The biggest thing that happens, and the least visible

Nothing. The close control and the escape key hide the panel, and hiding it is a message passed around inside the page rather than a request to anywhere. Nobody is told. There is no ending event, no duration, and no record that the panel was ever on screen.

A conversation record is written when the first question is sent, not when the panel opens. So a visitor who opened it, read the greeting, thought better of it and closed it again has produced no record of any kind. They are not a conversation with no messages. They are not there at all.

That is worth sitting with before reading any other number here, because it sets the boundary of what the whole system knows. The population everybody most wants to understand is outside every denominator on the insights page.

When it started, when the last message arrived, how many messages it holds, the address of the page it was opened from, a country where the request carried one, an email address if the visitor volunteered one, and a rating if they pressed a button. That is the shape of it.

None of those fields says the conversation ended. There is no closing step, nothing marks a thread finished, and a record that stopped an hour ago looks exactly like a record that stopped in March. The last message time is the only clue and it is a clue about when, never about why.

A visitor who got a perfect answer and left satisfied produces a conversation with one question and one answer. A visitor who got a useless answer and left annoyed produces a conversation with one question and one answer. Same shape, same length, same everything.

This is the reason a satisfaction figure cannot be inferred from behaviour, and the reason any product claiming to measure abandonment is measuring a rule somebody wrote rather than a thing that happened. Change the rule and the number changes, with no visitor having behaved differently.

The nearest honest move is to stop looking for a rate and start reading. A conversation holding one exchange is a candidate, and reading twenty of them tells you which of the two endings your assistant is producing. That answer is qualitative, it is bounded, and it is far more useful than a percentage would have been.

Three things, none of which is here. A signal when the panel closes, which means sending something on the way out and accepting that the browser will sometimes fail to deliver it. A way to recognise the same person on a later visit, which the widget deliberately avoids doing. And a way to know what they did next, which lives on your side of the fence rather than ours.

The third is the one that cannot be bought. Somebody who leaves the panel and telephones you has not been helped by the assistant, and no widget on earth can observe that call. Any abandonment figure is silently missing it, and it is usually the most expensive part.

Define it as a conversation holding only one exchange, and you can improve it in an afternoon by ending every reply with a question back. Visitors answer politely, conversations grow a second turn, and the figure falls without one person being better served. This is the most common proxy in the category and it is the easiest to move.

Define it as leaving without a handover, and the way to improve it is to push the handover harder at everybody, including the people who were perfectly happy. You have now converted a silent good outcome into an enquiry somebody on your team has to answer, and called it progress.

The accidental version needs nobody to decide anything. A month of visitors asking short factual questions produces short conversations, and any proxy built on length reports a worse month than one full of complicated back and forth. The traffic changed. The assistant did not.