"PRs take forever to merge" is a complaint, not a diagnosis. Since September 25, GitHub breaks PR review time into three stages, each with a median and a p90.
The repository-level Copilot usage metrics reports (enterprise and organization, repos-1-day) gained a pull_request_review_times array: for each day, a median and a 90th percentile for three stages. From "ready for review" to first review. From first review to final review. From final review to merge. Only PRs opened by a human and reviewed by at least one other human qualify; Copilot code review and bot reviews are not timed (a PR reviewed by both a human and Copilot still counts). There is no backfill: data accumulates from September 21, 2026 onward.
The value lies in the decomposition, not in one more number. Three stages are three different diagnoses, each with its own cure. A long "waiting for the first look" is a response-norm and routing problem: the PR cannot find its reviewer. A long "first to final" points at PR size and the count of required reviewers: the change is drowning in revision ping-pong. A long "approved to merge" is about automerge and permissions: the work is done, the change just sits there. While these three waits were fused into one "time to merge," every improvement was a ritual performed on a hunch. And the p90 next to the median separates systemic slowness from the tail: if the median is calm but the p90 climbs, the process is fine — a handful of slow PRs is skewing the picture, and those deserve a case-by-case look, not a process overhaul.
What this means for a team lead. First, turn the collection on now: there is no backfill, and every week of hesitation is a week of history lost. Second, spend the first week just watching: build a baseline for your repositories, no conclusions, no org decisions. Third, define your red zone against your own baseline, not someone else's benchmarks: a monorepo and a fleet of small services have different norms. Above all: these metrics are a thermometer for the process, not a per-person KPI. The moment a stage median becomes a reviewer's scorecard, people start optimizing the number instead of the flow — and the thermometer stops telling the truth. Measure the system, fix the system; conversations about people belong to Concordius, not to a dashboard.
📎 github.blog/changelog/2026-09-25-usage-metrics-api-adds-pull-request-review-stages