One word, two numbers, and the distance between them

One minus the count of assistant messages that returned the fallback, divided by the count of all assistant messages, over the last thirty days. That is the whole calculation and there is nothing else in it.

Note what is in the denominator: assistant messages, not conversations. A single conversation with six answers and one refusal contributes six sevenths, not a whole conversation marked good or bad. This makes the figure weight talkative visitors more heavily than brief ones, which is a real property of the number rather than a defect, but it is not how most people picture it.

When there are no assistant messages in the window the figure is zero rather than a hundred percent. An empty denominator has no honest answer, and showing a perfect score for an assistant nobody has used would be the least honest of the available choices.

A refusal is the fallback message, the wording the owner wrote, returned because the best retrieved passage scored below the match threshold. It is recorded as a flag on the message, so the count is exact rather than inferred from the text of the answer.

This matters because it excludes a whole category people assume is included. An answer that hedged, waffled, or told the visitor to check elsewhere in its own words is not a refusal by this definition. It reached the model, the model produced prose, and it counts as answered. The number cannot tell a confident correct answer from a confident wrong one, and it was never built to.

Thirty rolling days, ending now. There is no comparison against the previous period on that page, no split by topic, and no way to see whether the figure moved after you published something.

So the tile gives you a level and not a direction, and a level on its own is close to uninterpretable. If you want the direction, write the figure down on the same day each month. Two minutes of clerical work turns a decorative number into a usable one, and a day by day series of messages and refusals is available through the interface if you want to be more thorough than that.

The refusal decision happens before the model is called at all. The assistant compares the best retrieved passage against a threshold the owner chose, and below it returns the fallback without spending anything. Above it, the model is called and the message counts as answered.

The owner picks that threshold from three options: Cautious at 0.5, Balanced at 0.35 which is the default, and Willing at 0.22. That means part of this percentage is a preference rather than an observation. Two assistants with identical material and identical traffic will show different figures if their owners chose differently, and neither is more correct.

Move the threshold to Willing. Questions that would have been refused now reach the model, the model answers from weaker material, and the figure rises immediately. The interface is direct about the trade in its own wording: fewer refusals, and more thin answers. Nothing about the material changed. The number moved because the bar moved.

The accidental version is more common and harder to spot. Publish one long, general document that touches every topic loosely and match scores rise across the board, because there is now always something vaguely relevant to retrieve. Refusals fall, the figure improves, and the answers quietly get worse because the passage that won was general rather than right.

There is a third route that nobody intends: traffic mix. A week of visitors asking the three things your material covers well will produce a better figure than a week of visitors asking about the new product nobody has written up yet. The assistant did not change. The questions did.