Somewhere in your analytics this morning there is a checkout event that no human hand pressed. Maybe it was a person who let an assistant fill the cart and then took over to pay. Maybe it was software all the way to the payment page. Your dashboard cannot tell you which, and neither, for now, can the tool that just published a 5.2 million session study of it.
That would be a curiosity if checkout events stayed in analytics. They do not. They get imported into Google Ads and Meta as conversions, and the bidding systems learn from them. Black Friday is November 27. The question for this week is simple and slightly uncomfortable: what exactly are you teaching your bidders to buy more of?
What Microsoft actually counted
On October 29 Microsoft’s free analytics product, Clarity, published a study of more than 5.2 million sessions involving AI agents, recorded across 230,511 Clarity projects between August 13 and September 9. Sessions were flagged by detecting agentic browser providers from the OpenAI, Google and Anthropic ecosystems.
The headline: checkout was recorded in 25.21% of agent sessions, against 14.65% of human sessions on the same group of sites. Form submissions were almost level, 19.85% for agents and 18.56% for humans. Appointment bookings showed up in 8.42% of agent sessions and did not make the human top ten at all. Outbound clicks ran the other way, 26.99% for humans and 14.28% for agents. On-site search was higher for agents too, 7.16% against 4.79%.
Clarity also says agent session volume in its data “increased by 88% in just three weeks,” and it used the study to announce that its Agent Activity Dashboard is now generally available for projects that receive agent traffic.
Read the caveats, because Microsoft wrote them itself. An agent session “may also include a human handoff,” and Clarity “cannot currently separate the agent and human portions into different sessions.” The events are “independent, non-mutually-exclusive session counts” and “should not be interpreted as steps in a single conversion funnel or proof that an agent independently completed a task.”
So no, this does not prove robots are buying a quarter of the time. It proves something more useful for a marketer: sessions with agents in them are hitting the exact events you call conversions, at different rates than people do, and the tools you use to count those events were built for people.
The fine print is the story
PPC Land’s Luis Rijo went through the tables on November 1 and found two things worth carrying into your next reporting meeting.
First, the base. Divide each published count by its published percentage and every row lands on roughly 1.22 million sessions, not 5.2 million. By his math, if the percentages had used the full headline number, checkout would sit near 5.9% of agent sessions, not 25.21%. The study does not explain the gap. One possible reading is that the percentages only cover sessions where Smart Events recorded something, but nobody outside Microsoft can confirm that yet.
Second, the shape. In agent sessions, checkout outnumbers add to cart by more than ten to one. Order success, a separate event, appears on the human list at 1.79% and is below the agent top ten entirely. For humans, “checkout” at 14.65% next to “order success” at 1.79% tells you checkout means reaching or touching checkout, not paying. The numbers fit agents landing straight on the money page from a link or a prepared cart, then stopping, handing over to a person, or getting stuck. The published data cannot say which.
Neither point kills the study. Both point at the same operating problem. The event you probably treat as a strong buying signal is the event agents inflate most, and the event that proves money moved is the one where they barely register.
Most agents never say who they are
If the fix were “filter out the agents,” this would be a two line memo. It is not, because most agents do not identify themselves.
On October 20 Arkose Labs, which sells bot and fraud defence, reported that 97.7% of the AI agent sessions it judged legitimate between July 1 and September 30 carried no verified disclosure. It says Meta’s Muse “appears on networks as a Chrome browser.” PPC Land’s breakdown of the full report adds that ChatGPT agent was the one product signing at scale, on 78.4% of its sessions, using the Web Bot Auth standard. Arkose also says 65.2% of its customers saw agents on protected flows such as login, registration and checkout by September 30. Weigh those numbers knowing Arkose sells the cure, and that it publishes percentages with no raw counts.
Put the two reports side by side and the picture for a marketer is plain. Agents are present on the pages that fire conversions. Most of them look like ordinary Chrome. Clarity’s own sample covers OpenAI, Google and Anthropic and closed on September 9, one day after Meta launched Muse in the US, so the agent that security vendors watched surge through September is not in the count. Whatever share your analytics thinks is agent traffic, the real share is probably higher.
Your bidder cannot see the difference either
Here is where it stops being an analytics story. Google’s own Smart Bidding documentation says the system optimizes “for conversions or conversion value in every auction,” that you need conversion tracking to use it, and that it reads signals such as device, operating system and browser. It learns which kinds of visits turn into your chosen conversion and bids up for more of them.
Now picture an account whose primary conversion is a checkout page view or a begin checkout event, which some accounts do when the sale closes offline or the payment step is hard to tag. If sessions involving agents trigger that event at a higher rate than people, and if those sessions look like ordinary Chrome visits from the same places, the bidder gets a clean, cheerful signal that some slice of traffic converts well. It cannot know that part of that slice reached the payment page and stopped. You will not see a red flag. You will see a nicer conversion rate.
The same logic hits lead gen from a different angle. Clarity’s form submission rates for agents and humans are nearly identical. That means frequency gives you no clue. A pay per lead program, or any campaign optimizing on raw form fills, has no way to spot agent filled forms from volume alone. The tell will show up later, in contact rates and sales acceptance, long after the bidder has learned from them.
None of this says agent sessions are worthless. Some of those handoffs end in real orders from real customers who simply let software do the boring part. The point is that you do not know the ratio, and a bidder trained on a mixed signal at peak season is an expensive way to find out.
Retailers with leverage are already splitting the ledger
The cleanest model came from Walmart on October 21. In its announcement of Agent Profiles, the company wrote that today an agent “signs in using a customer’s account,” so purchases by the customer and the agent “appear together, without any differentiation.” Agent Profiles, due within a month, give each authorized agent its own profile so its purchases sit apart, with approval required by default and optional spending limits such as $50, $100 or $200. Walmart is also building an Agent Gateway so agents stop clicking through a site “built for people to browse, click and sign in themselves.”
Walmart wrote that for customer control. Read it as a marketer and it is a measurement design: label the agent at the door, keep its orders on a separate line, and stop pretending a person and their assistant are the same visitor. Most brands cannot build a gateway by Black Friday. Every brand can copy the idea of a separate line.
What to do this week
Black Friday is 25 days out. Changing conversion goals inside that window has a cost, because bidders need time to relearn, so the useful work is this week and next.
Look at your own number before you argue about Microsoft’s. If you run Clarity, open the agent dashboard and see what share of sessions on your checkout, booking and lead pages involve agents. If you run another analytics or bot tool, ask the team that owns it the same question. You want a per event share, not a sitewide one, because the damage sits on conversion pages.
Move primary conversions to proof of money. If any campaign optimizes on checkout page views, begin checkout or similar events, switch the primary goal to a server confirmed purchase, or at least the order confirmation, and demote the softer events to secondary. The Clarity tables suggest agents inflate the step before payment and barely touch the confirmation. Do it this week so the bidder has time to settle before the peak.
Make lead gen prove itself downstream. For form driven campaigns, start sending qualified or contacted leads back as the optimization event through your offline conversion import, rather than raw submits. If that is too big a change before peak, at least tag form fills where your bot tool or analytics flags an agent, and watch their contact rate for two weeks.
Do not reach for the block button. Arkose’s own warning is that blocking all automation turns away real customers acting through an assistant. A bot rule that kills agent sessions on checkout may cost you orders from people who were going to pay. Check what your bot defence currently does on checkout and booking pages, and make that a decision someone signed, not a default.
Split your test readouts. Any checkout or landing page A/B test reading out before Black Friday should be cut by agent flagged and unflagged sessions where your tools allow it. An agent that goes straight to checkout on one variant can make a page look better than it is.
Ask for the separate line. If you sell through retailers or marketplaces, ask your account team whether agent orders will be reported separately, the way Walmart plans for its own customers. If you run your own store, ask your commerce platform the same question. The answer tells you how long you will be measuring with one eye shut.
The agents are early, the numbers are messy, and Microsoft’s study has gaps it does not explain. That is exactly why the decision belongs to you this week and not to your bidder. Agents are already pressing your checkout button. Make sure the machine that spends your money is only rewarded for the receipts.
