An AI tries to become your customer. We record exactly where it stops.

What we observe

Four revenue-critical journeys, because these are the ones an agent has to finish on a customer's behalf.

• BookingReaching a bookable slot: finding the service, the location, the availability, and the point where a booking would be confirmed.
• QuoteGetting to a priced or scoped answer: configurators, eligibility questions, and the point where a quote request would be submitted.
• CheckoutReaching a completable basket: variant selection, delivery options, and the point where payment details would be entered.
• ClaimsReaching the right claim or support path: triage routes, policy or order lookup, and the point where a claim would be filed.

The five-step method

1. Declare the journeyWrite down the goal in the customer's terms and the checkpoints that have to be reached, before the run starts.
2. Attempt it with named agent familiesRun the declared journey with agents we name in the record, so the attempt is attributable rather than anonymous.
3. Record checkpoint statesMark each checkpoint reached or stopped, with the observation that decided it. Observation before implication.
4. Reproduce and controlRepeat the run, and compare against a control journey, so a one-off is not reported as a pattern.
5. Watch for changeRe-run on a schedule. A journey that changes is itself the finding; the date is part of every record.

What a record contains

The structure of a record, not a result. We publish no findings until there are runs behind them.

• Goal The declared journey, in the customer's words.
• Agent The agent family and version used for the attempt.
• Timestamp When the run started, so it can be placed against site changes.
• Checkpoints Each checkpoint and its state: reached or stopped.
• Final state Where the run stopped, and the observation that decided it.
• Likely fix owner Which team would hold the change: content, front-end, platform, or policy.
• 60-second self-check The steps to reproduce the stop yourself, without taking our word for it.

Public research and authorized tests are two different lanes

Public research stops at the form. Anything deeper needs written authorization. Every capability here is paired with its boundary.

• Lane A — public, read-only

  What any visitor could see

  Public pages only. No text typed into forms and nothing submitted. No accounts, logins, payments, or calendar holds. No bypassing of challenges, policies, or access controls. Paced, capped, and one session at a time.

• Lane B — written authorization

  Deeper testing, on your terms

  Only after written authorization naming the domains, the journey, the test identity, the rate, the synthetic data, and the consequential-action boundary. You set the limits before the first run.

Read the research disclosure in full

If you found us in your logs

Our sessions declare this user agent:

Concurrency

1 session per site at a time.

Pace

No more than 1 top-level navigation or action per 10 seconds per host.

Daily ceiling

No more than 40 top-level navigations per site per day.

Published signals

Robots, content signals, and terms are respected.

Opt out

Email research@agentjourneylab.com. We stop future public research on a verified request. No form, no account, no reason required.

Talk to us

Write to research@agentjourneylab.com — for an authorized pilot, an opt-out, or a question about the method. One address, read by the people running the research.

Discuss an authorized pilot Request an opt-out