Resources · 85
Customer reviews: provenance, moderation and consistent ratings
Explain verified-review labels and keep moderation fair from submission to structured data.
· 4 min
What this guide helps achieve
- Define what verification means
- Link reviews to the right item
- Moderate against explicit reasons
- Recalculate displayed ratings
- Investigate without premature conclusions
Quick check
- Should a negative review be removed?
- Can variants share reviews?
- Does automated detection prove a fake review?
Step-by-step method
- 01
Define what verification means
Distinguish a linked purchase, declared use and an unrestricted submission. In France, claiming reviews come from purchasers or users requires checking measures; ministry guidance also covers transparency about dates, ordering, incentives and moderation. Describe what your checks actually establish without claiming absolute assurance.
Deliverable: provenance labels and applicable rules.
- 02
Link reviews to the right item
Retain a stable product or service identifier. Explain whether reviews are shared across variants and why. During migration, check origin, experience date and publication date; do not present imported reviews as locally collected.
Deliverable: review-to-item and origin mapping.
- 03
Moderate against explicit reasons
Separate negative opinion, abusive content, personal information and off-topic submissions. Apply the same rules regardless of rating. Retain rejection reason, decision, date and available appeal. A generated summary must not become a new review attributed to a customer.
Deliverable: moderation rules and decision register.
- 04
Recalculate displayed ratings
Define scale, included population, rounding and withdrawn-review handling. Reconcile counts, visible averages and structured data for the same item. Follow Review snippet requirements; correct markup does not guarantee a rich result.
Deliverable: reproducible calculation and markup check.
- 05
Investigate without premature conclusions
Grouped submissions, similar wording or rating shifts may warrant investigation; alone they do not prove fraud. Examine campaign invitations or genuine incidents before deleting content. Let readers access ordering and verification rules.
Deliverable: signal-investigation procedure.
Acceptance case to reproduce
Fictional example: these inputs describe no customer or observed result.
View case inputs
{
"ratings": [
{
"value": 5,
"min": 0,
"max": 5
},
{
"value": 8,
"min": 0,
"max": 10
}
],
"raw_mean": 6.5,
"normalized_mean_out_of_5": 4.5,
"semantic_equivalence_verified": false
}Expected decision
6.5 mixes rating scales. Compare their endpoints and meaning before publishing a combined average.
Fictional variations to test
40 ratings average 4.5 and two ratings average 3.
With comparable scale, item and period, the weighted mean is (40 × 4.5 + 2 × 3)/42 ≈ 4.43, not 3.75.
A product has not received any rating.
Show that no rating exists. Zero reviews is not a zero rating; a mean without observations is undefined.
One person corrects their archived rating.
Following the publication rule, replace that review’s current state and retain its revision trail; do not count two votes.
Evidence to retain
- Scale, item, period and compared counts.
- Review identifier and revision rule.
- Weighted sum, denominator and missing rating.
Verification procedure
- Record scale endpoints, provenance and moderation rules.
- Compare ratings expressed on different scales.
- Keep normalization and its semantic limitations.
Ministère de l’Économie — Peut-on faire confiance aux avis en ligne ?
Reusable worksheet
Complete with your authorised observations. These fields are a working template, not observed results.
| Field | Information to record |
|---|---|
| Origin | Collection channel, item, experience and publication dates |
| Verification | Check performed and exact meaning of the verified label |
| Moderation | Rule, reason, decision and appeal route |
| Aggregate | Included population, count, average, rounding and markup |
Worked example
Illustrative situation
Fictional example: a store combines two colors and a new technical product version under one rating.
Decision and expected evidence
It retains color-level sharing where relevant but separates the technical version and identifies imported review origins.
Distinguish the mechanisms
| Mechanism | Purpose | Check or limitation |
|---|---|---|
| Linked purchase | Connect a review to a transaction | Does not prove every statement is accurate |
| Moderation | Apply publication rules | Must not select only favorable opinions |
| Average rating | Summarize a review population | Conceals spread, age and exclusions |
Management indicators
| Indicator | What it measures | First action |
|---|---|---|
| Rating consistency | Matching interface and structured-data aggregates | Correct item or calculation population |
| Justified decisions | Rejections linked to explicit rules | Reconsider decisions lacking a reason |
| Traceable imports | Reviews retaining origins and dates | Relabel unsupported verification claims |
Common pitfalls
- Does not prove every statement is accurate
- Must not select only favorable opinions
- Conceals spread, age and exclusions
Frequently asked questions
Should a negative review be removed?
Its rating alone is insufficient. Examine publication rules and handle the complaint without manufacturing an impression of universal satisfaction.
Can variants share reviews?
Define the scope and inform readers. A different characteristic may make a review less relevant to another variant.
Does automated detection prove a fake review?
No. Use signals for documented review with a route to correct mistaken decisions.
Official references
References consulted: . The method and worksheet propose checks to adapt to your context; they do not constitute certification.
