Agent vs person: why the two are counted apart
An AI agent's visit is real, and a sale it completes is a real sale. But its behavior is not a person's. The agents observed moved and typed differently from people. Counting agents as people distorts every number built on how people behave. So agents are counted as their own population, and measured on their own terms.
Last updated
What two questions does every visit answer?
Is this visit real, and whose behavior is it? The first decides whether a visit counts at all. The second decides which population it belongs to. An agent's visit is real, so its sale counts. It is not a person's behavior, so it stays out of anything that measures people.
| Question | A person's visit | An agent's visit |
|---|---|---|
| Does a completed sale count as revenue? | Yes | Yes |
| Does it count toward the effect on people? | Yes | No |
| Does it train what the system learns about people? | Yes | No |
| Where is its own result read? | The conversion rate | The Agent Checkout Rate |
How does an agent's behavior differ from a person's?
In almost every signal a site reads. The pointer of Meta's Muse jumped instead of traveling. xAI's Grok agent filled a checkout form with zero keystrokes. Bench agents scrolled past in-page menus, and in one model family closed every pop-up. Signals built to read a person's hesitation mean something else, or nothing, for an agent.
What goes wrong when agents are counted as people?
The effect on people is misread. On 2026-09-21, before a rule recognized its jumping pointer, a Muse signup on cromanion.com was counted as a person's conversion, in the control group, and credited as assisted. In another case, a headless scanner, a machine rather than an AI agent, produced 29 visits counted as people in five days, in both groups.
Each case moves a number that is meant to describe people. The first adds a conversion no person made. The second adds visits no person made, and dilutes the rate on both sides.
The separation is built into how every figure is counted: one rule for whether a visit is real, another for whose behavior it is. Each rule is defined once, so two figures cannot disagree about it.
Why not leave agents out entirely?
Because their sales are real, and their share of visits is a fact about the store. An agent that completes a purchase brings in the same money as a person. Dropping agents would understate revenue and hide a population the store really has.
So agents stay in every count of sales and revenue. They leave only the reads that describe or learn from people.
How are agents measured, then?
On their own, with their own control group. The Agent Checkout Rate compares agent visits where the system acted with agent visits where it held back. People are compared with people. Neither population's result is read into the other's.
Common questions
What about an agent that is not recognized?
It is counted as a person, because nothing distinguishes it from one. The separation covers recognized agents: those that sign their requests, that a network provider verifies, or that behave like a driven browser.
Is a person using an AI assistant to shop an agent visit?
When the assistant drives the browser, yes: the behavior on the page is the agent's. When the person reads the assistant's answer and then visits the store themselves, the visit is the person's.
See it on your own site
Paste one tag. Cromanion crawls your site, watches real sessions in Learn mode, and only acts when you switch it on — with a permanent 10% holdout proving what it caused. Free to start, no credit card: a 14-day or 1,000-session live trial, then keep measuring for free.