A developer studies a large software analytics dashboard with charts, status indicators, and code review panels.
GitHub has added a "Repository insights" page that shows what happens to Copilot code review feedback after it is posted. It is the first item under a repository's Insights tab. It tells you whether developers act on Copilot's comments, whether humans still ask for changes afterward, and which merged pull requests were never reviewed by Copilot. It is a measurement tool, not a quality guarantee, and GitHub's documentation is blunt about that.

A developer studies a large software analytics dashboard with charts, status indicators, and code review panels. What shipped​

Open the Insights tab of any repository you have read access to, then click Repository insights. The page is built for organization-owned repositories. The Copilot code review section requires the organization to have enabled both the Copilot usage metrics policy and the Copilot code review policy.

The launch notice lists four summary signals:

  • Comments addressed: Copilot comments that were acted on.
  • High-severity comments addressed: the same measure, limited to high-severity feedback.
  • Pull requests with no later human change request: Copilot-reviewed pull requests with no human "request changes" after Copilot's review.
  • Merged pull requests without Copilot review: merged work that Copilot never reviewed.

You can also dig into the addressed rate, with its numerator, denominator, unevaluable exclusions, trend and breakdowns. A comment categories view shows the types of feedback Copilot leaves.

The launch coincided with a wider Copilot code review update. GitHub's changelog bundle also covered Adaptive review effort, secret detection in public preview and Azure DevOps support. Repository insights is one piece of that bundle.

How "addressed" is decided​

This is the part to understand before you trust any percentage. GitHub's documentation says an AI model evaluates each published Copilot comment when the pull request merges. The model judges whether the issue was addressed. A comment counts as addressed if the developer did one of the following:

  • Applied a suggestion.
  • Asked Copilot to fix it.
  • Pushed a manual or agent-made fix.
  • Accepted the feedback and explicitly deferred the fix to a later pull request.

Unaddressed comments stay in the evaluated total but not the addressed total. This is broader than the older suggestion-application counts, which only track ready-to-apply suggestion blocks applied with the button. As GitHub's November 5 API changelog notes, developers act on most feedback in other ways than clicking apply.

Because an AI model makes the judgment, it is a classification, not a log of developer actions. GitHub's own guidance says a low rate is a reason to inspect examples. It is not proof that the feedback was unhelpful.

Reading the four cards correctly​

The four cards use different populations, so don't compare them directly.

SignalCalculationCaveat
Overall address rateAddressed ÷ evaluated commentsA low rate means inspect examples, not "Copilot is wrong"
High-severity address rateAddressed ÷ evaluated high-severity commentsCheck the count; small samples mislead
No human follow-up requestMerged Copilot-reviewed PRs with no human change request after Copilot's first review, ÷ all merged Copilot-reviewed PRsNot a measure of defect-free code
Merged without Copilot reviewAll merged PRs minus Copilot-reviewed merged PRs, ÷ all merged PRsCheck whether those PRs should have been reviewed

The "no human follow-up" rule is strict. A human change request made after Copilot's first review excludes the pull request, even if the request was later dismissed or the pull request was approved. A request made before Copilot's first review does not exclude it.

The coverage card has its own trap. The "Automatic review is off" message describes configuration. It does not explain every uncovered pull request, and manual Copilot reviews still count toward coverage.

Breakdowns and what they can't tell you​

The address rate can be split by:

  • Severity: High, Medium, Low.
  • Review effort: Max, Balanced, Lite. Adaptive-routed reviews are counted under whichever level Adaptive chose.
  • Customization: custom instructions, agent skills, MCP servers, or none.

Each rate is calculated within its own group. The customization groups overlap, so one comment that used custom instructions and an MCP server appears in both. Don't add those groups together. Some comments lack a recorded severity or effort, so breakdown totals may not match the overall figure.

GitHub also warns that differences between effort levels or customization groups are descriptive. They don't prove that more effort or more configuration causes a higher address rate or better code. If one repository shows MCP-assisted comments being addressed more often, that is a lead to investigate, not a reason to change policy.

The category chart ranks the number of addressed comments per category, not the address rate within each category. A tall "Documentation" bar doesn't mean documentation comments are more likely to be addressed. Examples of categories include Bug, Documentation, Security, Concurrency and Performance.

Time windows and data gaps​

  • The weekly cards cover seven calendar days ending on the latest fully published day. Check the dates shown on the page.
  • Outcomes are attributed to the day the pull request merged, not the day the comment was posted.
  • The monthly trend compares effort levels across up to six completed months. Six is a maximum, not a promise.
  • On GitHub.com, trend history begins in August 2026. Effective review effort has only been recorded since August 26, so August reflects requested effort.
  • Enterprises with data residency have coverage starting in October 2026. Trends are withheld until at least two completed months exist.

The documentation also separates missing data from zero activity. A "waiting" or stale indicator means the data hasn't arrived. A rate with no evaluated comments has no denominator, so it is not a measured 0%.

The page also shows pull request throughput and median time to merge for context. Opened and merged pull requests are separate populations, not a funnel. Daily medians shouldn't be averaged into a weekly median.

The API route for admins​

The same outcomes are in the daily repository reports (repos-1-day) of the Copilot usage metrics API. Both organization and enterprise scopes have them. According to GitHub's documentation, these reports hold one record per repository with pull request activity that day. Repositories with no activity are omitted. The copilot_code_review object holds the comment outcome counts and breakdowns.

For your own reporting, GitHub's guidance is:

  1. Sum addressed and evaluated counts across the days you want, then divide. Don't average daily percentages.
  2. For the no-follow-up share, divide total_merged_no_human_change_request_after_review by pull_requests.total_merged_reviewed_by_copilot.
  3. For the number merged without Copilot review, subtract pull_requests.total_merged_reviewed_by_copilot from pull_requests.total_merged.
  4. Treat a zero evaluated count as "no denominator", not a 0% rate.

Access to the reports goes to enterprise owners and billing managers, organization owners, and anyone with a custom organization or enterprise role that grants the View Copilot Metrics permission. The Copilot usage metrics policy must be enabled. For OAuth app tokens and classic personal access tokens, the REST documentation lists read:org for the organization endpoint, and manage_billing:copilot or read:enterprise for the enterprise endpoint.

Repository-level reports only count comments Copilot posts on GitHub pull requests. Comments from people or other tools are excluded, as are Copilot reviews done in the IDE.

Analysis: useful, with guardrails​

Most Copilot dashboards until now measured activity: seats, active users, suggestion clicks. This one tries to measure outcomes per repository. For administrators who have turned on automatic review everywhere and want to know whether it matters, that is a real step forward. The "merged without Copilot review" card alone is a quick way to find repositories where enablement quietly never happened.

The caveats matter, though. The address rate depends on an AI model's reading of what a developer did. "No later human change request" is a proxy that a small team with informal review habits could distort. And GitHub says plainly that none of these numbers show that Copilot caused better outcomes. Copilot code review also always leaves a "Comment" review rather than an approval, so human reviewers remain the gate. (That is the behavior third-party guides describe. GitHub has since added an approvals preview, so check your settings.)

A sensible way to use it:

  1. Start with the coverage card to find repositories with many unreviewed merges.
  2. Check whether those gaps are configuration problems or intentional exclusions.
  3. Sample unaddressed high-severity comments in repositories with low rates.
  4. Only then adjust effort levels, instructions or skills, and watch the monthly trend.

Treat the dashboard as a prompt for questions, not a scorecard to put in front of developers.

 

References

  1. Track Copilot code review outcomes in Repository insights GitHub Changelog 2026-11-17T18:00:35+00:00
  2. Repository insights metrics - GitHub Docs docs.github.com
  3. Data available in Copilot usage metrics - GitHub Docs docs.github.com