Metrics Methodology Update: What's Changing and Why

Updated: July 14, 2026

Summary

We're rolling out a set of refinements to how several engineering metrics are calculated. These updates keep our measurements accurate as AI-assisted development matures, and they respond directly to customer feedback asking for more precision.

This is a normal part of how we maintain the product. Your historical data is preserved, your dashboards work the same way, and nothing about how you use the product changes. What changes is that some metrics get more accurate, and a few will read differently as a result.

Most metrics in the catalog are unchanged. This article covers the ones that are moving, grouped by the improvement behind each.

The improvements, by theme

1. More precise AI attribution

A commit used to count as either "AI-assisted" or "not," based on a single threshold. We now measure how much AI contributed to each commit and count it proportionally. That reflects reality more faithfully: a commit where AI did a little and one where it did most no longer look identical.

Affected metrics:

• Activity Percentages (AI): each commit now contributes to the AI share in proportion to how much AI helped, instead of all-or-nothing.

• Activity Percentages (Coding): the mirror of the above. The non-AI share is now the proportional remainder.

2. Fairer, business-day-aware activity

Activity metrics can now normalize over the working week instead of the full calendar week, so a healthy Monday-to-Friday cadence isn't diluted by weekend days that were never expected to carry activity.

Affected metrics:

• Active Days Per Week: when business-day mode is enabled, the score normalizes over weekdays (out of 5) rather than 7 calendar days. Weekend commits still count as active days; only the denominator changes.

3. Cleaner human signal (consistent bot exclusion)

People-focused metrics should reflect people. We now exclude bot and automation accounts consistently, using our full bot list rather than only catching accounts with a [bot] name suffix.

Affected metrics:

• Activity Percentages (Reviewing): only human review comments are counted.

• Semantic Review Depth: bot-authored review comments are reliably excluded from both numerator and denominator.

• PRs Merged Per Developer: bot-authored pull requests are excluded from throughput.

4. More stable delivery timing (median instead of mean)

Delivery-speed metrics now use the median rather than the mean. The median resists outliers, so a small number of unusually slow pull requests no longer distorts the headline number. This was one of the most-requested precision improvements.

Affected metrics:

• Cycle Time: median days from PR creation to merge (the mean is retained only as a secondary hint).

• Queue Time: median age of open PRs.

5. Clearer review-load metrics

Review-load percentages now measure each reviewer against their own reviewed volume, rather than against all merged code in the window. The practical result: for any reviewer, AI Review Load and Non-AI Review Load now add up to 100%, which makes both far easier to read.

Affected metrics:

• AI Review Load

• Non-AI Review Load

6. Open-PR metrics now reflect all open work

Metrics describing the open-PR backlog used to include only PRs that carried AI-usage data, which could understate the true backlog. They now include every open pull request as of the measurement date.

Affected metrics:

• Open PRs Age (Median)

• Open PRs Age (P90)

• Queue Time (also benefits from this change)

7. Consistent measurement windows

Contribution Stability now uses the report's own window start and end dates consistently, and drops incomplete early windows, so the score is comparable across reports.

8. Retiring a few legacy metrics

A small group of metrics compared current activity to a "pre-AI baseline." That baseline means less and less as AI-assisted development becomes standard, so these are being retired and will no longer appear in the UI:

• AI Adoption

• Merge Uplift

• Quality Uplift

• Velocity Uplift

If any of these are load-bearing in your reporting, get in touch. We can discuss alternatives or custom equivalents.

Notable changes at a glance

The table below highlights the changes most likely to be visible in your dashboards. Directions reflect typical impact; the exact effect depends on your data.

Unchanged: the majority of metrics are not affected by this update, including AI Code Quality, AI Code Contribution, Qualitative Code Score, AI Usage %, AI Usage % by LOC, Batch Size, PR Size, Commits per Developer per Week, and the AI usage and volume family.

Frequently asked questions

Will my historical data change? Your stored history is preserved. Metrics recalculated under the new methodology reflect the improved definitions going forward, and we'll make the transition point clear in-product.

Do I need to do anything? No. The updates apply automatically as they roll out.

Why do some numbers look different now? The calculation got more accurate. Where a metric moves, it's because we removed a distortion (an outlier, a bot, a coarse threshold), so your team's underlying performance is the same.

Can I keep a metric that's being retired? Reach out. For enterprise customers we can often provide a tailored equivalent through custom tuning (below).

Enterprise custom tuning

These improvements also make our custom-tuning service more capable. If you're an enterprise customer, we can calibrate your metrics to the KPIs that matter most in your specific SDLC, adjusting thresholds, weightings, and definitions to match how your teams actually plan, review, and ship.

If that's useful, reply to the announcement email, contact your account manager, or email support@exceeds.ai, and we'll set up time to work through it with you.