Microsoft Clarity on October 29, 2026 published an analysis of more than 5.2 million website sessions involving AI agents, recorded across 230,511 Clarity projects between August 13 and September 9, 2026, and found that checkout was the action most often recorded in those sessions - in 25.21% of them, against 14.65% of human-only sessions on the same group of sites.

In Short

Microsoft looked at website visits where an AI assistant was doing the browsing for someone and counted which steps, such as checkout or filling in a form, showed up during those visits. Checkout showed up far more often than in visits by people alone, which matters to anyone who sells online or reports sales figures, because these AI-driven visits land in the same analytics as everyone else. Microsoft says it cannot tell which clicks the AI made and which the person made, and its percentages appear to rest on fewer than a quarter of the 5.2 million visits in its headline, so the numbers describe activity on sites, not proof that AI agents are buying things.

What Clarity measured

The study appeared on the Clarity blog under the byline "Clarity Staff," in a post that also announced a new Agent Activity Dashboard inside Clarity, which Microsoft describes as a free web analytics platform. The post sets out its premise early: "Most analytics stacks were built to understand human visitors, not agents acting on their behalf."

According to Microsoft, the dataset covers more than 5.2 million distinct "agentic-classified" sessions on 230,511 Clarity projects over four weeks, from August 13 to September 9, 2026, measured in UTC. Sessions were classified by detecting agentic browser providers across three operator ecosystems: OpenAI, Google and Anthropic. The study names no individual product, and it does not describe the detection method - whether it relies on declared user agents, cryptographic signatures, behavioural signals or something else.

That matters because a session in this data is not necessarily a session run entirely by software. "An agentic-classified session may also include a human handoff," according to the study. "Because Clarity cannot currently separate the agent and human portions into different sessions, the findings describe activity recorded during sessions involving agents, not necessarily actions completed autonomously by an agent." Handoffs are a routine part of how a computer-use agent works in practice: many pause and return control to the person when a login, a payment card or a confirmation step appears. Every action count in the study therefore blends what the agent did with what its user did.

Growth was the other headline. "In our own data, we noticed that agentic-session volume increased by 88% in just three weeks," according to Microsoft. The methodology describes the portfolio-wide growth figures as "observed weekly counts across all analyzed projects," but the study does not say which three weeks it compared, what the starting volume was, or whether the number of projects receiving agentic traffic changed over the period. An 88% rise from a small base reads differently from an 88% rise from a large one, and the study gives no way to tell which applies.

Smart Events, and what they are not

The behavioural comparison is built on Smart Events, a Clarity feature that the study describes as capturing "business-critical actions such as checkouts, form submissions, appointment bookings, searches, downloads, and logins." Microsoft compared which Smart Events were recorded, and at what rate, in agentic and human sessions during the same period.

Smart Events are not a fixed taxonomy applied identically on every site. According to Clarity's documentation, last updated in December 2025, the product automatically detects nine event types - purchase, add to cart, begin checkout, contact us, submit form, request quote, sign up, login and download - while project admins can hide or customise those automatic events and create up to 20 custom ones. A February 2026 guide on the Clarity blog describes automatic events as inferred from button text and interaction patterns, and lists lower precision and potential false positives among their limitations.

Several event names in the October study - Checkout, Order success, Outbound click, Book appointment, Check availability, Show more, Search and Find location - do not appear in that list of nine. The documentation may lag the product, but the study itself does not define any of them either. Nor does it say whether "Checkout" marks a click on a checkout button, a visit to a checkout page or a completed transaction. The human results suggest it is not the last of these: human sessions record "Checkout" at 14.65% but "Order success" at only 1.79%, so the two labels measure different things.

Checkout first for agents, outbound clicks first for people

The core of the study is two ranked lists of the ten most frequent Smart Events. For each event, Microsoft divided the number of sessions in which it was recorded by the total number of sessions in that group.

Action Agentic sessions (count) Agentic sessions (%) Human sessions (%)
Checkout 306,621 25.21 14.65
Submit form 241,391 19.85 18.56
Outbound click 173,655 14.28 26.99
Book appointment 102,401 8.42 not in top ten
Search 87,096 7.16 4.79
Login 83,096 6.83 8.84
Show more 52,456 4.31 6.37
Check availability 29,274 2.41 not in top ten
Add to cart 28,289 2.33 3.31
Download 20,593 1.69 2.97
Order success not in top ten not in top ten 1.79
Begin checkout not in top ten not in top ten 1.62

Outside the top ten, the study adds that 10,521 agentic sessions recorded Find location. It publishes no session counts for the human group, only percentages.

The two profiles overlap more than they differ. Eight of the ten actions appear on both lists. What changes is the weighting. Checkout sits at 25.21% of agentic sessions against 14.65% of human sessions, a ratio of roughly 1.7 to 1 by PPC Land's calculation. Search is about 1.5 times as frequent in agentic sessions, at 7.16% against 4.79%. Form submissions are close, at 19.85% and 18.56%.

The direction reverses for outbound clicks, which lead the human list at 26.99% and reach only 14.28% of agentic sessions. The human figure is "nearly twice" the agentic one, according to Microsoft. Humans also record more logins (8.84% against 6.83%), more "show more" interactions (6.37% against 4.31%), more add-to-cart events (3.31% against 2.33%) and more downloads (2.97% against 1.69%).

Microsoft's own summary is careful. "The clearest difference, then, isn't that agentic sessions perform a completely different set of actions," according to the study. "It's how heavily those actions are weighted."

Bookings in, order confirmations out

Two task-style events made the agentic top ten without appearing on the human list at all. Book appointment was recorded in 8.42% of agentic sessions and Check availability in 2.41%. Because the human list bottoms out at 1.62%, both must sit below that level for human sessions, which puts appointment bookings at more than five times the human rate in agentic sessions, by PPC Land's calculation. For travel, hospitality, healthcare and local services businesses, which run on bookings and availability checks, that is the most concrete signal in the data.

The opposite gap is just as telling. Order success, at 1.79%, and Begin checkout, at 1.62%, appear on the human list but not the agentic one. The agentic list ends at 1.69%, so both events were recorded in fewer than 1.69% of agentic sessions. The study gives no exact figure for either, which leaves open whether order confirmations are rarer in agentic sessions than in human ones or merely slightly below the cut-off.

Then there is the sequence problem. Read as a conversion funnel, the lists make little sense. In agentic sessions, Checkout (25.21%) outnumbers Add to cart (2.33%) by more than ten to one; in human sessions the ratio is about 4.4 to 1. Microsoft warns against that reading: the events "should not be interpreted as steps in a single conversion funnel or proof that an agent independently completed a task," according to the study. Several explanations would fit the numbers - agents reaching checkout pages directly from a link or a pre-built cart, a person taking over at payment, or site-by-site differences in how each event is defined - and the published data cannot separate them. Overlapping session counts cannot be turned back into a path, and Microsoft does not attempt to build one.

A 5.2 million headline, a 1.2 million base

The arithmetic of the tables raises a larger question. The footnote under the agentic list defines each percentage as "the number of agentic sessions that had a specific action divided by the total number of agentic sessions." Dividing each published count by its published percentage, however, returns the same denominator on every row: about 1.22 million sessions. Checkout gives 1,216,267; Submit form, 1,216,076; Outbound click, 1,216,071; Book appointment, 1,216,164. That is PPC Land's calculation, and it is consistent to within rounding across all ten rows.

About 1.22 million is roughly 23% of 5.2 million. Had the percentages been calculated against the full 5.2 million sessions in the study's title and methodology, Checkout would have been recorded in about 5.9% of agentic sessions, not 25.21%.

A second calculation points the same way. The ten agentic counts add up to 1,124,872. Because a single session can record several events, the number of distinct sessions that recorded at least one of the ten is lower still. Either way, at least 78% of the 5.2 million agentic sessions recorded none of the ten most common actions, again by PPC Land's calculation.

Why would the base be so much smaller? The study does not say. One possibility is that the percentages use only sessions on projects with Smart Events in place, or only sessions that recorded at least one Smart Event. If so, the human percentages would need to be built the same way for the comparison to hold, and Microsoft does not publish the number of human sessions or how that base was drawn. Without those figures, it is not possible to check whether the 25.21% and the 14.65% sit on comparable foundations.

There are smaller inconsistencies as well. The tables describe shares of sessions, while the accompanying text describes Checkout as "accounting for 25.21% of agentic actions" and Search as a larger "share of agentic activity" - a different measure, since the percentages are not shares of a total and do not sum to 100. The text says checkout appeared "in more than 306,621 agentic sessions," an exact figure presented as a floor. And a section on what site owners should watch states that "the findings point to two areas worth watching," then describes only one: task-oriented actions such as checkouts and appointment bookings.

Then there is what the study leaves out. It provides no breakdown by operator ecosystem, so OpenAI, Google and Anthropic traffic cannot be told apart. There is no split by industry, country or device. The study says human sessions were drawn "across the same cohort of Clarity projects that received agentic traffic," but not whether the comparison was weighted by site. If agentic sessions cluster on retail and travel sites while human sessions spread across every type of project in the cohort, part of the checkout and booking gap could reflect where agents go rather than what they do once there.

Which agents the count covers, and which it misses

The window closed on September 9, 2026. Meta released Muse in the US on September 8, and Meta is not among the three ecosystems the study names. That is not a minor omission when set against other vendors' data from the weeks that followed. HUMAN Security found Muse averaging 72% of daily AI agent requests on its network, with 84% of those requests aimed at product and search pages. DataDome put Muse at 33% of agentic browser requests in September despite measuring it for only nine days, with ChatGPT's cloud browser at 47% of the category and Perplexity's Comet falling to 0.8 million requests. Perplexity is not named in the Clarity study either. Nor is Microsoft's own Copilot, which gained a checkout inside Copilot.com in January 2026.

The three ecosystems that are covered have each shipped browser agents. Anthropic began piloting Claude for Chrome with 1,000 Max plan users in August 2025. Google introduced Chrome auto browse for AI Pro and Ultra subscribers in the United States on July 14, 2026, describing it as able to book travel and fill in forms. OpenAI discontinued its Atlas browser on July 9, 2026, a month before the window opened, and folded it into a new ChatGPT Work agent. OpenAI's update letting ChatGPT control Chrome tabs to fill forms and send messages arrived in mid-October, after the window closed. Which of these products, or which versions, the 5.2 million sessions represent, the study does not say.

Detection is the harder problem underneath. Arkose Labs found that only 2.3% of legitimate AI agent sessions on its network presented a verified signature, although ChatGPT agent signed 78.4% of its own sessions. The signatures in question use Web Bot Auth, which lets an agent prove its identity cryptographically; agents that do not use it can only be identified by inference. Any analytics product that classifies agentic sessions will count the agents that identify themselves, or behave distinctively, more reliably than those that do not. Microsoft does not say how much agentic traffic its classification might miss.

From bot logs to agent sessions

The study extends a run of AI-related measurement features that Clarity has added over the past year, most of them free. Microsoft Clarity launched Bot Activity tracking on January 21, 2026, built on CDN and server integrations whose connection might carry additional charges. In May it made its Citations dashboard generally available, tracking how often AI answers cite a site. June brought robots.txt violation flags inside Bot Analytics, and on August 13 a card comparing how much each AI operator scrapes with how many visits it refers, combining CDN logs with sessions recorded by Clarity's JavaScript tag.

The October study sits on the other side of that divide. Crawler requests show up in server logs; agentic sessions show up because an agent loads pages in a browser and Clarity's on-page tracking records what happens. Bot Activity counts how often automated systems fetch content. The new dashboard is meant to show what agents, and the people working alongside them, do once they are on a page.

Microsoft says the Agent Activity Dashboard is "now generally available for projects with agentic traffic" and helps businesses "identify agentic traffic and understand how AI agents interact with their websites." For ecommerce and travel and hospitality businesses, it adds, "simulations can also help uncover opportunities to improve experiences for an increasingly agent-driven web." The study does not explain what the simulations involve. The name is close to that of Clarity's existing Agent Dashboard, which reports on conversations with a retailer's own on-site brand agent rather than on outside agents arriving at the site.

The passage introducing the dashboard ends with a line that frames the study as much as the product: "Understanding agentic traffic is the first step toward understanding its business impact."

Why the numbers matter to advertisers

For marketers, the immediate issue is contamination of the numbers they already rely on. Checkout events, form submissions and appointment bookings are the actions many advertisers count as conversions, import into ad platforms and use to train automated bidding. If a meaningful share of those events now occurs in sessions where software did some of the clicking, conversion rates, cost-per-lead figures and attribution reports mix two very different kinds of visitor. Clarity's own caveat - that it cannot separate the agent from the human within a session - applies just as much to tag-based measurement systems that do not look for agents at all.

Form submissions deserve particular attention. At 19.85% of agentic sessions and 18.56% of human ones, they are the action on which the two groups look most alike. For lead-generation advertisers who pay per lead and score leads after the fact, near-identical rates mean that frequency alone gives little sign of whether a form was filled by a person or by an agent acting for one.

None of this tells marketers whether agentic sessions are worth more or less than human ones. Earlier Clarity research found that AI referrals converted to sign-ups at 1.66%, against 0.15% for search, but that measured people clicking through from AI answers - AI referral traffic - not agents browsing on their behalf. Walmart's experience points the other way: conversion for products sold directly inside ChatGPT was three times lower than for click-out links. Retailers are now writing their own rules, as Walmart did on October 21 with Agent Profiles that let AI agents buy within customer-set spending limits, while Amazon cut Muse off from buying on Amazon.com in September.

What the Clarity study does establish is narrower but still useful. Across a large sample of sites, sessions involving OpenAI, Google and Anthropic agents recorded checkout and booking events at higher rates than human sessions on the same properties, and outbound clicks at lower rates. Whether those sessions ended in orders, how many of them a person finished, and how the percentages relate to the 5.2 million sessions in the headline are questions the published data leaves unanswered.

Timeline

Summary

Who: Microsoft Clarity, Microsoft's free web analytics product, through a study attributed to Clarity Staff. The agentic sessions came from agentic browser providers in the OpenAI, Google and Anthropic ecosystems. The findings concern website owners, advertisers and analysts who rely on on-site events as conversion signals.

What: An analysis of more than 5.2 million agentic-classified sessions across 230,511 Clarity projects. Checkout was recorded in 25.21% of agentic sessions against 14.65% of human sessions; outbound clicks in 14.28% against 26.99%; form submissions in 19.85% against 18.56%. Appointment bookings (8.42%) and availability checks (2.41%) made the agentic top ten but not the human one, while order success and begin checkout made only the human list. Microsoft says sessions may include human handoffs it cannot separate. By PPC Land's calculation, the published percentages imply a base of about 1.22 million sessions rather than 5.2 million, a gap the study does not explain. Microsoft also made its Agent Activity Dashboard generally available.

When: Published October 29, 2026. The data covers August 13 to September 9, 2026 (UTC), a window that closed one day after Meta released Muse in the US.

Where: Websites running Microsoft Clarity. The study gives no geographic, industry or device breakdown.

Why: AI agents now browse, fill in forms and reach checkout pages on people's behalf, and their sessions land in the same analytics as human visits. Microsoft argues that measuring traffic volume is not enough and that site owners need to see what happens once an agent session begins. The study also serves as the launch material for Clarity's new dashboard.