Suppose a weekly report shows twelve mentions from fifty scheduled runs. Only forty runs actually returned an answer. If you quote 24% to the marketing director, you have quietly decided what ten unread answers said. Before acting on a drop, check the collection record as well as the chart.
Report both outcomes
In this example, completion is 40/50, or 80 percent. Presence among observed answers is 12/40, or 30 percent. Keep a valid answer that names no brands in the denominator. Separate an actual refusal from a network failure, and document whether refusals count for the business question you are measuring.
Look for patterned gaps before comparing
Check whether failures came from one engine, region or question type. If all ten were long comparison questions, the remaining forty are a poor proxy for the full batch. Retry under the same conditions, keep the original run IDs and replace missing observations only once. A retry must not earn an extra vote simply because it succeeded.
Give the decision a range
In this example the missing answers could move the final rate anywhere from 12/50 to 22/50: 24% to 44%. A decision that depends on clearing 35% cannot be made yet. If the missing runs are all pricing comparisons, recover those first; asking ten easier brand-name questions does not repair the sample.
