What Is Voice of Customer? An Amazon Seller's Guide to VOC

What VOC Actually Is

In one sentence: voice of customer, or VOC, is the practice of systematically collecting what buyers actually said, classifying those words into groups, counting each group, and turning the counts into specific actions on your product or your listing. Run that loop once, end to end, and you have done VOC.

The term came out of product development work in the 1990s, where big companies used it to gather requirements before building anything new. It drifted into the Amazon seller toolkit only in the last few years, for a simple reason: review volume got big enough that a single category, with a few thousand text reviews, can tell you exactly what buyers complain about and how often.

Most sellers believe they already do VOC. Reading reviews, sure, everyone reads reviews. The difference is the word "systematically." Reading reviews looks like scrolling the page for half an hour, remembering the review that stung, and walking away with nothing you can act on. VOC spreads those few thousand reviews out, sorts every complaint into a group, counts which group dominates, who said it, in what words, and then decides whether the fix belongs in the product, the listing, or the customer service script.

The same product, described at two levels of depth, sounds like two different products. A 3.6-star average says "mostly fine." VOC says: this product collected 112 expressions of dissatisfaction, 74 of them landed on 3 stars, and the complaints cluster around cleaning hassle and too many small parts. Meanwhile the competing product, also sitting at 3.5 stars, has a similar pile of complaints centered on dying after a few months. Nearly identical scores, completely different diseases, completely different fixes. That example is real — the humidifier report on our site, which I will unpack in section five.

This article covers three things: how VOC differs from reading star ratings, what each of the four steps involves, and where the whole practice usually goes wrong.

How VOC Differs from Reading Star Ratings

Ratings are numbers, VOC is text. That sounds trivial. It is not, because the number itself is broken in a specific way.

We pulled the rating distributions from 26 public reports covering 52 products on our site, and buyers are simply not using the 1-to-5 scale. The 4-star slot holds a median share of only 10.5%; 24 of the 52 products sit below 10%, and the lowest, a USB cable, shows 1.4% — with 52.4% at 5 stars and 32.7% at 3 stars. The middle of the scale is basically empty. Buyers actually use three slots: happy goes to 5 stars, has complaints but still uses it goes to 3 stars, unusable or deceived goes to 1-2 stars. Five stars plus three stars together take a median 66.1% of all ratings; almost everything else piles into one and two.

Which means the average is fake smooth. A 3.9-star rating sounds like "most buyers are satisfied," but it is often just half perfect scores averaged with a third of the reviews sitting at 3 stars. If you want to know what buyers are actually thinking, the score cannot tell you. You have to go back to the words.

The text side has its own trap, and it is a practical one: most "negative review analysis" only pulls 1-2 star reviews. But the 3-star slot hides the biggest pool of dissatisfaction. We add 3-star and 1-2 star reviews together and call it the dissatisfaction pool. Across our 52 products, the pool takes a median 45.7% of all reviews, while the negative review rate (the 1-2 star share) has a median of only 24.0% — a factor of 1.74, and the widest gap among these products reaches 5.31. On 21 of the 52 products, 3-star reviews outnumber 1-2 star reviews combined.

The most extreme case is an electric toothbrush with a negative review rate of 8.6%, the lowest among all 52 products. Sounds very healthy. But 37% of its reviews sit at 3 stars against that 8.6%, so its dissatisfaction pool reaches 45.7% — nearly half of all buyers expressed some dissatisfaction in text, they just did not go down to 1-2 stars. A seller who only reads negatives sees "no real problems here." A seller who runs VOC sees "over forty percent of buyers have complaints, concentrated on charging."

So VOC does not replace star ratings; they are two layers. The rating tells you whether the outcome is good. VOC tells you why, and where. They are only useful together, which is why every report on our site puts the rating distribution, the negative review rate, and the pain-point breakdown on the same page.

The Four Steps of VOC

The full loop is four steps: collect, classify, count, act. Skip any one of them and the earlier ones are wasted.

Collect — gather the buyer's own words. Amazon sellers have four channels ready at hand: text reviews (negative and 3-star both), the questions buyers ask on the listing, return reasons from your returns dashboard, and the messages your own customer service receives. Reviews are the biggest and the most structured of the four, but negatives alone miss the 3-star pool described above. There is a sample floor: below roughly 100 text reviews on a single product, do not draw conclusions. A "top complaint" counted from thirty reviews is usually luck.

Classify — sort scattered sentences into a small set of groups. This is the technical core of VOC and the most tedious part to do by hand. "Stopped working after three months," "won't charge," "the motor makes a grinding noise" are one thing, and they belong in a durability group together. There is no universal right set of groups, but the categories must stay consistent within one analysis, or the counts cannot be compared. The reports on our site break complaints down by physical cause — the humidifier report, for instance, uses six groups: durability, maintenance, leakage, performance, smart features, and a long tail.

Count — for each group, three numbers: how many mentions, against which denominator, and whether the share is moving over time. Counting looks like the easy step. It is where most analyses quietly go wrong, because picking the wrong denominator bends every conclusion. The next section is about exactly this.

Act — translate the counts into owned actions. This is where VOC parts ways with ordinary review analysis. A review analysis ends in a summary. VOC ends in a list where every item has an owner: this complaint goes to product development (who, what change, what acceptance metric), that one goes to the listing ("easy to clean" into the first bullet point), another one goes to customer service (proactively state the cleaning schedule). Every report on our site ends with an action page like this — business-side suggestions carry an owner and a KPI, engineering-side suggestions carry a fix, a quantified acceptance metric, and a timeline. Until something lands on someone's desk, the loop is not closed.

The first three steps are method plus grunt work. The fourth is management. Most sellers die halfway through step three: collected the reviews, scrolled through them, never classified or counted, and ended at "we should improve quality," a conclusion everybody already had.

A Real Example: Same Score, Different Disease

Here is the full picture on one report. Our humidifier report compares Levoit against DREO across 437 text reviews — every number below is checkable on the report page. The two products have negative review rates of 17.3% and 18.4%, average ratings of 3.6 and 3.5. Practically identical. Now break the complaints down by physical cause:

At the rating layer, nothing here looks wrong: both products sit between 3 and 4 stars with negative rates around 18%. Any selection tool would score them as roughly equivalent. At the VOC layer, these are two different products with two different diseases.

Levoit's top complaint is maintenance — 21 mentions, nearly all about too many small parts, awkward cleaning, tank seams you cannot reach — and its performance group carries another 13 mentions about mist volume and noise. The product fundamentally works; it loses on daily experience. DREO's top complaint is durability: 24 mentions, 34.8% of its own denominator, with the typical quote being "stopped working after a few months," plus 12 mentions of leakage. It loses on reliability and sealing.

The actions could not be more different. If you sell a Levoit-type product, the fix is structural — fewer parts, a wider tank opening — and the listing leads with "easy to clean." If you sell a DREO-type product, motor life and the seal come first, and no amount of listing work survives "dead in three months." And if you monitor reviews, the metrics you watch differ too: for the first, whether the maintenance share drops after the redesign; for the second, whether "stopped working after X months" phrases become rarer.

Across the 26 head-to-head reports on our site, only 6 have both brands' top complaint in the same group. The other 20 all look like this one. Same score, different disease is the norm — which is exactly why the star rating cannot substitute for VOC.

Without a Denominator, VOC Is Just Vibes

Once classification gives you mention counts, the critical question is: divide by what?

Take durability in that same humidifier report. The group was mentioned 35 times. Three different denominators produce three different numbers, all correct, all saying different things:

All three are right. They answer three different questions. Choosing what to fix first — use the first denominator, it shows the structure of complaints. Judging what negative reviewers care most about — the second. Estimating how broad the blast radius is — the third.

The classic mistake is mixing them: taking "24.1% of pain-point mentions" and reporting it as "a quarter of buyers complain about durability." The buyer-level number is 8%. Getting this wrong downward makes your problem look three times smaller than it is; getting it wrong upward manufactures a crisis.

A subtler trap is comparing same-named groups across reports. "Durability" in report A might only capture total failures, while report B also counts "a bit loose." Different boundaries, and the percentages cannot be compared directly. Every report on our site prints its page denominator — all pain-point mentions — right under the breakdown table, precisely so every percentage can be traced to its base. For the full treatment of the three denominators, our article on analyzing competitor reviews goes deeper.

The Biggest Goldmine Sits at 3 Stars

Three stars keep coming up in this article, and they deserve their own section, because they are the most wasted data in this entire practice.

The 3-star reviewer occupies a special psychological position: not angry enough for 1-2 stars, not satisfied enough for 5, and exactly because of that, willing to write down both what works and what does not. The information density of a 3-star review is often higher than a 1-star — 1-star reviews are frequently pure emotion, 3-star reviews are assessments.

The numbers back this up. The median 3-star share across our 52 products is 17.9%, unremarkable on its own, but the spread is wide — up to 37.5% at the top — and on 21 of the 52 products, 3-star reviews outnumber 1-2 star reviews combined. That electric toothbrush with the 8.6% negative rate has 90 three-star reviews against 21 one-and-two-star, meaning 81% of its dissatisfaction pool sits at 3 stars.

How to actually use it: export every 3-star review on your main product — usually a few dozen to a few hundred — and count one thing, what these buyers are hesitating about. "Great product but the tank is small" and "great product but it's loud" point to two completely different improvements, and both buyers left 3 stars, so the average shows no difference at all. The most frequent "but" in your 3-star pool is usually the exact step between your product and 4.5 stars.

Negative-review-only analysis cannot reach this layer. That is a real, practical gap between "review analysis" and VOC: VOC's collection range is all text reviews, and negative reviews are just the loudest subset.

The Three Most Common Fake VOC

Having said what to do, here are the three fake versions we see most often. Each one is seductive.

First, watching the score and never the text. Decisions made on the average rating and the negative rate; relief when the score rises, panic when it drops; not a single review opened. This is the fake-smooth problem from section two — nobody knows what the 3.9 is covering, let alone what to fix.

Second, screenshot reading. Someone finds a sharp negative review, posts it in the ops group, everyone discusses it for half an hour, and nothing follows. However vivid, a single review is an anecdote with a sample size of one. Without classification and counting, you cannot know whether it is the most typical of two hundred or the most extreme. The whole difference between VOC and scrolling reviews is that VOC earns the right to ignore individual reviews, because the overall structure is known.

Third, acting on a sample below the threshold. Thirty reviews with ten noise complaints reads like 33%, but the confidence interval on thirty reviews runs something like 17% to 53%, which supports no decision at all. Below 100 text reviews on a single product, we draw no conclusions, and that floor is written into the methodology of our reports. If you want a fast small-sample signal, watch the trend in the 3-star pool and repeated questions in the listing Q&A — do not open a new mold because of thirty angry reviews.

All three fakes share the same flaw: they stop at "looking." VOC is a methodology precisely because it forces you through the boring steps, classification and counting, and turns feelings into numbers.

What VOC Cannot Prove

Every article in this series states its limits, and VOC has more than most.

Text reviews carry survivorship bias. Buyers who spend minutes writing are a minority; unrated star clicks vastly outnumber written reviews, and writers tend to be either delighted or furious, while the moderate majority stays silent. So VOC tells you what vocal buyers think, not what all buyers think. The negative review rate shares the same scope — 1-2 star text reviews over all text reviews — which is not the number shown in the star pool on the front end, and neither of them is a return rate. Most people who return a product never write a word.

Mention share is not impact. "Maintenance" at 27.6% of pain-point mentions describes the structure of complaints, not the share of buyers affected — silent buyers may not have the problem, or may have it and not bother writing. Converting mention share into "churn" or "return rate" has no basis whatsoever.

Our data covers 26 reports, 19 categories, 50 products, all of them electronics and small home appliances. That is not a census of Amazon, and extrapolating these numbers outside those categories deserves caution. The same product can also appear in several reports at different snapshot times; check which one you are citing.

VOC answers "what are buyers complaining about, and in what structure." It does not answer "how much will sales rise if we fix it." That takes post-change data. VOC's job is to point the direction.

Where to Start

If you have never done this systematically, do not build an apparatus. Start with one small thing: export all text reviews on your main product — negatives and 3-stars both — sort them into five or six physical groups, and count the mentions in each. One afternoon. The most common first reaction, in our experience, is discovering that the real top complaint is not the one you guessed.

To see what the structured version looks like, open the public reports on our site — the humidifier report (Levoit vs DREO, 437 reviews in six groups, each with mention counts and shares, followed by typical buyer quotes) or the electric toothbrush report (Aquasonic vs Philips Sonicare, 857 reviews, 20.4% negative rate). Every figure is checkable on the page, including which denominator sits under each percentage.

For the method in full depth, the pillar guide on analyzing Amazon reviews lays out the four-layer breakdown, and the 30-minute competitor review workflow shows how to run it on a schedule. VOC is not hard. What is hard is upgrading "reading reviews" into "counting reviews," every week — do that, and you hold a demand radar that sellers watching only the average rating simply do not have.