---
title: "AI agent analytics: what agents do on your site | PostCliq"
description: "Microsoft Clarity studied 5.2 million AI agent sessions: agents reach checkout and booking more often than people. How to measure agent visits and fix gaps."
url: "https://postcliq.com/blog/ai-agent-analytics"
---

# AI agent analytics: what 5.2 million agent sessions did on websites

Microsoft Clarity's study of 5.2 million AI agent sessions shows agents head for search, booking and checkout far more than people do. Here is what the data says, what it does not, and how to measure agent visits on your own site.

By [Nirmal Gyanwali](https://postcliq.com/blog/author/nirmal-gyanwali) Updated 13 November 2026 8 min read

On this page

1. The short answer
2. What is AI agent analytics?
3. What did Microsoft Clarity find about AI agents?
4. Does a checkout in an agent session mean the agent bought something?
5. What does Clarity's Agent Activity dashboard show?
6. Why does agent behaviour change what you measure?
7. Where do agents get stuck?
8. Where is AI agent analytics heading?
9. What to do next
10. Frequently asked questions

## The short answer

Key takeaways

- Microsoft Clarity studied more than 5.2 million agent sessions across 230,511 sites from 13 August to 9 September 2026. Checkout appeared in 25.21% of agent sessions and 14.65% of human sessions on the same sites.
- Agent sessions also searched more and booked appointments often enough to enter the top ten actions. They clicked away to other sites about half as often as people did.
- Clarity's Agent Activity dashboard and task simulations launched on 4 November 2026. Supported agents come from the OpenAI, Google and Anthropic ecosystems.
- An agent session can include a human handoff, so these numbers show activity during agent visits, not purchases agents completed alone.
- Our view: agents arrive with a job to do. The site that lets them finish it wins the enquiry. Agent visits now need the same post-click measurement as people.

We have argued that what an agent can find, read and act on is becoming the conversion layer. Our guide to [AI agent detection](https://postcliq.com/blog/ai-agent-detection) covers who these visitors are and why some hide. This piece covers the next question: what they do when they arrive, and how to measure it.

## What is AI agent analytics?

AI agent analytics is measuring the visits that AI agents make to your website for a person: which agents arrive, which pages they reach and which actions they take, such as searching, submitting a form, booking or checking out. It sits beside human analytics, not inside it, because an agent's visit has a different shape.

One naming note. In the data industry, "agentic analytics" usually means AI tools that analyse your business data for you. That is a different subject. Here we mean analytics about agents, the visitors, which Microsoft calls Agentic Analytics in Clarity.

The difference from ordinary bot reporting matters. Bot reports count crawlers that copy pages for search indexes or model training. Agent analytics watches software acting for one person, often in a real browser, trying to complete a task. A crawler reads your returns policy. An agent tries to lodge the return.

## What did Microsoft Clarity find about AI agents?

Clarity found that sessions involving an AI agent are more task-focused than human sessions. Checkout appeared in 25.21% of agent sessions against 14.65% of human sessions, searches in 7.16% against 4.79%, and form submissions in 19.85% against 18.56%. Outbound clicks were far lower, 14.28% against 26.99%.

The study covered more than 5.2 million sessions classified as agentic, across 230,511 Clarity projects, from 13 August to 9 September 2026. Clarity identified them by detected agentic browser providers in the OpenAI, Google and Anthropic ecosystems, and compared them with human-only sessions from the same projects ( [Microsoft Clarity](https://clarity.microsoft.com/blog/agentic-analytics-what-agents-do-on-websites/) ).

Microsoft Clarity agent session study

**5.2M** agent sessions analysed across 230,511 sites, 13 August to 9 September 2026 [Microsoft Clarity, Oct 2026](https://clarity.microsoft.com/blog/agentic-analytics-what-agents-do-on-websites/)

**25.21%** of agent sessions included a checkout event, against 14.65% of human sessions [Microsoft Clarity, Oct 2026](https://clarity.microsoft.com/blog/agentic-analytics-what-agents-do-on-websites/)

**8.42%** of agent sessions booked an appointment, an action outside the human top ten [Microsoft Clarity, Oct 2026](https://clarity.microsoft.com/blog/agentic-analytics-what-agents-do-on-websites/)

**Nearly 2M** weekly agent sessions in Clarity's data, nearly doubling in three weeks [Microsoft Clarity, Nov 2026](https://clarity.microsoft.com/blog/clarity-connects-agents-llms-people/)

| Action | Agent sessions | 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 the top ten |
| Search | 7.16% | 4.79% |
| Login | 6.83% | 8.84% |
| Add to cart | 2.33% | 3.31% |
| Download | 1.69% | 2.97% |

Source: [Microsoft Clarity, What AI agents do on websites, 29 October 2026](https://clarity.microsoft.com/blog/agentic-analytics-what-agents-do-on-websites/)

Read the table as a behaviour profile, not a funnel. Clarity says these events are not steps in one journey, and they do not prove an agent finished anything. What the profile shows is intent. People browse, compare and leave for another tab. Agents go to the search box, the booking form and the checkout, because a person sent them there with a job. Some of that checkout work is moving off the page altogether: the [Universal Commerce Protocol (UCP)](https://postcliq.com/blog/universal-commerce-protocol) lets an assistant build a cart and pay without opening your website.

## Does a checkout in an agent session mean the agent bought something?

No. Clarity is explicit: "An agentic-classified session may also include a human handoff." Its findings describe "activity recorded during sessions involving agents, not necessarily actions completed autonomously by an agent" ( [Microsoft Clarity](https://clarity.microsoft.com/blog/agentic-analytics-what-agents-do-on-websites/) ). A checkout event is also not an order. The study lists order success as a separate event, and it does not appear in the agent top ten.

The handoff is often by design. Google says auto browse in Chrome will "explicitly ask for your confirmation" before sensitive actions such as a purchase ( [Google](https://blog.google/products-and-platforms/products/chrome/gemini-3-auto-browse/) ). So a typical agent visit may look like this: the agent searches, filters, fills the cart and the delivery details, then stops at payment and waits for the person to approve. Whether that person approves on your site depends on how much work the agent got done before the pause.

That is the useful reading for a business. The agent does the tedious part of the conversion. The person does the final click. If the agent stalls on step two, there is no final click to make.

## What does Clarity's Agent Activity dashboard show?

The Agent Activity dashboard shows how agent traffic changes over time, which agents visit, and what they do once there: searches, form submissions, logins, downloads, appointment bookings and checkout. Clarity says it is generally available for projects that receive agent traffic, and Clarity itself is free ( [Microsoft Clarity](https://clarity.microsoft.com/blog/clarity-connects-agents-llms-people/) ).

Microsoft announced it on 4 November 2026 alongside two other features ( [Microsoft Advertising](https://about.ads.microsoft.com/en/blog/post/november-2026/whats-new-in-microsoft-clarity-agent-insights-faster-answers-and-ai-discovery) ). Simulations test whether an agent can complete a key task on your site and report whether it finished, how long it took, how many interactions it needed and where it struggled. They start with commerce and travel journeys. Brand Agents can make eligible Shopify and WooCommerce products discoverable in Microsoft Copilot and Bing, with the purchase still completed on the merchant's site.

Two limits are worth knowing before you rely on it. Supported agents come from Google, OpenAI and Anthropic, so agents from other companies are not counted. And detection depends on the agent being recognisable. Our [AI agent detection](https://postcliq.com/blog/ai-agent-detection) guide covers assistants that now browse looking like people, which no dashboard will label.

Clarity is not alone. Cloudflare added a Personal Assistants category to its bot analytics on 4 November, with per-agent controls to allow or block each supported agent. Its own framing is blunt: "Blocking agents altogether can feel like the safest choice, but it can mean losing customers" ( [Cloudflare](https://blog.cloudflare.com/personal-agents-visibility-control/) ). When the company that sells bot protection says that, the default has shifted.

## Why does agent behaviour change what you measure?

Because the metrics built for people mislead you about agents. Bounce rate, time on page and scroll depth describe human attention. An agent that lands, books and leaves in twelve seconds looks like a bounce. An agent stuck in a loop on a broken date picker looks like engagement.

The measures that matter for agents are the ones that matter most after the click anyway. Did the visit reach the action it came for? Where did it stop? What was on the page at that moment? We cover the human version in our piece on [dead clicks and rage clicks](https://postcliq.com/blog/dead-clicks-and-rage-clicks) . For agents, the same friction shows up as repeated interactions, abandoned forms and sessions that end one step short.

There is a revenue reason to care now, before the numbers are large. Agent sessions concentrate on the highest-value pages on a site: checkout, booking and enquiry. A small share of total traffic can be a large share of the visits that were about to convert. Treating them as noise means losing exactly the visitors who arrived ready to act.

## Where do agents get stuck?

Clarity has not yet published failure rates by cause, so we cannot rank them with data. What its simulations measure, completion, time, interactions and friction points, points to the usual suspects. These are the patterns we see stop software on real sites:

- **Content that only appears after scripts run.** Some agents read the server's HTML rather than the finished page. If prices, availability or the form itself load later, the agent may see an empty page.
- **Custom controls without names.** Drop-downs, date pickers and quantity steppers built from styled boxes rather than real form controls. Agents rely on the accessibility tree, and an unnamed control is invisible there. Our guide to [Chrome's Gemini autofill](https://postcliq.com/blog/chrome-gemini-autofill) covers how to label forms so software fills them correctly.
- **Challenges in front of the task.** A CAPTCHA or a firewall rule that blocks automated browsers stops the agent before it starts. That decision deserves care, not a default. Our guide to [whether to block AI agents](https://postcliq.com/blog/block-ai-agents) sets out the trade-off.
- **Forced accounts.** A checkout or booking that demands a new account mid-flow gives the agent a step it often should not take without its person.
- **Vague confirmation.** A form that submits with no clear success message leaves the agent, and its person, unsure whether anything happened.

On WordPress and WooCommerce sites, most of these live in a form plugin, a checkout extension or a security rule rather than in the content. That is the kind of fix a WordPress specialist such as WP Creative makes in the same pass as the human version of the problem, because the same broken field usually loses both.

## Where is AI agent analytics heading?

Our view: within a few years, agent sessions will be a standard row in every analytics report, next to organic, paid and direct. The question will move from "how many agents came?" to "what share of agent tasks did we complete, and what did those tasks become?" These are forecasts, not facts.

- **Completion becomes the agent metric.** Simulations are an early form of it. We expect agent task completion to sit next to conversion rate in reports, and to be compared between competitors.
- **Agents will choose the sites that work.** An assistant that fails on one site and succeeds on another learns where to send the next task. Sites that block or confuse agents will not see a drop in traffic. They will see the tasks go elsewhere, with nothing in their reports to say so.
- **Identity and analytics will merge.** Clarity identifies agents by provider. Cloudflare identifies them through its verified bot directory. Agents that sign their requests will be measured and trusted. Agents that hide will be measured as people, and blamed when their sessions look strange.

The direction is the same one we have been describing since October. The website is becoming an interface that software operates on a person's behalf. Measurement has started to catch up. The sites that look at the data early will fix the stuck steps while competitors still call them bounces.

## What to do next

1. **Install Clarity, or check it is still firing.** Run [is my tracking installed](https://postcliq.com/free-tools/is-my-tracking-installed) on your key pages. Agent Activity only appears for projects that see agent traffic.
2. **Open Agent Activity and write down three numbers:** weekly agent sessions, the share that reached your main action, and the agents sending them.
3. **Run a simulation if you sell or book.** Commerce and travel journeys are supported first. Note where the agent took the most interactions.
4. **Check what an agent can read without scripts.** Run the [agent readiness check](https://postcliq.com/free-tools/can-ai-agents-read-my-site) on your contact, booking or product page.
5. **Test the enquiry path end to end.** Run [can visitors contact you](https://postcliq.com/free-tools/can-visitors-contact-you) , then submit your own form and confirm it arrives and shows a clear success message.
6. **Review your firewall policy per agent, not all at once.** If you use Cloudflare, look at the Personal Assistants category before blocking anything. Our guide to [Cloudflare's AI agent controls](https://postcliq.com/blog/cloudflare-personal-agents) covers which personal agents to let in.

## Frequently asked questions

### Can Google Analytics see AI agent visits?

Not as a separate category today. Agents that run in a real browser can execute your analytics tag and be counted as ordinary visitors. Clarity's Agent Activity dashboard is one of the first tools that labels supported agent sessions and shows what they did.

### Is Microsoft Clarity's Agent Activity dashboard free?

Clarity is free, and Microsoft says Agent Activity is generally available for projects that receive agent traffic. Simulations started with commerce and travel journeys. Ask Clarity, the conversational feature announced the same day, is a closed beta for retailers.

### Which AI agents does Clarity detect?

Clarity says its agentic analytics supports agent operators from Google, OpenAI and Anthropic. Agents from other companies, and agents that present themselves as ordinary browsers, are not labelled, so the dashboard shows a floor rather than a full count.

### Should I block AI agents that try to check out?

Usually not. Clarity's data shows agent sessions concentrate on checkout and booking, which are the visits closest to revenue. Block abusive traffic by behaviour, and allow identified personal agents through to the step where their person confirms.

## Can an AI agent finish the job on your site? Check it free.

PostCliq checks whether people and agents can read your pages, reach your contact path and complete your form, and puts three fixes first.

Free. No credit card.

![Nirmal Gyanwali, founder and chief executive of WP Creative.](https://postcliq.com/authors/nirmal-gyanwali.jpg)

Written by

[Nirmal Gyanwali](https://postcliq.com/blog/author/nirmal-gyanwali)

Founder and CEO, WP Creative · Sydney, Australia

Nirmal Gyanwali has spent more than twenty years building websites, and most of that time watching them fail at the one job the business actually bought them for, which is turning a visitor into an enquiry. He founded WP Creative in Sydney in January 2016.

- [LinkedIn](https://www.linkedin.com/in/nirmalgyanwali/)
- [YouTube](https://www.youtube.com/@nirmalgyanwali)
- [Instagram](https://www.instagram.com/nirmalgyanwali)
- [TikTok](https://www.tiktok.com/@nirmalwpc)

## Keep reading

- [**AI agent detection: Muse, Instinct and Grok Bot stopped saying who they are**](https://postcliq.com/blog/ai-agent-detection)
- [**Chrome's Gemini autofill: your enquiry form now has an AI filling it in**](https://postcliq.com/blog/chrome-gemini-autofill)
- [**Should you block AI agents? Amazon says yes, Walmart says come in**](https://postcliq.com/blog/block-ai-agents)

[More on AI search and visibility](https://postcliq.com/blog/topic/ai-search) · [All articles](https://postcliq.com/blog)