Agent sessions look like bounces: how to tell them apart
An agent that gives up at the address form is logged as a bounce, same as a person who left. Three checks that separate agent traffic from human traffic in data you already have.
Three actors abandoned at the address form on Tuesday. Same page, same field, three separate takes. In the analytics they read as three sessions, each around forty seconds, no error recorded. Nothing in the row says the visitor was not a person.
That is the whole problem. Not detection for its own sake, but the difference between a person who left and an agent that could not proceed, because both produce the same row.
What an agent session is
An agent session is a browser session driven by a goal rather than a person. It loads pages, reads what the DOM exposes, fills fields, submits forms, and checks whether the outcome matches the goal. It does this without a pointer, without a scroll, and without patience for a modal that has no reachable dismiss control.
The instrument you already have was built for the other visitor. Heatmaps measure pointer movement and click density, and an agent does not move a pointer, so the heatmap is blank. Session recordings replay browser events a human generated, and an agent produces a different event stream. Funnels count completions and attribute drop-offs to steps, and an agent that fails mid-task is structurally identical to a human who left by choice.
So the agent that gave up at the address form is counted as a bounce, and the bounce is counted as a human decision.
What the volume actually is
Four figures, each from its own source.
AI-driven traffic to retail sites grew 4,700% year over year, according to Adobe Analytics. Firecrawl expects 20% of e-commerce tasks to be agent-handled. MCP public server adoption has crossed 9,400 servers, after a 35% uplift in usage in a single month in early 2026. PwC finds 79% of companies have adopted some form of AI agent technology.
None of those numbers describe a future. They describe traffic that already reaches the address form, the filter panel and the cart summary, on products whose analytics have no column for it.
Three checks that separate the two
You do not need new instrumentation for a first cut. Three things separate an agent session from a human one in data you already collect.
Request cadence. A person pauses: page load, then a gap while they read, then an interaction. An agent issues its requests in a burst, parse and act and submit with little idle time between them. A distribution of time-to-first-interaction with a visible second cluster near zero is the first signal.
Session shape. Agents run longer than people and repeat the same flow in sequence, because the visit is one step in a larger orchestrated task. Claude Mythos Preview reached a 16-hour autonomous task horizon in March 2026. A product that assumes one session maps to one human intent reads a repeated flow as a returning customer.
Event gaps. No pointer events, no scroll events, fields filled by attribute rather than by tab order. A form where a value appears without a focus event ever firing is the signature of a client that filled the field directly.
Each check is a filter, not a verdict. Together they give you a short list of sessions worth reading, which is the point: you are looking for the take where the scene broke, not a traffic report.
What to capture when you find one
A pass or fail tells you the outcome. It tells you nothing about the path. The useful artefact is the trajectory: what the agent did, in order, with the moment it changed course marked.
That is what a checklist produces. A checklist is a list of requirements an actor runs, step by step, and each finding pairs what the actor did with a screenshot of the state it reached. Run the same checklist before and after a release and you have a regression signal rather than a story about a bounce.
Start with one flow
Pick a flow an agent would attempt on someone's behalf: checkout, a returns request, account access. Define the goal rather than the steps. Send one actor through it, then read the trajectory against the session log for the same minutes.
If the agent stalls at a label that reads clearly to a person, you have a product decision, not a measurement problem. Fix the label and rerun the take.
The address form is a good first scene.