By Naeela
Use Amazon Voice of the Customer to identify offer-level product and listing problems, then verify the cause before funding a correction. The dashboard's CX Health compares your offer with similar offers, while 3 related NCX rates describe different signals. A poor rating needs investigation, but a color alone can't tell you whether to repair packaging, correct an expectation, hold a batch or challenge an unexplained record. The commercial job is to protect customers and spend on the fix the evidence supports.
Key Takeaways
- Amazon's current US guidance defines 5 CX Health categories: excellent, good, fair, poor and very poor. They compare offers with similar offers, not a universal percentage target.
- All 3 NCX rates discussed here use total orders as their denominator, but their numerators differ. NCX Review Rate is not negative reviews divided by all reviews.
- In a hypothetical 1,000-order example, 30 issue-return orders give a 3% NCX Return Rate. That doesn't establish the overall NCX rate or the offer's CX Health category.
- A hypothetical $600 packaging change that avoids $12 per preventable incident needs 50 prevented incidents to break even. Safety and mandatory corrective obligations override that commercial test.
Where should you start in Voice of the Customer?
Open the Voice of the Customer dashboard for the relevant account and marketplace. Begin with offers marked poor or very poor, as Amazon recommends, then read the comments and trend behind each flag. Keep the marketplace, ASIN, offer and observation date with the record so the team investigates the same problem.
Amazon's current US help page describes CX Health as performance relative to similar offers, based on recent orders and customer feedback. Excellent and good sit above comparable performance, fair is comparable, and poor and very poor sit below it. A numerical rate without its benchmark doesn't reproduce that comparison.
Insufficient feedback is not a clean bill of health. The page says new offers or offers without recent sales may not appear, and NCX rates and CX Health require a recent sale. Don't turn a missing row into evidence that a product has no defects. Use the product's actual inspection and customer records where the dashboard cannot yet support a judgment.
The “last updated” date also needs context. Amazon says it reflects the latest order or negative customer experience, not the date someone completed a quality audit. Deleted listings may remain visible for a few months with their previous feedback. Repeatedly deleting a row isn't a credible substitute for resolving its cause.
What do the three NCX rates actually measure?
Preserve the numerator and denominator together. The current primary definitions distinguish the overall issue signal from the return and review subsets.
| Metric | Numerator in Amazon's definition | Denominator | What it doesn't mean |
|---|---|---|---|
| Overall NCX rate | Orders where a customer reported a product or listing issue | Total orders | Every return or every low-star review |
| NCX Return Rate | Returned orders where customers reported a product or listing issue | Total orders | All returned orders divided by total orders |
| NCX Review Rate | Negative 1- and 2-star reviews reporting a product or listing issue | Total orders | Negative reviews divided by all reviews |
Amazon says the return metric covers the NCX-related subset, not all returns. The review metric similarly covers qualifying reviews, not every negative opinion. The page describes the return feedback and review inputs for these metrics as starting after January 2024. That historical starting point doesn't establish a fixed current reporting window.
Take an invented 1,000-order period. Suppose 30 returned orders carry qualifying product or listing issues, and 8 qualifying negative reviews appear. The corresponding rates are 3% and 0.8% under the stated definitions. These are hypothetical numbers, not healthy-performance thresholds or an ALFI result.
You cannot add them to claim 38 distinct affected orders. A review may concern an order already represented among the returns, and the review numerator counts reviews. Overall NCX also draws on other channels. Without an appropriate deduplicated record and comparable coverage, the sum describes neither unique customers nor overall NCX.
A review-monitoring report may use a different definition of negative. For example, FeedbackFive's technical documentation groups negative written reviews as 3 stars and below. Amazon's NCX Review Rate uses qualifying 1- and 2-star reviews divided by total orders. Match the star filter, issue scope, ASIN coverage and date range before comparing the numbers. The software documentation explains its own fields; it doesn't establish Amazon policy or prove a correction worked.
Which customer evidence belongs to the problem you're investigating?
Read across the available channels without treating every comment as a new incident. Amazon says its NCX signal listens to returns, refunds, customer-service contacts, buyer-seller messaging, seller feedback, A-to-z claims and product reviews. An issue can therefore appear even when the physical return count doesn't explain it.
There is also a provenance limitation worth retaining. Amazon's primary page says the feedback concerns items sourced by your business, but under the manufacturer-barcode program the items may not have been sold by you. That's a reason to establish scope before assigning fault. It isn't permission to dismiss customer concerns or assume another seller caused the problem.
Historical seller questions show why that distinction is commercially uncomfortable. In a December 2024 US discussion, a seller asked why a fit-related issue remained after a return was cancelled. That is an account-specific question, not proof of a platform error or a right to remove feedback. A cancelled return and a reported experience aren't necessarily the same record.
Keep what you can establish, what the customer reported and what remains uncertain in separate fields. A photograph of a leaking unit, an inspected seal and a quoted complaint carry different evidential weight. If the record is unexplained, preserve it and ask the precise question through the applicable support process. Don't invent a cause to make the action list look complete.
How do you distinguish a product defect from an expectation gap?
Test the strongest plausible explanation against the actual item. Amazon's root-cause summaries group customer concerns and suggest potential causes using machine learning. They help you find a pattern to inspect; they aren't an independent physical examination or a finding of supplier fault.
A “damaged” theme could lead to inspection of the product, its immediate packaging, preparation and the point where damage appears. Compare affected units with unaffected stock from the same and other lots where that evidence exists. If the only evidence is a broad comment category, the correct next step may be an inspection, not a catalog rewrite or a full-batch disposal.
For a fit complaint, compare the dimensions and version of the received product with the offer. Then inspect what the customer was reasonably shown: variation selection, images, quantity, compatibility and size context. Correct measurements in a technical field may coexist with an image that implies a different use. The conversion diagnostic covers the wider traffic-to-offer problem; keep this investigation focused on the specific post-purchase complaint.
Suppose, hypothetically, complaints cluster around a closure that opens in transit. If inspection confirms a packaging fault, changing the title won't fix it. If inspection instead finds two product versions mixed under a confusing variation, a stronger carton may solve the wrong problem. Write down what observation would distinguish those hypotheses before funding either intervention.
Safety or compliance evidence changes the pace. Assign the qualified product owner and the required process immediately; don't wait for a statistically tidy trend before addressing a potentially unsafe product. Conversely, one unverified preference complaint doesn't automatically prove every unit is defective. Scale the investigation to the consequence and the evidence, without pretending uncertainty is reassurance.
What makes a proposed correction worth funding?
Name the affected scope, the intended customer outcome and the cost before approving the change. A useful decision might authorize revised packaging for a defined batch, a corrected image for one variation, or a hold while a defect is examined. “Improve customer experience” doesn't tell finance what was purchased or operations what must change.
For a purely commercial, hypothetical example, suppose an extra protection step costs $0.30 across 2,000 units. The incremental cost is $600. If preventing one relevant incident avoids $12 of incremental expense, the change needs 50 prevented incidents to cover that cost. At the assumed 2,000-unit scope, that is equivalent to 2.5 incidents per 100 units, but only if the cost assumptions and exposure match.
The $12 must contain distinct avoidable expenses. Don't add the same refund and revenue reversal twice, call all product cost avoidable, or book an expected recovery as cash received. If some expense would occur anyway, leave it out of the incremental comparison. Use your actual contribution economics rather than an attractive revenue total.
Then ask whether the intervention can credibly prevent that type of incident. A cheaper fix has no economic advantage when it targets the wrong cause. Where evidence is incomplete, a bounded inspection or trial may be the sensible first purchase. Define its cost limit and the decision its result will settle instead of committing the whole catalog to an untested change.
This financial test doesn't authorize unsafe or noncompliant conduct. Mandatory corrections and customer-protection obligations can require action regardless of the modeled payback. The calculation allocates commercial resources among permissible options; it doesn't decide which obligations apply.
How will you know the correction helped?
Separate the changed units or offer version from the earlier exposure as far as your records allow. Record when the correction actually reached customers, not just when somebody approved it. A revised pack specification sitting with a supplier hasn't changed the inventory already being sold.
Keep the old and new product or packaging versions identifiable, along with their order exposure and observation age. A short post-change period may contain too few outcomes or unresolved returns to support a useful comparison. Don't declare success because a week passed without a new comment, and don't combine a mature old cohort with an immature new one as though they were equivalent.
Amazon says NCX and CX Health reflect historical performance, so changes may take several weeks to appear depending on order volume. That lag isn't a universal waiting period or a promise that the color will recover. Monitor the physical/customer outcome while the dashboard catches up, and respond to actual notices through their own process.
Define a decision in advance. Continued instances of the confirmed defect after the revised stock reaches customers should reopen the cause or execution check. A lower observed incident rate with comparable exposure can support continuing the correction, while changes in traffic, mix or seasonality limit a causal conclusion. If tracking cannot distinguish the units, label the result uncertain and improve traceability.
Keep the seller-support record for unexplained dashboard issues separate from your internal product-action log. One documents the question Amazon must answer; the other documents what your business changed and whether customers experienced it.
What should happen to advertising while the issue is open?
Base the instruction on the affected offer and customer risk. A confirmed material defect may require stopping the affected sale and following the appropriate process. A bounded expectation issue may call for correcting the offer and reassessing spend. Healthy offers elsewhere don't automatically need the same action.
We would refuse to buy additional affected demand merely to dilute an NCX denominator. More orders can change a rate without repairing the experience that produced it. Advertising needs a clear instruction from the quality and commercial decision, not an independent target to make the dashboard green.
At ALFI, Naeela and senior operators own the connected Amazon mandate across listings, inventory, advertising and economics. The 18-brand ceiling protects the attention needed to retain that context. We accept the uncomfortable recommendation, including slower spend, when it protects the business and its customers.
For private brands generating about $5M+ a year on Amazon that need that connected authority, our operating model explains the fit. We don't accept isolated PPC responsibility while product and stock decisions remain outside the mandate. A strong internal product-and-operations team may need no agency, and specialist safety or legal work still belongs with the qualified owner.
Is CX Health the same as the product's star rating?
No. Amazon describes CX Health as an offer's performance compared with similar offers using recent orders and customer feedback. Product reviews are one input channel, alongside others. The star rating alone cannot reproduce that assessment. Read the relevant comments, metric definitions and benchmark context before choosing a corrective action.
Is NCX Return Rate the same as the overall return rate?
No. Amazon defines it using returned orders where customers reported product or listing issues, divided by total orders. It covers a qualifying subset of returns, not every return. Preserve that distinction when comparing dashboards, and don't infer the overall NCX rate from the return subset alone.
How long will a correction take to change CX Health?
Amazon says historical performance can take several weeks to reflect corrective actions, depending on order volume. It doesn't provide a guaranteed recovery date in the inspected guidance. Track the changed units and customer outcomes while monitoring the dashboard. A slower metric response doesn't justify ignoring a continuing confirmed defect.
Should I increase orders to bring the rate down?
Don't buy more affected demand solely to dilute the denominator. It can improve a percentage while exposing more customers to the same unresolved issue. Verify the cause, make the appropriate correction and give advertising a scoped instruction. Maintain unaffected business where the evidence supports doing so, rather than applying a blanket stop.
What to do this week
Choose one poor or very poor offer. Save its metric definitions and customer evidence, identify what would verify the leading cause, and give one owner a funded corrective decision. Record which units or offer version changed and what outcome will reopen the investigation. The next review should inspect customer experience, not just whether the color moved.
