This one belongs to what you wrote, not to what answers

Whether it is written down at all. A policy that lives in the head of the person who has been there longest cannot be covered by anything. This is the largest gap in most small businesses and it is the one nobody thinks of as a software problem, because it is not one.

Whether a source can reach it. Material behind a login, in a folder nobody connected, inside an email thread, or on a page nothing links to is outside every crawl and every connection. It exists, it is written down, and it is not coverable in its current location.

Whether there was room. A crawl takes up to the pages your plan allows, less whatever is already indexed, and stops. A large site under a modest allowance is covered in part, and which part depends on where the crawl got to rather than on what mattered.

Whether a retriever can find the answer inside what was taken in. A sentence about refunds buried in a nine thousand word general page is indexed and effectively uncovered, because the passage that wins a match is about six things and commits to none of them.

The number of documents indexed is shown per source, and it is a count of what was taken in. It says nothing about what fraction of your visitors' questions has an answer somewhere inside them, and the two can move in opposite directions.

Adding a large undifferentiated document set raises the count and can lower real coverage, because everything now matches everything weakly. A short set of focused pages, one topic each, can cover far more of what is asked with a tenth of the volume. If you are going to watch one thing about your material, watch whether each document answers a question somebody would ask.

The refused questions are the honest part. Each one is a question that arrived and found nothing close enough, with a figure showing how near the best passage came, and together they describe your coverage against the traffic you actually received.

The share of answers that were not refusals is the same information as a level rather than a list. Read as coverage it is badly behaved: it moves with the mix of questions in a given week, and it moves with the threshold setting, neither of which is a fact about your documents.

The measurement that works is one you construct. Write down the twenty things a customer most often needs to know, ask each in the widget in a customer's words, and record whether the answer was right and which document it cited. That is coverage measured against a denominator you chose deliberately rather than one your traffic chose for you, and it is repeatable next quarter.

Upload everything. The document count climbs, the sources screen looks substantial, and matches get vaguer across the board because there is now always something loosely relevant. Refusals fall, which reads as coverage, and the answers that replaced them are built on the least specific material you own.

Report one minus the refusal share and call it coverage. It is a figure about your visitors rather than your documents: an easy week reads as good coverage and a week where somebody linked to you from a new place reads as bad. Nothing about the material moved either time.

Loosen the threshold and the same questions that were uncovered yesterday are covered today. This is the cleanest demonstration that the figure people use as coverage is a setting rather than an observation.