I just finished analyzing 802,979 merged pull requests from GitHub Archive. The findings... aren't great.
The headline stat that haunts me
90% of PRs over 1,000 lines ship without any code review.
Not "light review." Not "quick approval." Zero approvals, zero change requests, zero review comments. Nothing.
And it's getting worse - up from 83% in 2024.
The self-merge reality
71% of all code on GitHub ships without a second pair of eyes. The author writes it, the author merges it. Done.
I'm not saying self-merging is always bad (solo projects exist), but this is all of GitHub. The "code review culture" many teams claim to have is increasingly fiction.
The size paradox broke my brain
You'd think bigger PRs would get more scrutiny. More code = more risk, right?
The data says the opposite:
| PR Size | Review comments per 100 lines |
|---|---|
| Tiny (<10 lines) | 0.98 |
| Massive (1000+ lines) | 0.05 |
That's 20x less scrutiny per line for the biggest changes. Reviewers see a wall of code and just... approve.
The first-timer tax
New contributors to a repo wait 38% longer to get merged than repeat contributors (22h vs 16h median).
Good news: this improved from 53% in 2024. Teams are getting better at onboarding. But the gap is still real.
Wednesday is the new Monday
This surprised me. In 2024, Monday was peak merge day (the "clear the backlog" effect). In 2025, Wednesday dominates at 23.5%.
Teams seem to be moving toward more distributed workflows rather than the Monday morning merge rush.
The bot collapse
Remember when Dependabot PRs were everywhere?
Bot PRs went from 62% of all PRs in 2022 to just 15.5% in 2025. The automation boom is over. Teams got tired of the noise.
What actually works (from the data)
For teams that do practice code review, the median cycle time is 3 hours. That's the benchmark to aim for.
Things that correlate with healthier review practices:
- Smaller PRs (the data is unambiguous here)
- Branch protection rules requiring approval
- CODEOWNERS files for domain expertise
The full research
I've published the complete methodology, interactive charts, and YoY comparisons here: 2025 Engineering Benchmarks: Year in Review
Includes breakdowns by language, a spotlight on how AI tool repos (Codex, Gemini CLI, Claude Code) ship code, and the raw numbers for benchmarking your team.
What's your team's self-merge rate? Curious if these numbers match what you're seeing.
Top comments (0)