?
The post does not say. It reports no country breakdown, so nothing in it can be read as a Thailand figure.
If you want to know what agents are doing on your own checkout or booking pages, Relevant Audience can help set up and read the data.
