A store’s average rating is the most watched number from its reviews, and the least instructive. A store can hold a consistently high score and still read, week after week, about damaged packaging, a late carrier or goods that do not look like the photo. The average will not notice, because it is made up mostly of delighted ratings. It is more useful to read what customers write about, separate a one-off complaint from a recurring pattern, and fix the cause before you start polishing public replies.
Why the average hides what matters
Reviews are not written by a random sample of customers. Research published in MIS Quarterly describes two self-selection effects: people who expect a good experience are the ones who buy, and those with extreme opinions are more likely to write a review than those who are moderately satisfied. The result is a J-shaped distribution – many top scores, few in the middle and a smaller cluster at the bottom – and an average that is a biased estimate of quality. An analysis of more than 280 million reviews from 25 platforms found that most reviews on most platforms are polarised in this way, and that polarity reduces how informative reviews are.
Another study compared the average ratings of 1,272 products with independent quality tests. The average often failed to match objective quality and was frequently based on too few ratings. For a store owner the conclusion is simple: a one-tenth change in the average is a poor signal. When a problem appears, the average moves late and little, because the satisfied majority holds it up.
Yet customers do read negative text. A study of book sales at Amazon and Barnes & Noble found that, in most samples, a one-star review had a larger effect than a five-star one, and that buyers read the review text rather than relying only on the summary score. Research on electronics reviews showed that the information in a review cannot be captured by a single number: the text describes specific features and influences choice beyond the rating and the number of reviews. Both studies are older and American, but their shared direction is enough for the decision: what matters is in the text.
Where to find review themes
Heureka Verified by Customers (Ověřeno zákazníky). On Heureka, a Czech price-comparison site, the post-purchase questionnaire contains three open questions: what advantages the store offers, what its weaknesses are and how the customer would sum up their opinion. Heureka also shows sub-scores for delivery time, site clarity, communication and delivery quality over the last 180 days, and the share of customers who recommend the store over the last 90 days. Ratings without text count towards the percentage but are not shown on the profile. The admin offers an export of completed questionnaires for the last six months. The “weaknesses” field is the fastest route to themes.
The questionnaire goes to customers from the orders the store passes to the service, and the terms require all orders to be passed on. That sample is less self-selected than voluntary reviews – the polarity research found that people asked to rate their last purchase give less extreme ratings than those who choose what to review. It is not complete, though: customers can decline or ignore the questionnaire.
Google and on-site reviews. These are written voluntarily, so they tend to be more polarised. They work well as a second source: when the same theme appears on Heureka and on Google, it is less likely to be a quirk of one channel.
Operational data. Warranty claims, returns with reasons, support enquiries and carrier delays are not reviews, but they are the best test of whether a review theme reflects reality.
A one-off complaint or a recurring pattern
One harsh review hurts, but on its own it is information about one order. Before turning it into a project, go through four questions:
| Question | One-off complaint | Recurring pattern |
|---|---|---|
| How many independent customers say it? | one or two | several different people |
| Over what period? | once | across several consecutive periods |
| In how many channels? | only one | on Heureka, Google and in support |
| Does operational data confirm it? | no, or cannot be traced | claims, returns or delays show the same |
Beware of small numbers. An illustrative example with invented figures: in a quarter you have 40 written reviews and three of them mention damaged packaging, or 7.5%. A standard statistical estimate (the Wilson confidence interval) says the true share could lie roughly between 3% and 20%. A comparison with two mentions in the previous quarter therefore proves neither growth nor decline. With 400 reviews the same 7.5% would mean a range of about 5–11%, and the comparison would start to say something. For a small store it is more reliable to add up themes over a longer period and look for confirmation in operational data than to declare a trend from a handful of comments.
Assign each complaint to one theme and one owner. A short, consistent list is enough: delivery and packaging, availability and lead time, match with the description, product quality, communication and claims, website and ordering. A theme without an owner does not get fixed.
Your store’s data in one place
Korzaro connects a store’s sales, marketing and operational data so you can see what is really happening in the business.
Fix the operational cause first, then reply
Once a pattern is confirmed, the fix belongs where the problem arises. Complaints about damaged goods lead to packaging or the carrier; “looks different from the photo” leads to descriptions and images; repeated “nobody got back to me” leads to the support process. After the fix, watch whether the theme fades in the following periods. That is the only honest proof that the change worked.
Replying to reviews has value. A study of hotels found that those that started responding saw ratings 0.12 stars higher and 12% more reviews on average, and received fewer but longer and more detailed negative reviews. These are hotels, and the size of the effect cannot be transferred to an online store. The research also shows the key point: a reply changes how a problem is written about, not whether it exists. Heureka lets the store attach a reply to a review; a specific reply such as “we have changed how we pack fragile items” carries more weight than a generic apology, but only if it is true.
What not to do with reviews
Delete, cherry-pick or invent. Since January 2023, Czech consumer protection law has required sellers that display reviews to state whether and how they check that reviews come from genuine buyers. It prohibits claiming verification without reasonable measures, publishing fake reviews, commissioning others to write them and misrepresenting reviews. The Czech Ministry of Industry and Trade adds that this information must be clear and prominent – a mention in the terms and conditions alone is not enough – and that reviews given in exchange for a product must be distinguished. The EU directive behind these rules explicitly counts publishing only positive reviews and deleting negative ones as misrepresentation.
Buy or filter reviews on Google. Google’s rules prohibit offering a discount, gift or other incentive for a review, discouraging negative reviews, selectively asking only satisfied customers and posting reviews with a conflict of interest. Asking for reviews without an incentive is fine. In practice, if you send review requests, send them to everyone, not only to customers you expect to praise you.
Work with people instead of themes. Reviews with names and order details are personal data. You do not need a reviewer’s name to make an operational decision. In a public reply, do not discuss order details; move the specific resolution to private communication.
What reviews cannot show
Reviews describe the experience of those who spoke up, not of all customers. The silent customer who simply never came back is not in them, so read reviews alongside repeat purchases and returns. Be equally honest about what the customer did not write: a rating without text does not say why, and a short “fine” is not enthusiasm. Automatic summaries or theme grouping save time, but check the resulting groups against a few original reviews before you change operations on their basis.
What to do now
- Export written reviews from the last 90 days. The questionnaire export from Heureka, plus reviews from Google and your own site.
- Assign each complaint to one theme. Use a short, consistent list of themes and keep customers’ names out of the overview.
- Separate one-offs from recurring issues. Treat a theme as a pattern when it comes from several independent customers, periods or channels.
- Test the pattern against operations. Compare it with claims, returns, support enquiries and carrier delays.
- Fix the cause, then reply. Give each theme an owner and check in three months whether it is fading in reviews.