Chromium Open Source Health - Value Exchange

What's counted. Review comments on chromium/src CLs targeting main, each in the week it was written:

• on merged, abandoned and still-open CLs (review that stops a CL landing counts too);
• only on human-owned CLs, so autorolls and other bot CLs are left out;
• not from bot or AI-agent accounts, nor owners on their own CLs (a reviewer's AI-assisted comments count as theirs);
• only on the paths listed under Contribution by directory.

Votes with no comment, such as a bare +1, don't count: they don't show how much review went into them.

Commits are the other side: commits on main (from Chromium's git history) that touch at least one of those paths, each in the week it landed and credited to its git author. Commits of bot-owned CLs, such as autorolls, are left out.

Organization. Each reviewer's organization as of the comment's date (for reviewed CLs, the close date), and each commit author's as of the day the commit landed, by the first rule that applies:

1. their affiliation in Chromium's affiliations.json5, which credits @chromium.org and personal addresses to employers and follows job changes;
2. other @chromium.org accounts count as google.com;
3. otherwise their email domain (fb.com counts as meta.com).

Left out of usage shares. This chart shows them, but the value ratio and the table don't, so each organization's share is of the remaining Chromium usage:

• Android System WebView (ChromeMobileWebview): embedded by many browsers and apps, so it can't be credited to one organization.
• ChromeDerivative, Radar's bucket for Chromium browsers it can't name. It jumped from about 0.3% to about 10% of Chromium usage on 9 September 2026, taken from every browser in proportion (Safari and Firefox too), which looks like a labelling change at Radar. Cloudflare hasn't explained it.

In-app browsers. Facebook and Instagram count for Meta, Microsoft 365 for Microsoft. Others (Snapchat, Pinterest…) stay in the total but count for no organization.

iOS is excluded, since every iOS browser must use WebKit. Radar can't filter it out, so each day's iOS traffic is subtracted using Radar's iOS-only browser breakdown and iOS share of requests.

Brave's usage is unknown. Brave sends Chrome's User-Agent and names itself only in the Sec-CH-UA client hint, so Radar reports it at about 0.002% of Chromium usage, roughly a thousand times too low; the rest is most likely counted as Chrome. brave.com gets no usage share or ratio. With Brave's real share moved out of Chrome, Google's usage share would be a few points lower and its ratio slightly higher.

A thread is a reviewer's comment and every reply to it.

• Addressed: adopting all or part of a suggestion, a different change because of it, a TODO, moving code out of the CL, or agreeing to make the change (“I'll rephrase this”).
• Not addressed: the owner answered, explained or pushed back; the comment was an approval, thanks or FYI; the work went to another CL or a bug; or the reply was a bare “Acknowledged”. This isn't bad review (questions are part of good review), just review that didn't change the code.

How it's decided, first rule that applies:

1. The owner started the thread or never replied: not addressed.
2. The owner's first reply starts with “Done”, “Fixed”, “Fix applied.” or similar: addressed.
3. Otherwise Claude Haiku reads the whole thread and decides.

Scoring. Every comment in an addressed thread that isn't the owner's, up to the owner's last reply, earns its author one point, so back-and-forth and a second reviewer joining in earn more.

Accuracy. Checked on 800 random threads with an owner reply (10 Aug – 1 Nov 2026), labelled blind by Claude Opus with disagreements settled by hand. The regex plus Haiku has a precision of 0.98 (98% of threads it counts as addressed were) and a recall of 0.99 (it finds 99% of addressed threads), each within about ±1 point. The regex settles 60% of threads (precision 0.98); Haiku the rest (precision 0.97, recall 0.96). 4 threads too unclear to call were left out. These are per-thread figures; points are per comment, so long threads weigh more. Details: docs/comment-classifier.md.

Placing comments on a file. Gerrit doesn't give us each CL's file list, so:

• an inline comment counts for its own file;
• a patchset-level comment (a reviewer's reply summary) counts for the file that reviewer commented on most in the CL;
• if the reviewer left no inline comment on the CL, it shows as (no file) above and is left out of the other charts. Accounts that have never left an inline comment are left out entirely.

Unplaced comments are mostly sign-offs, not review: in a sample, 39% of those on CLs with no inline comments were just “LGTM”, “thanks” or similar (against 2% of inline comments), and most of the rest were OWNERS sign-offs (“mojom lgtm”), requests to route the CL elsewhere, or spam. Leaving them out barely changes any organization's share.