# Customer Review Analysis: A Worked Example

Learn how to analyze customer reviews: document the platform, period, and selection, check individual cases, and keep conclusions within the corpus.

- Canonical URL: https://www.harmate.com/en/blog/customer-review-analysis-a-worked-example
- Author: Harmate Team
- Published: 2026-11-07
- Updated: 2026-11-07T13:51:28.895193+00:00
- Language: en

## Content

# Customer Review Analysis: A Worked Example

Customer review analysis can help a team identify a step in the customer journey to examine and choose an improvement to test. The worked table in this guide separates ratings, aspects of the experience, the reviewer’s words, and the next check. It turns scattered reviews into a decision memo whose sources can be revisited.

This guide is for a small team reviewing public feedback about a service to decide what to check next. It is not a way to measure satisfaction across an entire customer base from visible reviews, nor a method for automatically monitoring platforms.

## Define the question and the corpus limits

Begin by writing down the decision the reviews might inform. For example, you may want to check whether instructions sent before an appointment are understood. Then define the service, platforms, and period in scope. A bounded question makes searches and exclusions explicit.

Record what your review can actually cover: accessible pages, language, locations or branches selected, the period shown, and any filters used. Also state what the corpus leaves out. Public reviews do not show people who did not post, why they stayed silent, or a complete customer population.

The [France Num guide to customer reviews](https://www.francenum.gouv.fr/guides-et-conseils/developpement-commercial/gestion-de-la-relation-client/avis-clients-pourquoi-et) advises reading positive as well as negative reviews and notes that a single review cannot establish a trend. A study of reviews for purchased products also finds that reviewer self-selection can bias average ratings as a measure of quality; that result is specific to this setting and should not automatically be generalized to every sector ([Hu, Pavlou, and Zhang, 2017](https://scholarship.miami.edu/esploro/outputs/journalArticle/On-self-selection-biases-in-online-product/991032069201202976)). These sources call for caution; they do not measure your customer base.

## Record how the reviews were found

For another person to retrace your reading, record for each reviewed batch:

- the platform and relevant page or branch;
- the date you consulted it and the period covered by visible reviews;
- the selection method: filters, displayed sort order, pagination, or search terms;
- whether a review appears to have been solicited, and what evidence supports that observation;
- moderation or publication rules only when documented;
- visible edits, replies, or republications;
- a reference that lets you find the review again, consistent with access terms and privacy.

If the platform does not explain its display order, or you cannot tell whether an author was invited to post, write “unknown.” Do not turn missing information into a claim about authenticity, spontaneity, or moderation. This record can be made manually for a defined set; it does not assume exhaustive collection or automated extraction.

## Distinguish a review, an experience, and a person

A published entry is a review; it does not necessarily represent a distinct experience or a person you can identify. A pseudonym does not prove an author’s identity. Two similar statements may describe the same episode or different situations. Merge reviews only when you have verifiable evidence, such as the same case reference explicitly shown. Similar wording or names are not enough.

The fictional table below shows a limited way to take notes. Its references and ratings are invented and do not point to real pages.

| Reference and context | Rating and aspect | Fictional excerpt | Reported fact | Hypothesis to check |
|---|---|---|---|---|
| Platform A, review A-01, period consulted in March, “newest” sort | 5/5, welcome | “Warm welcome and clear explanations.” | The author describes a positive welcome and understood explanations. | Are explanations also understood at a first appointment? |
| Platform A, review A-02, same period | 2/5, wait and information | “I waited a long time and did not know whether my appointment was still on.” | The author reports a wait and uncertainty about the appointment. | At what point was information about the delay or appointment status missing? |
| Platform B, review B-07, visible date, invitation status unknown | 4/5, follow-up | “Very good overall; follow-up after the visit could be clearer.” | The review combines a positive assessment with a concern about follow-up. | Which part of follow-up is worth reviewing, and under what conditions? |

These three fictional reviews do not show that the welcome is generally good, that waits are frequent, or that follow-up is a widespread problem. They suggest different checks. The third review is mixed: reducing it to a positive or negative rating would erase its qualification.

## Move from excerpts to a check

For each selected review, separate four things: the published words, the reported fact or experience, your hypothesis, and a check you can carry out. Keep the source link or reference beside the excerpt. If known, distinguish the publication date from the date of the experience.

Revisit the review on its page before quoting it or interpreting it. Check whether a business reply, visible edit, or nearby context changes the reading. If a review was edited, several copies appear, or you cannot retrieve the source, record that uncertainty; do not reconstruct an earlier version by assumption.

A working table can add “verification status,” “missing information,” and “next check” columns. For example: “A delay is reported; arrival time and information received are unknown; review the notification process for this appointment.” This keeps the claim attributed to the review and turns interpretation into an examinable question. For a worked case that separates reported facts from follow-up and resolution, see our [customer complaint analysis method](/en/blog/customer-complaint-analysis-case-history).

## Compare cases without summarizing them with an average

Read positive, negative, and qualified reviews in context. Look for differences in period, service, location, channel, and publication conditions. An average rating can hide opposing experiences; it does not replace reading cases or describing how they were selected. Even if several reviews mention the same issue, the visible corpus alone cannot estimate how common it is among all customers.

Look for counterexamples and evidence that challenges your first hypothesis. If several reviews describe missing information, ask whether they concern the same journey and step. A positive review may reveal a process that works under some conditions; a negative review may identify a serious case to address without showing that it is common. The importance of a check may depend on its consequences, not only on the number of mentions.

## Write a bounded memo and choose what happens next

The final memo can fit on one page: question, scope and selection method; known or unknown limits; priority action; missing evidence; and planned check. For example: “Two of the three fictional reviews mention uncertainty in the service journey, one about the appointment slot and one about follow-up. Priority: check the information sent before and after an appointment. Missing evidence: the messages actually received in these episodes. Next step: review the notification template and examine a recent case.” This supports a check, not a frequency estimate or a conclusion about all customers.

If a broader satisfaction question remains important, plan a separate collection with its population and invitation method made explicit: it offers another perspective and does not automatically confirm public reviews. The [Customer Feedback — Experience & Improvements](https://www.harmate.com/en/product/template-library/1e2429e4-289e-4775-ba48-12ccd4a7c759) model offers adaptable questions about context, experience, and the change to prioritize. If you have a response table prepared for analysis, you can [import it into Harmate](/en/product/response-import), mapping the columns explicitly; public reviews are not imported from a platform. To organize several sources in a listening process, see [the voice-of-customer feedback loop](/en/blog/voice-of-customer-feedback-loop-small-team). Harmate does not collect or monitor public reviews, or reply automatically.

Useful analysis does not turn a few visible pages into a portrait of all customers. It leaves a checkable record of what was reviewed, what remains unknown, and the next question to resolve.