What each week held
Withheld: something was read in 1 week of the last 12. A series needs at least 2 to say anything a single reading does not.
What it declares
The reading listed none. That is what the reading listed, not a statement that this product carries nothing.
What users raised
- comment volume / nitpick flooding
- false positives eroding trust
- confidence scores perceived as unreliable
- reviewers get disabled or ignored ('cry wolf')
- real issues buried among noise
- cross-tool false-positive rate disputes
Themes people raised in public. A complaint is somebody’s opinion, not a measurement of the product.
Everything on record
3 observed claims, and 2 inferences counted apart from them. Newest first; the tag on each row is what kind of claim you are reading.
- DISCOURSEINFERENCESep 30, 2026
A cluster of 2026 blog posts and a GitHub Action example show the ecosystem response to bot-noise complaints converging on concrete mitigations: capped comment budgets (e.g., a default max of 5 comments ranked by severity), confidence/severity filtering, dedup and 'resolved' tracking, and in at least one case (Scanity) abandoning PR comments entirely in favor of failing the GitHub check to block merges above a severity threshold.
Indicates the noise complaint has crossed from niche annoyance to a recognized design problem prompting structural product changes (budgets, blocking checks) rather than just prompt tweaks.
radar.offseq.com/threat/we-noticed-most-ai-code-review-tools-just-comment-which-get-ignored-pretty-often-92ca9f714888e863 - DISCOURSEINFERENCESep 30, 2026
Multiple independent comparison sites converge on a benchmark claiming Greptile catches roughly 82% of bugs versus CodeRabbit's ~44-46%, but with Greptile generating roughly 11 false positives per run versus CodeRabbit's roughly 2, though one source notes Greptile's own primary benchmark page publishes no per-tool false-positive counts and states only that 'false positives, style suggestions, and unrelated comments did not affect the catch rate.'
The specific 11-vs-2 figure circulating widely across secondary blogs appears to be unverified/uncorroborated by the primary source, which is material for anyone citing it as fact.
particula.tech/blog/greptile-vs-coderabbit-vs-qodo-ai-code-review-2026 - DISCOURSECOMPANY CLAIMSep 30, 2026
Greptile's own blog frames signal control as 'non-negotiable,' stating that a reviewer that floods every PR with low-value comments will get turned off within a month, and that severity controls, comment-type filtering, and learning from feedback are table stakes.
Vendor acknowledgment that noise is an existential product risk, shaping how review-bot vendors now market restraint as a feature.
www.greptile.com/blog/ai-code-review - DISCOURSEUSER OPINIONSep 30, 2026
A commenter who switched from CodeRabbit to a different reviewer reported CodeRabbit had 'worked really well and caught a lot of bugs / implementation gotchas,' and its 'learnings' feature meant it avoided repeating comments on intentional codebase patterns — finding CodeRabbit was if anything too quiet rather than too noisy.
Shows a conflicting, positive independent account of CodeRabbit's noise level, indicating the complaint theme is not universal and experiences vary by team/config.
news.ycombinator.com/item?id=46766961 - DISCOURSEUSER OPINIONSep 30, 2026
A developer shopping for PR review bots reported CodeRabbit added too much noise to PRs with only a small percentage of comments being useful, and that instructing it to ignore nitpicks did not stop the nitpick comments.
Establishes a first-hand complaint sighting specifically naming CodeRabbit as noisy, contradicting CodeRabbit's low-noise positioning elsewhere.
news.ycombinator.com/item?id=42451968
A record of what Forge read in public about Noise from review bots — its own pages and what people wrote about it. Nothing here is a test of the product, a ranking or a score, and Forge has no relationship with this company.