A detailed, fact-checked Fashion Nova case analysis with controls for review intake, moderation, publication, ratings, exceptions and audit records.
Short answer: answer: The Fashion Nova case shows that review governance must preserve the distribution of opinions received, not merely prevent fabricated reviews. Automatically publish positive and negative reviews under the same content-neutral rules, record every rejection reason, reconcile received versus published counts weekly, and stop displaying aggregate ratings when the missing-review population could materially change them. A vendor’s default workflow never transfers accountability away from the retailer.
The easy lesson is “do not delete bad reviews.” The operational problem is broader. Bias can enter at intake, automated publication thresholds, manual queues, fraud rules, syndication, rating calculations or staff capacity. A company can present only authentic reviews and still create a misleading picture if critical opinions wait indefinitely while praise appears immediately.
CDM’s position is that review completeness is a marketing claim even when no sentence says “we publish every review.” A familiar star display invites buyers to treat the visible sample as representative. If the system systematically excludes a class of genuine experience, the evidence shown no longer represents the evidence received.
What the FTC record establishes
In January 2022, the FTC alleged that Fashion Nova used a third-party review-management interface configured to automatically publish four- and five-star reviews while holding lower-starred reviews for approval. The complaint said that from late 2015 until November 2019 the company did not approve or publish hundreds of thousands of lower-starred reviews.
The Commission finalised its order in March 2022. Fashion Nova agreed to pay $4.2 million and was prohibited from suppressing customer reviews or misrepresenting that displayed reviews reflected all purchasers who submitted them, were unedited, appeared regardless of rating, or were incorporated into aggregate scores in a way different from reality. The order also imposed reporting and recordkeeping obligations.
Fashion Nova disputed the FTC’s characterisation. A company spokesperson told Time that the allegations were inaccurate, attributed the issue to a third-party platform’s auto-publish option and said the manual-release process was not completed during rapid growth. That response is relevant because it presents the company’s position; it does not nullify the final consent order. The matter was settled rather than tried, so regulatory allegations should not be presented as judicial findings.
The Review Integrity Checkpoint
The Review Integrity Checkpoint is a six-control chain. Each control must produce an auditable record. If a team cannot reconcile the chain, it should remove or qualify the aggregate score until it can.
Checkpoint 1: intake completeness
Give every submitted review a unique identifier, receipt timestamp, product identifier, rating, source and status before moderation. Preserve the original submission separately from the publishable version. Monitor broken forms, API failures and syndication delays, because selective loss is not the only way a review population becomes distorted.
The key reconciliation is simple: opening backlog + reviews received = reviews decided + closing backlog. Break it down by star rating, source, language and product. A total that balances can still hide a one-star queue that never moves.
Checkpoint 2: content-neutral moderation
Write allowed rejection reasons around content, not sentiment: spam, no relevant product experience, personal data, threats, illegal content or a clearly disclosed length rule. Apply the same rule to praise and criticism. “Brand damaging,” “unhelpful” and “customer-service issue” are not neutral reasons merely because they appear in a dropdown.
Keep a reason code, reviewer, timestamp and before-and-after text for edits. If the system permits only stars and no text after removing prohibited content, define that treatment consistently.
Checkpoint 3: equal publication timing
Set one service level for all ratings—for example, 95% of ordinary reviews decided within two business days. CDM offers this as an operating example, not an FTC requirement. Measure the 50th and 95th percentile by rating; an identical average can conceal a long negative-review tail.
Automatic publication can reduce bias if it applies to every rating and exceptions enter the same moderation queue. Auto-publishing only high ratings while low ratings await human action creates a structural asymmetry even before anyone intentionally rejects a review.
Checkpoint 4: representative rating maths
Document which reviews enter the displayed average, how syndicated reviews are handled, whether verified-purchase status changes weighting, and what happens after returns or product variants merge. Recalculate from the underlying published records and compare the displayed value.
Also compute a shadow rating from all eligible received reviews. If the visible and shadow values diverge beyond a predeclared tolerance, pause the aggregate or disclose the limitation while investigating. The threshold should reflect volume and buyer consequence; a universal 0.1-star rule would create false precision.
Checkpoint 5: exception governance
Fraud waves, litigation holds and unsafe-content events may justify temporary queues. An exception needs a scope, owner, start time, expiry and customer-facing treatment. It cannot become an unowned warehouse for inconvenient opinions.
Escalate any rating-correlated backlog, manual override without a reason, reviewer complaint about disappearance, or vendor configuration change affecting publication. Product, legal and customer-service teams may contribute evidence, but marketing should not be able to hide a review because it lowers conversion.
Checkpoint 6: durable records
Retain the policy versions, vendor settings, original reviews, decisions, edits, complaints and aggregate-rating calculations for the required period. The Fashion Nova order specifically required records relating to reviews, complaints and marketing representations. Your retention period must be set with counsel for the applicable law and order, not copied from this article.
Run a monthly independent sample: select published and unpublished reviews across every rating without letting the moderator choose the sample. Reperform the decision and report disagreement.
Reader asset: review-governance control sheet
| Control | Evidence to retain | Weekly test | Red flag |
|---|---|---|---|
| Intake | Original review, ID and receipt timestamp | Reconcile received, decided and backlog | Missing IDs or source-specific loss |
| Moderation | Rule, reason code, reviewer and edit history | Reperform a stratified sample | Rejection correlated with rating |
| Timing | Queue entry and decision timestamps | Compare percentiles by star | Low ratings wait materially longer |
| Publication | URL/status and syndication record | Match approved to visible | Approved reviews remain invisible |
| Rating | Formula, eligible population and displayed value | Recalculate visible and shadow scores | Unexplained divergence |
| Exceptions | Owner, rationale and expiry | Review every open exception | Expired or unowned queue |
| Oversight | Vendor configuration and change log | Diff settings against approved baseline | Silent threshold change |
The owner signs one statement: “The visible review population and rating are produced by the documented rules, and known exceptions are disclosed.” If that statement cannot be supported, the star display should not remain business as usual.
Why outsourcing does not outsource the claim
Review software can collect, filter, syndicate and calculate, but the retailer chooses the configuration and publishes the resulting representation. A vendor error may affect contractual responsibility between companies; it does not make the customer’s evidence problem disappear.
Procurement should therefore test the data model, export access, moderation logic, permissions, audit log and termination path before purchase. Growth teams should never receive a conversion tool that compliance cannot inspect. That is the uncomfortable position: a review system optimised only for social proof is not a trustworthy review system.
Related guides
Frequently asked questions
What did the FTC allege Fashion Nova did with negative reviews?
The FTC alleged that Fashion Nova’s third-party interface automatically published four- and five-star reviews while holding lower-starred reviews for approval, and that hundreds of thousands of lower-starred submissions were not published from late 2015 through November 2019. The final consent order prohibited review suppression and specified misrepresentations about publication and ratings.
Fashion Nova disputed the allegations and said a manual-release process was not completed during rapid growth. Because the case settled, accurate wording distinguishes the FTC’s allegations, the company’s response and the binding order.
Is it legal to reject any customer review?
Businesses can generally moderate reviews for content-neutral reasons such as spam, irrelevance, threats, personal information or unlawful material, subject to applicable law and their stated policy. The risk is suppressing genuine reviews because they are negative or applying nominally neutral rules unevenly. Publish the moderation standard, use reason codes and test decisions across ratings.
The caveat is jurisdiction and platform context: consumer-review, defamation, privacy and sector rules differ. Qualified counsel should approve the policy, especially when incentives, employee reviews or contractual restrictions are involved.
Must positive and negative reviews publish instantly?
No, but they should pass through equivalent rules and comparable service levels. Pre-publication fraud or content checks can be legitimate. The governance question is whether rating predicts delay or publication without a content-neutral explanation. Compare median and tail decision times by star rating, not just the overall average.
An emergency fraud hold may justify temporary delay, but it needs an owner, documented scope and expiry. If the backlog could materially distort the visible rating, suspend or qualify the rating until review processing is current.
How should an ecommerce team calculate its star rating?
Define the eligible population, weighting, product variants, syndicated sources, deleted reviews and rounding rule, then calculate the score reproducibly from retained records. A simple unweighted mean of eligible published ratings is understandable, but other designs can be valid if they are not misleading and are disclosed where material.
Also calculate a shadow score using all eligible received reviews to detect moderation bias. The caveat is fraud: suspected fraudulent reviews should not automatically enter the shadow population as genuine; retain them in a separately labelled class and validate the detection rule independently.
Who should be able to change review-platform settings?
Limit configuration changes to named administrators, require approval for publication thresholds and rating logic, and log every change. Marketing can propose experiments, but an independent commerce, trust or legal owner should approve changes that alter which opinions buyers see. Review the approved baseline automatically or weekly.
The exception is emergency action to stop spam or harmful content; even then, use a time-limited change and retrospective approval. Shared administrator accounts are unacceptable because they destroy the evidence needed to attribute and reverse a configuration change.
What should we do if we discover unpublished negative reviews?
Preserve the records, stop any biased rule, quantify the affected products and period, and obtain legal advice before altering evidence or contacting customers. Process legitimate reviews under the documented neutral policy, recalculate ratings and correct any material representation. Record what changed and monitor publication.
Do not silently dump a backlog live without checking privacy, relevance and product mapping. The caveat is that the correct notification, refund or regulator response depends on jurisdiction and facts; the operational steps support, but do not replace, counsel’s advice.
Next decision: How Do You Correct a Wrong Price or Offer Published Across Multiple Channels?
Related reading: Creator Commerce: The Complete Guide From Influence to Revenue · How Do You Measure Creator Cohorts and Repeat Purchase With Shopify and Klaviyo? · How Do You Build a Claim Evidence Register for Marketing Content?
Sources and research notes
- FTC Fashion Nova case docket — complaint, final order and case chronology; checked 26 September 2026.
- FTC settlement announcement — regulator’s account of the review workflow and alleged scale; checked 26 September 2026.
- Fashion Nova final decision and order — binding prohibitions, reporting and recordkeeping; checked 26 September 2026.
- FTC Consumer Review Fairness Act guidance — related US guidance on honest consumer reviews; checked 26 September 2026.
- Time report including Fashion Nova’s response — contemporaneous company position and procedural context; checked 26 September 2026.
- Limitations: This US consent matter did not produce a contested judicial finding on each allegation. The control sheet and operating thresholds are CDM recommendations and require jurisdiction-specific legal review.
This article is editorial guidance. Apply the principles in proportion to your market, evidence, and responsibilities.
