proof · the honest metrics manifesto
Honest numbers only.
Retention software grades its own homework. These are the seven rules that stop us.
Every number wears its provenance.
Backtest, holdout, demo store, or a named merchant — the origin travels with the figure, in the same breath, with a link to the method. A number that can't say where it came from doesn't get published.
The floor is always subtracted.
About ~2.7% of lapsed customers come back with no outreach at all — software that counts them is billing you for the tide coming in. We measure the floor first and claim only what clears it.
Observational is not causal.
A tool that watched a sale happen did not cause the sale. Causal credit requires a control group that got nothing — we label which is which, everywhere, including inside the product.
The machine drafts. You decide.
Everything our AI writes appears in violet — the mark that the machine wrote it — and nothing sends until you approve it. An engine that acts without asking gets to invent its own results.
Smaller-but-real beats bigger-but-imaginary.
Our numbers will usually look modest next to the category's. A small figure you can defend in a board meeting is worth more than a large one that dissolves under a single honest question.
Real screenshots or none.
Every interface shown on this site is the actual product on labeled demo-store data. The first quote we publish will be a named human and one holdout-verified number.
The method is public — and allowed to improve.
How we measure is published, versioned, and corrected in the open. An honest number is one you can check.
If we ever publish a number that breaks these rules, hold us to this page.
— SK, founder, Retrics
retrics.ai · july 2026