Some of the visits to your shop are no longer made by a person. They are made by software working for a person.
On 29 October, Microsoft published a study from its free website analytics tool, Clarity. It looked at more than 5.2 million website sessions involving AI agents. These are AI tools from the OpenAI, Google and Anthropic families that open a website, click around and fill in forms on someone’s behalf. The most common action recorded in those sessions was checkout.
That matters for anyone who reports conversion rate to a board. These visits land in the same analytics as your customers. Most of the time, nobody can tell them apart.
For ecommerce and digital directors, the lesson is short: some of your visitors are now software shopping for a person. Find out how many before you trust your conversion rate.
What Microsoft Clarity found
Microsoft opened its study with this line:
“Most analytics stacks were built to understand human visitors, not agents acting on their behalf.”
The study covered sessions on 230,511 Clarity projects (websites) between 13 August and 9 September 2026. In plain numbers:
Checkout was recorded in 25.21% of sessions involving an AI agent, against 14.65% of sessions by people on the same sites.
Form submissions were almost the same: 19.85% for agent sessions, 18.56% for people.
Outbound clicks went the other way: 14.28% for agent sessions, 26.99% for people.
Add to cart was recorded in 2.33% of agent sessions and 3.31% of people’s sessions.
Microsoft says agent session volume in its data “increased by 88% in just three weeks.”
Microsoft is careful about what this proves. A session can start with the agent and end with the person taking over, and Clarity “cannot currently separate the agent and human portions into different sessions.” It also says the numbers “should not be interpreted as steps in a single conversion funnel or proof that an agent independently completed a task.”
Microsoft also made a new Agent Activity Dashboard generally available in Clarity, for sites that receive agent traffic.
Source: Microsoft Clarity — What AI Agents Do on Websites: Insights From 5.2 Million Sessions (Clarity Staff, 29 October 2026, checked 2 November 2026).
What the experts say about the numbers
On 1 November, Luis Rijo at PPC Land took the study apart line by line. Three points matter for directors:
“Checkout” is not “order”. In people’s sessions, Checkout was recorded in 14.65% but Order success in only 1.79%. So Checkout here means reaching or clicking into checkout, not paying.
The base looks smaller than the headline. By PPC Land’s calculation, the percentages work out against about 1.22 million sessions, not 5.2 million. Microsoft has not explained the gap.
Some agents are missing. The study names only the OpenAI, Google and Anthropic families. PPC Land notes that Meta’s Muse and Perplexity are not named, and that Arkose Labs found only 2.3% of legitimate agent sessions on its network carried a verified ID.
Source: PPC Land — Checkout appears in 25% of AI agent sessions vs 15% for humans, Clarity finds (Luis Rijo, 1 November 2026).
What it means, in plain terms
An AI agent is not the same as a search bot. A search bot reads pages. An agent uses a browser like a person: it searches your site, adds products, opens checkout and fills in forms.
Your analytics may count it as a normal visit. Google Analytics removes “known bots and spiders” using Google’s research and an industry list. Google’s help page also says you “cannot disable known bot traffic exclusion or see how much known bot traffic was excluded.” Google has not said how Google Analytics treats AI agents that browse in a normal browser. If the agent loads your pages and your tags, your tags can fire as if a person were there.
Many agents do not prove who they are. PPC Land reports Arkose Labs found only 2.3% of legitimate agent sessions on its network carried a verified ID. Without one, tools can only guess from behaviour.
This is different from two earlier notes. The AI Assistant channel is a floor, not the full AI discovery story covered people who click through from ChatGPT or Gemini. The sale your tag never sees still has to be counted covered orders placed inside AI apps. Today is the third piece: AI agents visiting your own site.
Why it matters for the P&L
Conversion rate can move for the wrong reason. If agent visits are counted as sessions, they change both the top and bottom of your conversion rate. A board may see a drop or a jump that has nothing to do with your customers.
Ad platforms may learn from the wrong signals. If Google Ads or Meta bid on checkout starts or add to carts, agent sessions can feed those signals. You could end up paying to find more software, not more buyers.
Remarketing can chase software. An agent that reaches checkout and leaves could land in an abandoned-basket audience, depending on your setup.
A/B tests can be skewed. Agents may behave very differently from people on the same page.
Peak week is close. Black Friday is 27 November. If your numbers shift between now and then, you want to know whether it is people or agents.
What is not clear yet
How many of your visits are agents. The Clarity figures are an average across many sites. Your share could be much higher or lower.
How Google Analytics, Google Ads and Meta treat these sessions. None of them has said, as far as we can find.
Whether agents are buying. The study cannot tell who pressed the final button, the agent or the person.
How Microsoft detects agents. The study does not explain its method or give a split by agent, country or industry.
Director checklist
Judge sales on your own orders. Your order system or payment provider records real paid orders. Use that, not sessions or checkout starts, as the number the board sees.
Check what your ads bid on. In Google Ads and Meta, make sure purchases are the main goal. Keep add to cart and checkout starts as watch-only signals.
Compare checkout starts with orders, week by week. If checkout starts rise but orders do not, look closer before you celebrate or panic.
Look for agent traffic. If you use Microsoft Clarity, open the new Agent Activity Dashboard. If not, ask your website host, CDN or bot-protection supplier what they can see.
Try it yourself. Ask your team to shop on your own site with one of these AI browser agents, then check what Google Analytics and your ad pixels recorded.
Do not block agents in a panic. Some agents bring real buyers who take over at payment. Decide with your payments and fraud teams what to allow.
Mark the date. Add a note in Google Analytics for 4 November and take a before picture now: sessions, checkout starts and paid orders by day.
Tell the board what conversion rate means. Say whether it is based on all sessions or people only, and that agent visits may be included.
Questions to ask your team this week
Which number do we report to the board: checkout starts or paid orders?
What do our Google Ads and Meta campaigns bid on?
Can any tool we use tell us how many visits are AI agents?
Has anyone tried buying from our site with an AI agent, and what did our tags record?
Are our checkout starts and paid orders moving together?
Caveats
This is one study. It covers four weeks of Clarity sites, ending 9 September 2026. Agent use may have grown since.
Agents are not all bad. Many work for real shoppers who still pay. The aim is to measure them, not to fear them.
Detection is not perfect. Agents that do not identify themselves are hard to spot, so any count is likely to be a floor.
Your books win. Whatever your analytics say, judge performance against paid orders in your own system.
Final thoughts
#27/100 — Some of your visitors are now software shopping for a person. Find out how many before you trust your conversion rate.
AI agents will change how people shop. For now, the practical job is simple: keep the board’s numbers tied to paid orders, keep ad platforms learning from purchases, and start watching for agent traffic before peak week.
Not sure what your Google Analytics and ad tags are counting? Run our free GA4 and GTM tracking scan at https://audit.taggurus.co.uk/, email info@taggurus.co.uk, or book a meeting today to discuss.
FAQ
Q: What is an AI agent visit?
A: A visit where an AI tool uses a web browser for someone. It can search your site, add products, open checkout and fill in forms.
Q: Is this the same as the AI Assistant channel in Google Analytics?
A: No. That channel shows people who click a link in ChatGPT, Gemini or similar tools and land on your site. An agent visit is software browsing your site for a person.
Q: Does Google Analytics filter out AI agents?
A: Google Analytics removes known bots and spiders automatically, and you cannot see how many it removed. Google has not said how it treats AI agents that browse in a normal browser.
Q: Are AI agents buying from us?
A: Possibly, but nobody can say for sure yet. Microsoft says its sessions may include a person taking over, and its “Checkout” count is not the same as a paid order.
Q: Should we block AI agents?
A: Not in a panic. Some bring real buyers. Talk to your payments and fraud teams first, and measure before you decide.
Q: How can we see agent traffic?
A: Microsoft Clarity has a free Agent Activity Dashboard for sites with agent traffic. Your website host, CDN or bot-protection supplier may also be able to help.
Q: What should we do first?
A: Make sure the board sees paid orders, and that your ad platforms bid on purchases, not checkout starts.
Sources
https://clarity.microsoft.com/blog/agentic-analytics-what-agents-do-on-websites/ (Microsoft Clarity, “What AI Agents Do on Websites: Insights From 5.2 Million Sessions”, Clarity Staff, 29 October 2026: “Most analytics stacks were built to understand human visitors, not agents acting on their behalf.”; 5.2 million+ agentic-classified sessions across 230,511 Clarity projects, 13 August–9 September 2026 (UTC), OpenAI, Google and Anthropic ecosystems; Checkout 25.21% of agentic sessions vs 14.65% of human sessions; agentic-session volume “increased by 88% in just three weeks”; sessions “may also include a human handoff”; Agent Activity Dashboard generally available)
https://ppc.land/checkout-appears-in-25-of-ai-agent-sessions-vs-15-for-humans-clarity-finds/ (PPC Land, Luis Rijo, 1 November 2026: percentages imply a base of about 1.22 million sessions by PPC Land’s calculation; Order success 1.79% of human sessions; Meta Muse and Perplexity not named in the study; Arkose Labs: 2.3% of legitimate AI agent sessions presented a verified signature)
https://support.google.com/analytics/answer/9888366?hl=en (Google Analytics Help, “Known bot-traffic exclusion”: known bots and spiders automatically excluded using Google research and the IAB International Spiders and Bots List; “you cannot disable known bot traffic exclusion or see how much known bot traffic was excluded.”)
