Ask inside the experiment

It is the right thing to worry about, and the caps are built around it rather than bolted on afterwards. At most one survey is shown per session across every survey you are running, not one per survey, because two widgets in one visit is how you teach people to dismiss on sight and that damage is permanent. Somebody who has answered is never asked that question again, with no re-ask window at all. Somebody who dismissed it is left alone for a cooldown you set, and after a small number of prompts is left alone regardless. If the browser cannot tell whether they have been asked before, nothing is shown at all.

If all you want is a score, a dedicated tool is fine. What you cannot get from one is the thing this exists for: every response records which variant of which experiment the visitor was in when they answered, so the same question can be read per arm. An experiment tells you variant B converted better and cannot tell you why. Read the answers per variant and it often can, and the why is what decides what you test next. It also catches the case a conversion metric hides entirely, a variant that wins on conversion while scoring worse on effort.

Not on pages with no survey running, because nothing is requested at all. The survey code is a separate file from the experiment runtime, and a website with no survey never asks for it: the flag that would trigger the request is simply absent rather than set to false. Where it does load, it loads last, after the page is already visible and after measurement has been requested, so a failure inside it cannot cost you a render or a conversion event.

This is the honest failure mode of most feedback programmes: the free-text answers are the valuable part and nobody reads them past the first fifty. Every written answer is classified positive, neutral or negative and the response feed filters on it, which makes the negative bucket reachable, and that is where the specific, actionable complaints are. It is per answer rather than an overall mood, so a high score sitting next to negative comments shows up as the disagreement it is instead of being averaged away.

Each one is defined by its scale and by the moment it belongs in. NPS asks about loyalty on nought to ten, after a completed purchase or a support interaction. CSAT asks about satisfaction on one to five, immediately after a specific interaction. CES asks how easy it was on one to seven, after checkout or a self-service task, and it is usually the most useful to conversion work because it finds friction rather than feeling. All three arrive with the standard wording already in place, because an NPS question that does not ask about recommending is not NPS whatever scale it renders.