A post-chat survey is the cheapest feedback instrument you own. It arrives at the one moment the visitor still remembers what happened, it costs them a tap, and it produces a number you can track for years.
Which is why it is worth understanding the several distinct reasons a visitor may finish a conversation and never be asked anything at all. Some of them are settings. One of them is a deliberate product rule that no settings screen mentions.
There are two post-chat surveys
The embedded widget and the popup chat window each have their own post-chat survey screen, with their own settings behind them. Configuring one does not configure the other. This is the same split that catches people out with pre-chat and offline forms, and the diagnosis is identical: a form change that appears to do nothing is usually a change made on the screen your visitors are not using.
So the first question is never which switch is wrong. It is which of the two surveys your visitors actually reach. Settle that before you touch anything, because half the time the setting you are staring at is correct and simply belongs to the other surface.
The embedded survey is the narrower of the two: whether it is shown at all, and whether it asks for a rating, a comment, and an email address. The popup version carries the same four decisions plus its own wording, headline and template.
Every switch starts on
All four survey switches are treated as enabled unless something explicitly turns them off. A site that has never opened the screen has a survey with a rating, a comment box and an email field, because an unset value means never configured rather than off.
That is a friendly default and an awkward one to reason about. It means you cannot tell from the screen whether a setting is on because somebody chose it or on because nobody has ever looked. It also means turning something off is a real edit that has to be saved, not merely a matter of leaving a box empty.
The practical consequence: if you have inherited a site and want to know what visitors are being asked, open the widget and finish a chat yourself. The settings screen tells you what was configured. Only the widget tells you what is rendered.
The rule that withholds the survey entirely
This is the one that surprises people, and it is not a bug.
When a conversation is an AI conversation and no human agent is available to take over, the survey is suppressed. Not hidden by a setting, not skipped by the visitor: withheld by the runtime, whatever the four switches say.
The reasoning is sound. A rating collected on a conversation that a human never touched is not a measure of your support team, and mixing it into the same score as human conversations makes both numbers harder to read. If a human is ready to step in, the survey behaves normally.
The consequence for measurement is large and easy to miss. A site running AI answers outside staffed hours collects no survey responses for those conversations at all. If you compare this month against a month before you enabled AI, you are not comparing like with like: the denominator has quietly changed. Judge AI conversations with the measures built for them rather than expecting them in your CSAT feed.
Submit needs a rating; Close is the skip
When the rating control is shown, submitting the survey requires a star to be chosen. A visitor who submits without picking one is asked to choose a rating rather than having a zero recorded.
This exists because the earlier behaviour posted a zero-star record, which is not a rating at all: it is the absence of one, filed as the worst possible score. Enough of those and your average is measuring hesitation rather than satisfaction.
Skipping is still allowed, and the control for it is Close rather than Submit. That distinction is worth knowing before you read your response rate: a visitor who closes the survey is a deliberate non-answer, not a lost record.
A comment-only survey is perfectly valid. Turn the rating off and the comment box carries the survey on its own, with nothing required. On low-volume sites that is often the better instrument, because a handful of sentences beats a handful of stars.
The email box and the transcript box are not the same thing
The email input appears whenever the survey is set to ask for an email, and it is independent of the checkbox that offers the visitor a copy of the transcript. They look related on screen and they are not.
The validation follows from that. An email address is required only when the visitor asks for a transcript copy, because that is the case where an address is needed to deliver something. But an address that has been typed is checked for a valid format whether or not the copy box is ticked, so a malformed address does not get recorded simply because the visitor changed their mind about the transcript.
The format check is the strict one used elsewhere in the product rather than a loose pattern, which means addresses that look plausible but are not deliverable are rejected at the point of entry instead of polluting your contact records.
What all this does to your ratings
Put the rules together and your ratings feed has three distinct kinds of gap, and they are not interchangeable.
There are conversations that were never offered a survey, because they were AI conversations with nobody available. There are conversations offered a survey that the visitor closed. And there are conversations where the survey was configured off on the surface that visitor used.
Only the middle one is a real non-response. The other two are conversations that were never in the sample at all, and counting them as non-responses understates a response rate that is probably fine. This is one of the more common ways two reports about the same week disagree.
A diagnosis order that works
When somebody reports that the survey never appeared, work in this order and you will usually land in a minute or two.
First, establish which surface the visitor used, embedded or popup, and open that screen rather than the other one. Second, check whether the conversation was handled by AI with nobody available, which explains the absence completely and needs no setting changed. Third, confirm the survey is switched on for that surface, remembering that on is the default and off is the deliberate act. Fourth, finish a real chat yourself and watch what renders, because that is the only test that covers all of the above at once.
Only after those four would I start looking at wording, templates or anything cosmetic. In practice the answer is one of the first three almost every time.
What to measure
Track survey responses as a share of conversations that were eligible for a survey, not of all conversations. Once AI is answering some of your traffic, the two diverge, and the second number will drift downwards while nothing is actually wrong.
Watch the comment-to-rating ratio as well. Comments are where the actionable material lives, and a survey that collects plenty of stars and almost no sentences is usually asking at the right moment with the wrong instrument.
And if you change anything on this screen, note the date somewhere you will find it later. A change in average rating that lines up exactly with a settings change is nearly always the settings change, and that is far easier to see in your chat analytics when you know which day to look at.