Methodology
How the numbers are made.
Written for the buyer who asks how, not whether. If a claim on this site has a number behind it, this page is where the number comes from.
One loop, five stages
A guest posts a review. Replio reads it, categorizes it, and drafts a reply in your voice. You tap to publish. The pattern behind it lands in the next morning’s Brief. What changed afterward is measured and written down. Every feature in the product belongs to exactly one of those stages.
How replies are drafted
Drafts are written from what the guest actually said and what happened in that store, in the voice your published replies already use. No reply reaches a platform without a human tap from your team, ever. There is no bulk-approve and no silent mode.
What counts as a pattern
Reviews are categorized into a fixed operational taxonomy: order accuracy, food quality, speed of service, hospitality, and their peers. A theme only counts as a pattern when at least three complaints land in the same category at the same store inside the measurement window. One bad night is a bad night; three of the same complaint is a pattern.
How outcomes are measured
Every play gets a receipt: the complaint rate in the window before the play against the window after, at 30, 60, and 90 days. Both windows need at least five reviews and the baseline at least three complaints in the category, or the receipt is marked insufficient and is never presented as a verdict. A win means the complaint rate dropped at least 25 percent. If it did not, the problem has recurred, and the play escalates instead of being celebrated.
The two human gates
The system automates reading, drafting, pattern detection, and measurement. It never automates two things: no guest reply publishes without a human tap, and no outcome is ever called verified unless the person who ran the play confirmed they ran it. Trust in the numbers starts with refusing to invent the human parts.
What we do not claim
Complaint rates move for reasons outside any single fix: seasons change, staff turns over, menus rotate. Receipts are operational evidence, not controlled experiments, and they are labeled that way wherever they appear. A same-period reference from our benchmark panel of public restaurant data provides context, not proof.
Auditability
Every published reply is attributable to the person who approved it. Outcome math is versioned: when a calculation changes, history is recomputed under a new version rather than silently rewritten. Snapshots are append-only. If a number on your screen changed, there is a record of why.
The thresholds above (three complaints, five reviews per window, a 25 percent drop, 30/60/90-day windows) are the live constants in the production system, not marketing choices. Data handling lives on the security page.