The customer needs an appointment for a tyre change
“Find an appointment for a tyre change on Wednesday.” A customer sends her agent off with this request. The scene is invented, but it makes a practical design question clear: can the agent merely recommend a garage on your website, or does the customer end up with a booked appointment?
The agent needs to establish whether the garage offers the service, when it is available and what it costs. In this example, it should suggest an appointment first and book only after the customer agrees. We use Agent Experience to examine that whole journey from the agent’s perspective. This is our working definition, not an industry standard or universal score.[S1]
The agent suggests Wednesday at 2 pm and asks for approval
In this example, the agent finds an available appointment on Wednesday at 2 pm. It presents the service, price and conditions and asks: “Does that suit you?” Only after the customer agrees does it submit the booking. If information is missing or the appointment has been taken, it must ask a question or suggest another time. This approval step is chosen for our example.[S1]
The website has work to do too. A field labelled “Choose an appointment” and a clear response help the agent take the right step. Google describes website agents using HTML, screenshots and the browser’s map of controls. Clear labels and appropriate controls remain useful foundations.[S3]
The garage’s confirmation establishes that the appointment is booked
After submission, the customer needs confirmation from the booking system: tyre change, Wednesday, 2 pm. It must match her request. A completed form or an agent saying “Done” is not enough evidence. If confirmation does not arrive, the agent must not claim a successful booking.
This is how we would set up the first test. The table lists proposed checks, not measured results.[S1]
| What happens? | What should the test establish? |
|---|---|
| The agent looks for a tyre change. | It finds the right service and reads its price and conditions correctly. |
| It suggests Wednesday at 2 pm. | The appointment is available; the agent asks the customer before booking. |
| The customer agrees. | The booking system receives the approved service and appointment. |
| The appointment is taken or confirmation is missing. | The agent reports the problem and asks what to do, rather than claiming success. |
Include a taken appointment, missing approval and a missing confirmation in the test. Chrome’s WebMCP evaluation guidance likewise examines individual function calls and the complete journey to the result.[S5]
A recommendation alone does not show whether booking works
For the request to reach your garage, the agent has to find it first. SEO still matters. Google says its usual SEO foundations also apply to its generative search features. It does not use llms.txt to improve Search visibility or rankings; other systems may use such files.[S2]
Test both questions: does the garage appear for the relevant request, and can the customer’s agent complete the booking on your website? A mention in an answer addresses only the first question. A usable form does not explain how the customer finds you.
Bring a customer task like this to our Agent Experience work . Together, we can trace where the agent proceeds and where it needs help. For agents inside your own team, Hybrid Organisation addresses the complementary question: who supervises them and owns their work?
