Back to Explore
NewsVibeFix editorialUpdated Nov 20, 2026

Your Next Customer May Not Be Human: Microsoft Clarity Now Counts What AI Agents Do on Your Site

On Nov 4, 2026, Microsoft added three things to its free tool Clarity: a dashboard that watches what agents do on your site, a brand assistant that puts products into Copilot answers, and an analyst you can question. Checkout appears in 25.21% of agent sessions — but that number has three caveats.

A shopper holds a blue credit card while a laptop screen shows a checkout page (real photo)

On November 4, 2026, Microsoft added three things to Clarity, its free website analytics tool: a dashboard that shows what AI agents do on your site, a brand assistant that puts your products inside Copilot and Bing answers, and an analyst that answers questions about your data directly. Each one serves a different web, in Microsoft's framing: the same customers now arrive three ways — some browse your site themselves, some ask an AI for a recommendation first, and some send an agent to book or buy on their behalf. The real meaning of this update is that website analytics is, for the first time, treating that third kind of arrival — that third kind of "visitor" — as a group worth counting and understanding on its own.

It is worth reading carefully, not because Microsoft shipped three more features, but because of one number it published alongside them: in sessions involving an AI agent, the checkout action appeared in 25.21% of sessions, against 14.65% in human-only sessions. That number is easy to misread as "agents love buying things more than people do," but Microsoft's own research report immediately pours cold water on that reading: these sessions may include a human handoff, and the events are not steps in a funnel. The number is real; the interpretation needs care. This article follows that order: first what the three things are, then the report's numbers taken apart one by one, and finally the boundaries that must not be crossed — and what site owners should do first.

Three things, one for each web

The first is called Agentic Analytics, and its core is an Agent Activity dashboard. It answers three questions: how agent traffic to your site changes over time, which agents are visiting, and what they do once they arrive. Microsoft lists the actions as searches, form submissions, logins, downloads, appointment bookings, and checkout. The dashboard has a second ability, simulation assessments: you can send a test agent through a key task on your site — buying a product, booking a trip — and see whether it finished, how long it took, how many interactions it needed, and where it got stuck. Simulations start with commerce and travel scenarios, and Microsoft says it plans to expand them to more kinds of customer journeys over time.

The dashboard is not entirely new. Microsoft had already made it generally available on October 29, alongside the same research report, for projects that actually receive agent traffic. The November 4 announcement folds it under the Agentic Analytics name and adds the simulation piece. In other words, on October 29 Microsoft first produced data proving that agents are genuinely placing orders, and on November 4 it packaged "seeing" and "testing" into a formal feature. The order itself says something about Microsoft's judgment: prove the phenomenon first, then sell the tool.

The second is Brand Agents, which serves the "LLM web." It started as an AI shopping assistant on your own website, grounded in your live product catalog, shipping and return policies, and brand voice. What is new is distribution: eligible Shopify and WooCommerce merchants can now have their products appear across Microsoft Copilot, Bing, MSN, and more when shoppers ask AI for recommendations, with the shopper then completing the purchase on the merchant's own website. Merchants who already have Brand Agents installed need no additional catalog upload or configuration, and during setup they can choose to use Brand Agents only for discovery across AI experiences, without putting the assistant on their own site.

The same product is also expanding to local service businesses. When a potential customer arrives at your site and asks, "Can you help me fix a leaking pipe under my kitchen sink?", Brand Agents can identify the right service and offer to book it. It connects with the booking and scheduling tools local businesses already use — Square, Jobber, HubSpot, Calendly, Housecall Pro, and Salesforce — so leads arriving after hours can be captured and appointments scheduled directly, instead of waiting for a callback the next morning. Eligible WordPress businesses can get started today.

The third is Ask Clarity, which serves the "human web." You ask Clarity a question about your business, and it analyzes your own data much like an analyst would: exploring possible explanations, testing them against the evidence, and identifying what is really happening. It then recommends what to do next and links its findings to the relevant session recordings, so you can see the evidence for yourself. Microsoft's example: European shoppers are not converting as well as other regions, and you think price is the problem; Clarity tests that assumption against real visitor behavior and finds something you were not looking for, like untranslated pages or a delivery option that only works in one market. The capability is currently in closed beta for retailers — join the waitlist to try it — and Microsoft plans to expand it to more use cases and data sources, including AI visibility and advertising data.

The research numbers, taken apart one by one

Before judging the three features, it is worth reading the research report behind them. The study Microsoft published on October 29 analyzed 5.23 million sessions involving AI agents across 230,511 Clarity projects, covering August 13 to September 9, 2026. Sessions were identified using detected agentic browser providers across the OpenAI, Google, and Anthropic ecosystems. The same study says weekly agent-session volume rose 88% in just three weeks, to almost 2 million. Put together, the message is clear: agents visiting websites are no longer a sporadic curiosity, but a flow of traffic that nearly doubled within weeks.

Now the actions. The most frequently recorded action in agent sessions was checkout, appearing in 306,621 sessions, or 25.21%; second was form submission, 241,391 sessions, or 19.85%; then outbound click at 14.28%, book appointment at 8.42%, search at 7.16%, login at 6.83%, show more at 4.31%, check availability at 2.41%, add to cart at 2.33%, and download at 1.69%. Compared with human-only sessions, the difference jumps out: for humans, the top action was outbound click at 26.99%, nearly twice the 14.28% share in agent sessions, while checkout accounted for only 14.65% and ranked third. Form submissions were close on both sides, 19.85% against 18.56%; search ran higher for agents, 7.16% against 4.79%; add to cart ran higher for humans, 3.31% against 2.33%.

The most interesting thing about this comparison is not which side is higher, but what it says about the "shape" of an agent session versus a human one. In human sessions, outbound clicks take the largest share — people click out to compare prices, read reviews, and jump elsewhere; in agent sessions, checkout and form submission take the largest share — the agent arrives with a specific task, clicks out less, and heads straight for the steps that get the job done. Microsoft's own conclusion is measured: the two groups perform largely the same actions, but those actions carry very different weights. Agent sessions are not doing a completely different set of things; they are doing the same things with different emphasis. That is a far more accurate — and more restrained — reading than "agents love buying things."

Agent sessions vs human sessions by action: checkout 25.21% vs 14.65%, form submission 19.85% vs 18.56%, outbound click 14.28% vs 26.99%, search 7.16% vs 4.79%, add to cart 2.33% vs 3.31%

Two more numbers deserve to be pulled out on their own. One is book appointment, at 8.42% of agent sessions and ranked fourth. Agents are not only buying things; they are also booking time on people's behalf — reserving tables, booking services, arranging schedules. The other is "show more" at only 4.31%, against 6.37% for humans: agents do not browse a page slowly the way people do. They either head straight for the goal or get stuck and leave. Together, the two numbers draw a concrete portrait of the agent: goal-directed, direct in its steps, and not fond of wandering.

Three boundaries not to ignore

The first boundary is the list of names. Agentic Analytics currently recognizes agents from only three operators: Google, OpenAI, and Anthropic. Microsoft's own Copilot is not on the list. That is a detail worth pausing on: Microsoft's tool for measuring "whose agents visited your site" does not count its own product. There are plausible reasons — Copilot's traffic may not travel through the same browser channel, or Microsoft may have chosen to get third-party measurement right first — but the announcement does not explain it. For readers, the practical meaning is simple: the dashboard's numbers are not "all agent traffic," but "agent traffic from three operators," and that qualifier belongs in every reading.

The second boundary is methodological, and it is the most important sentence in the whole report: a session classified as "involving an agent" may also include a human handoff, and Clarity cannot currently separate the agent's part from the human's. So every percentage in the report describes "what actions were recorded during sessions involving agents," not "what actions an agent completed on its own." Microsoft repeats this in the methodology section and adds that these events are independent, non-mutually-exclusive session counts: they should not be read as steps in a single conversion funnel, nor as proof that an agent independently completed a task. Checkout at 25.21% means a checkout event was recorded in a quarter of agent-involved sessions — not that a quarter of agents checked out by themselves, and certainly not that the agent checkout success rate is 25.21%.

That qualifier also undermines the tempting reading that "agents are better shoppers than people." The gap between 25.21% and 14.65% is real, but its cause is unclear: agents may genuinely be better at completing a checkout flow, or people may be using agents as "pick it out for me, I will check out myself" helpers, with the human completing the final step. Both explanations fit the data, and Microsoft does not pick one. For site owners, the implication is that the dashboard can tell you "agents came, and where they stalled," but it cannot yet answer "who actually completed that step."

The third boundary concerns Ask Clarity's eye-catching number. Microsoft says that in a survey of early-access users, 88% reported discovering something new about their business or website. That sounds persuasive until you read the footnote: the survey ran in July and August 2026 with only 16 participants — fourteen of sixteen. Sixteen people is not fake data, but it supports only a limited conclusion: some early users found it useful, not that most websites will surface new discoveries. More credible, and more useful, is the named example Microsoft quotes in the post: Kelly Wellman, eCommerce Manager at Pela Case, says the feature helped her find that 80.2% of pageviews on a new product personalization landing page ended with no onward click. That is a specific story about a real person finding a real problem — worth more than the 88%.

What this means for people who run websites

Consider a hypothetical example to see how simulation assessments work — the following is an imagined small shop, not a measured result. Suppose you run an online store selling plumbing fittings, and Brand Agents has put your products into Copilot's answers, so a customer sends an agent to buy one specific elbow fitting. A simulation lets you send a test agent through the same journey first: can it find that model? Can it recognize the promotional price? At the checkout page, does it get stuck because a required field's hint is hidden inside a collapsed panel? If it stalls, you fix that panel instead of guessing "why don't agents buy from me." The value of the feature is not that it is clever, but that it turns "the agent experience" into something testable and localizable to a specific step — much like usability testing a checkout flow with real people used to be.

Zoom out a little further. In the same week, United Airlines announced that it would not allow AI agents to book flights or manage reservations on passengers' behalf, citing safety and federal regulations (covered on this site on November 11). On one side, an airline is shutting agents out; on the other, Microsoft is handing site owners tools to welcome them. The two moves are not contradictory — they are two sides of the same fact: agents have grown too large to pretend away, and the remaining choice is only whether to block or to receive. Those who block must answer "how, and for how long"; those who receive must answer "what happens after they arrive." Clarity's update is a measuring tape for the second group.

One more angle is worth thinking about: the measuring tape is free. Clarity was already a free product, and these three capabilities are available globally to all users — Ask Clarity excepted, still in closed beta. Free means any small site can start using it immediately, and measuring agent traffic is no longer a privilege of large companies. But it also means your site's behavioral data flows into Microsoft's systems. Microsoft does not discuss data use in this announcement; Clarity's privacy terms are a separate matter. For small teams, the trade is easy to state: free visibility in exchange for accepting Microsoft's data terms. Anyone who finds that unacceptable can opt out — and will simply not see this slice of traffic.

Three things to do first

First, check Clarity to see whether your site receives any agent traffic at all. If it does, look at what the agents did and where they stalled before deciding what to fix. If it does not, do not worry yet: run a simulation on your most important task line — checkout for e-commerce, booking for travel, appointment scheduling for local services. Simulations currently cover only commerce and travel scenarios; if your site is a different kind, the tool cannot help you yet, and you will have to wait for it to expand.

Second, check whether your product information can be read by machines. Brand Agents works by answering shoppers' questions from your catalog. If your product titles are vague, specifications are buried inside images, and prices and stock are out of sync, the AI cannot answer on your behalf, and the exposure will go to someone else. This resembles SEO, but the audience has changed: it used to be about being understood by people; now it is about being readable by models. Eligible Shopify and WooCommerce merchants can do this as a side effect of turning on distribution, since it requires no fresh catalog upload.

Third, do not treat the dashboard's numbers as a report card. A high checkout share does not mean your site is agent-friendly; a low one does not mean it is broken. What is worth watching is the specific stall point a simulation finds: where the agent stopped, and why. Fixing one real stall point is worth far more than watching percentages move. That is also where Microsoft's tools offer their greatest value: for the first time, "can an agent get something done on my site" has become a question that can be measured, located, and fixed — not just a vague feeling.

Back to the opening point. For twenty years, the logic of website analytics has been built around how people browse: clicks, dwell time, bounces, conversions. Clarity's update does not overturn that logic; it simply adds a column beside it: after agents arrive, what did they do? That column is still rough — it recognizes only three operators, cannot separate human handoffs, and simulates only two scenarios — but it is there now. For anyone whose business runs through a website, that is something to take seriously: your next customer may not be a person, but a program acting on someone's behalf, and whether it can get through your checkout page is, starting this month, something you can finally see.

Views 7Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates
A developer types at a dual-monitor workstation; the left screen shows a dark code editor, the right a testing dashboard with green checkmarks and bar charts
News
Did Your Agent Get Better or Worse? AWS Open-Sources UAEF to Measure Agents from Every Framework with One Ruler

On 11 November 2026, AWS open-sourced UAEF: adapters turn traces from LangGraph, LangChain, Strands and Bedrock AgentCore into one format, then 36 metrics across 7 dimensions score them, flagging regressions past 5%. The judge model is hardened too — scores must cite verbatim evidence, and never decide safety.

Developer WorkflowProduct LaunchTesting & Quality