AI Agents Are Visiting Websites: What Microsoft’s Clarity Data Shows

AI agents are now browsing websites on people’s behalf, and Microsoft has put some large numbers on it. In a 4 November 2026 post on its Agent Activity dashboard, Microsoft Clarity, its free website analytics tool, said that in its own analysis “weekly sessions from AI agents nearly doubled in three weeks to almost 2 million” [1]. Microsoft says the dashboard shows which agents visit a site and what they do once they arrive, “ranging from searches and form submissions to logins, downloads, appointment bookings, and checkout”, and that the new capabilities are “available globally to all our users”.

A week earlier, on 29 October 2026, the Clarity team published the study behind those numbers. We build sites, search and conversational commerce for retail and D2C brands, so we wanted to see what agents actually do when they land on a site, and how much of it a retailer can see today. We’ve also written about what Google Connected Apps means for UK retailers, the other half of this shift.

Microsoft’s figure for weekly AI agent sessions across Clarity sites. Source: Microsoft Clarity [1].

What Microsoft measured

The study looked at more than 5.2 million sessions involving AI agents across 230,511 Clarity projects, between 13 August and 9 September 2026 [2]. Sessions were identified using detected agentic browser providers across the OpenAI, Google and Anthropic ecosystems. Separately, the Clarity team says that in its own data agentic-session volume “increased by 88% in just three weeks”.

The study is careful about what it can and can’t show, and we think that care matters as much as the headline numbers:

• A session can include a person. Microsoft says an agentic session “may also include a human handoff”, and that Clarity “cannot currently separate the agent and human portions into different sessions”. So the figures describe activity recorded during sessions involving agents, not necessarily actions an agent completed on its own.
• The actions aren’t a funnel. Each figure is a separate count of sessions in which an action was recorded, across different websites and industries. Microsoft says they “should not be interpreted as steps in a single conversion funnel or proof that an agent independently completed a task”.
• Only some agents are covered. The supported agents for the dashboard are, in Microsoft’s words, “Google, OpenAI, and Anthropic agent operators”.

Microsoft also doesn’t break the figures down by country, so the study can’t tell us how much of this traffic reaches UK sites.

The actions come from Clarity’s Smart Events, which group clicks, page views and other signals into business actions. Clarity’s documentation lists nine automatic event types, from Purchase and Add to Cart to Login and Download, and project admins can create up to 20 custom smart events without code [3]. The study also reports actions such as outbound click and book appointment, and Microsoft doesn’t say how each one is defined.

Share of agent and human sessions in which each action was recorded. Source: Microsoft Clarity [2].

What agents did in those sessions

Checkout was the most frequently recorded action in agentic sessions. It appeared in 306,621 of them, 25.21% of agentic sessions, against 14.65% of human sessions. For people, the top action was an outbound click, at 26.99% of human sessions, compared with 14.28% of agentic sessions.

Submit form looks much the same for both groups, recorded in 19.85% of agentic sessions and 18.56% of human sessions. Search was more common with agents, at 7.16% against 4.79% for people. Book appointment came fourth for agents, at 8.42% of agentic sessions.

One thing stood out to us as people who build on-site search and checkout. Add to cart was recorded in only 2.33% of agentic sessions, far fewer than checkout and below the 3.31% for human sessions. We read that with Microsoft’s caveats in mind: these are counts of recorded actions, not journeys, and a person may have taken over part way through.

Microsoft’s own reading is that task-oriented actions, such as checkouts and appointment bookings, occur at a higher rate in agentic sessions, and that understanding what causes that gap “will be important as agentic commerce evolves”.

How agents identify themselves

Analytics only helps if a site can tell an agent from a person. The platforms are starting to publish ways to do that.

Google lists a user-triggered fetcher called Google-Agent, which it says “is used by agents hosted on Google infrastructure to navigate the web and perform actions upon user request” [4]. It uses IP ranges from Google’s user-triggered-agents.json file, and Google says it is also experimenting with the Web Bot Auth protocol, using the https://agent.bot.goog identity. Two details on that page matter for retailers. Because these fetches are requested by a user, Google says they “generally ignore robots.txt rules”. And Google warns that the user agent string can be spoofed, so the name alone doesn’t prove a visit came from Google.

OpenAI’s agent browser signs its requests instead. OpenAI says ChatGPT Work’s Cloud browser uses Web Bot Auth to sign outbound HTTP requests under the HTTP Message Signatures standard, RFC 9421, with a Signature-Agent header set to "https://chatgpt.com" and public keys published for checking each request [5]. OpenAI also says that “at launch, Cloud browser cannot sign in to websites or complete payments”, and that legitimate requests can be blocked if a proxy strips the three signature headers.

Ways agent visits can be recognised, as the platforms describe them. Sources: Google for Developers [4], OpenAI Help Center [5], Microsoft Clarity [1].

How we’re looking at agent traffic on the sites we build

None of this needs a new kind of website. In our experience, the things that trip up agents are often the things that already trip up people and search engines. This is how we’re approaching it.

1. Seeing it first. We start by checking whether agent visits are visible at all, in analytics and in server or CDN logs, and we keep agent sessions apart from human ones before reading any conversion figures. A rise in “checkout” that’s partly agents tells a different story from a rise in shoppers. We cover the log side in our guide to measuring AI search and referrals.
2. Letting the right agents in. We check that bot protection, firewalls and CDN rules aren’t blocking verified agents by accident, and that proxies pass signature headers through untouched. Equally, we don’t treat a user agent name on its own as proof of who is visiting.
3. Making the key tasks machine-friendly. Search, product pages, forms and checkout are where agents spend their time in Microsoft’s data. We look for clear labels, predictable form fields, prices and stock in the page itself, and on-site search that copes with a plain-language query rather than only an exact product name. That’s the same work behind our AI website search builds. We go through readiness in Is your e-commerce website ready for AI shopping agents?
4. Testing the journey end to end. Microsoft says Clarity’s simulations show whether a task was completed, how long it took and how many interactions it needed, starting with commerce and travel scenarios. We treat that kind of test as part of conversion rate optimisation: if an agent stalls at the same step as a person, fixing it helps both.
5. Reading the numbers carefully. Microsoft’s figures cover three agent operators and can’t separate the agent from the person. We keep the same caution with client data and don’t plan around agent conversion rates nobody has measured yet.

Firney is a UK web development and AI engineering company. We build fast websites, AI site search, product recommendations and conversational commerce for retail and D2C brands. Firney is a Prismic partner and a Google Cloud partner.

Our Free GEO audit shows how AI systems read a site today, with results in a few minutes, and we’re happy to talk it through on a call with us.

References

1. Microsoft. “Agents act, LLMs recommend, and people browse. Microsoft Clarity connects all three journeys.”. Microsoft Clarity Blog. 4 November 2026. https://clarity.microsoft.com/blog/clarity-connects-agents-llms-people/
2. Microsoft. “What AI Agents Do on Websites: Insights From 5.2 Million Sessions”. Microsoft Clarity Blog. 29 October 2026. https://clarity.microsoft.com/blog/agentic-analytics-what-agents-do-on-websites/
3. Microsoft. “Smart Events”. Microsoft Learn. Accessed 11 November 2026. https://learn.microsoft.com/en-us/clarity/setup-and-installation/smart-events
4. Google. “Google User-Triggered Fetchers”. Google Crawling Infrastructure, Google for Developers. Last updated 19 August 2026. https://developers.google.com/crawling/docs/crawlers-fetchers/google-user-triggered-fetchers
5. OpenAI. “ChatGPT Work’s Cloud browser allowlisting”. OpenAI Help Center. Accessed 11 November 2026. https://help.openai.com/en/articles/11845367-chatgpt-works-cloud-browser-allowlisting