The share of questions that got the refusal message

Refused assistant messages divided by all assistant messages in the window. The refusal is recorded as a flag written at the moment the answer was produced, not guessed later by matching the text, so the count does not drift if the owner edits their fallback wording.

Refused messages carry no model and no cost, which is what makes the arithmetic downstream work: a refusal is distinguishable from a cached answer even though neither called a model.

The retriever finds passages, the best score is compared against the threshold, and below it the fallback is returned before any model is involved. That ordering is the reason a refusal costs nothing and returns instantly.

So every fallback is one of three things: material that does not exist, material that exists but uses vocabulary the visitor did not, or a threshold set higher than the material can clear. Those need three different fixes, and the rate alone tells you which you have exactly never.

The insights page uses the last thirty days and counts messages. A fallback rate computed per conversation would be a different number and usually a larger looking one, since a conversation only needs one refusal to be tainted.

With no assistant messages in the window the figure is zero, which reads as a perfect record and means no data. Check the message count before believing a flattering rate.

Most people treat this as a number to minimise, and past a point that instinct is wrong. An assistant that never refuses is an assistant that answers questions your documents do not cover, which is the failure mode that produces confident nonsense.

A rate near zero on a modest set of documents means either your visitors ask an unusually narrow set of questions, or the threshold is loose enough that thin matches are being answered from. The second is far more common, and the way to tell is to read a sample of the weakest answers rather than to look at the rate again.

Lower the threshold. Every question that scored between the old bar and the new one now reaches the model, and every one of those becomes an answer rather than a refusal. The rate falls the same afternoon and not one document changed.

Rewrite the fallback message into something that sounds like an answer and the count does not change at all, but the people reading transcripts will stop noticing them. This is the accidental version and it is worse than the deliberate one, because it hides the signal rather than moving it.

Indexing a very large general document has the same effect as lowering the threshold without anybody deciding to lower it. There is always something to match, so there is rarely a refusal, and the answers that replaced those refusals are built on the vaguest passages you own.