Review moderation that catches health claims before they publish

In this retailer’s market, a single published review that says a product treats, cures or relieves a condition can create legal risk. A review that reports a side effect has to reach the right team quickly.

General review apps publish by star rating or by hand. Reading every submission by hand costs staff time, and the review app the retailer used could not tell which reviews were risky.

1. 01Two passes. A fast keyword and pattern check runs first and marks hits. Then a language model reads the review and returns a recommendation, a confidence level, the quoted text behind each flag, and a short reason.
2. 02If the pattern check finds a health claim, the review cannot publish on its own, even when the model says "publish".
3. 03If the model cannot be reached or returns a bad answer, the review goes to a person instead of publishing.
4. 04Reviews are processed in the background, so a burst of submissions does not slow anything down. Older reviews can be checked again when the policy changes.

The two-pass check, the queues, the alert routing and the automation levels carry over to other stores.

The rule set does not. Which claims are blocked and what counts as a side effect were written for this retailer’s market. Another seller of regulated or claim-sensitive products needs its own rules.