AI Transparency Policy

This page explains, in detail, how the AI engine that calculates the AI Objective Score works — because we believe an automated rating is only trustworthy if how it works can be audited.

1. What the AI never does

The AI does not edit, summarize, censor, or hide a review's original text. What a customer writes is published in full. The only output the AI produces is a structured score calculated from that text.

2. How the score is calculated

When a review is submitted, the text is analyzed with a language model (OpenAI's GPT-4o mini) instructed to independently evaluate three dimensions, each from 1.0 to 5.0:

• Product: quality of the core product or service.
• Service: treatment, responsiveness, and resolution of customer service.
• Shipping: whether delivery timelines and conditions were met.

The model also receives the business's category (e.g. restaurant, fashion, electronics) and a short description, so it can interpret each dimension with criteria suited to that type of business.

The final score uses a fixed weighting: 40% Product + 30% Service + 30% Shipping. The goal is that a single-dimension problem (e.g. a late shipment) doesn't erase a positive experience in the others.

If a review doesn't mention one of the three dimensions, that dimension isn't given an invented score: it's left unrated and excluded from the calculation, which reweights across only the dimensions the review actually describes.

3. Penalty for unresolved recurring issues

If the same problem repeats across several of a business's reviews (e.g. delays with the same shipping carrier) and the business doesn't resolve it within 30 days, new reviews mentioning that problem receive an additional penalty to their score. This exists so the rating stays honest with customers when a known problem goes uncorrected.

4. Appeals and human review

No AI model is perfect. If a business believes a review is fake, defamatory, or breaks the rules, it can appeal by attaching evidence. A human reviewer evaluates each appeal and can only make one of two decisions, always binary: remove the review (if the evidence confirms it's fake or defamatory) or reject the appeal, in which case the review stays exactly as it is. There is no third option to manually adjust or correct a score — the score always comes from the same automated calculation based on the text, never from a case-by-case human decision.

5. Known limitations

• The model can misinterpret sarcasm, local slang, or ambiguous text.
• Very short or vague reviews give the model less information to infer accurately.
• The score is an automated estimate, not a legal judgment or an independent fact-check.

6. Auditability

Every published review shows the per-dimension breakdown for whichever dimensions were actually scored, visible both in the business's dashboard and in the public widget — anyone can see how the final score was reached. On each business's public reviews page, the "Corrected" filter shows exactly which reviews have a real gap between what the customer rated and what the AI read, so anyone can compare the score against the original text and draw their own conclusions.