A retail chain with hundreds of stores collects thousands of reviews a week, and the star rating tells you almost nothing about which ones matter: a four-star review can contain a hygiene complaint, while a one-star review may be nothing more than a customer who could not find parking. This post explains how AI review classification works in BranchVoice, which labels every review receives, and what a regional manager actually does with them.
Why the star rating is not enough
Stars only exist on map platforms. Reviews from Şikayetvar (Turkey's largest consumer complaint site), Ekşi Sözlük (a high-traffic Turkish collaborative dictionary where brands are discussed) and X carry no rating at all. Even where a rating exists, it often disagrees with the text: customers habitually leave five stars and then write that the fitting rooms were dirty, or give two stars while praising the staff and criticising only the price.
The question head office needs answered is not "how many stars" but "what is in this review and who should handle it". Classification answers exactly that. Every review text, whatever its source, is processed with the same set of labels, so a Google review and a Şikayetvar complaint can be queried with one filter.
How AI review classification works
Once new reviews arrive through a sync, they are queued for analysis in the background. The AI reads each review together with your brand context: brand name, sector, product categories and known brand aliases. That is how sector-specific phrases are interpreted correctly, and how text that has nothing to do with your brand is kept out of the analysis.
Each review receives the following labels:
- Sentiment: positive, neutral or negative. It is independent of the star rating and based only on the text. Reviews that contain both praise and complaint are marked as mixed.
- Risk level: topics that require a fast response (hygiene, safety, misleading pricing, discrimination) receive a separate risk score. Reviews above a defined level count as "critical" and are passed to the alert engine.
- Response urgency: how quickly the review expects an answer. The "To do" card on the dashboard counts high-urgency reviews that have not been answered.
- Topic labels: staff attitude, checkout waiting time, fitting rooms, returns and exchanges, stock and sizing, and so on. The default list depends on your sector; you can add your own topics through analysis groups.
- Intent signals: churn intent ("I will never come back") and recommendation intent ("I would recommend this to anyone") are measured separately.
- Responsible unit: store staff, store hygiene, e-commerce logistics, e-commerce product or customer service. This is the field that routes a review to the right department.
- Suggestion and category: when a customer makes a concrete suggestion (product design, store experience, pricing, staff training), it is flagged separately.
- Employee mention: whether a member of staff is referred to by name or by title. See employee mentions in reviews for the details.
- Other flags: offensive language, suspicious (fake-looking) reviews, competitor mentions and the language of the review.
What happens to empty and irrelevant reviews
Many Google customers leave a rating and write nothing. Those reviews are not analysed and are counted separately as "no text" in the coverage strip on the dashboard. Text that has nothing to do with the brand (reviews left on the wrong business, unrelated chatter) is skipped as "irrelevant" and never sent again.
This transparency matters. When a regional manager sees the total number of reviews, the number with text and the number analysed side by side, they know exactly what every figure on the dashboard covers. Reviews whose analysis fails are retried a few times and then marked as skipped.
Where the labels show up
The labels flow into every layer of the product:
| Label | Where it is used |
|---|---|
| Sentiment | Experience score, review flow chart, platform and store breakdowns |
| Risk and urgency | To-do card, alert rules, anomaly detection |
| Topic labels | Topic Analysis page, impact matrix, topic-by-store heatmap |
| Churn / recommendation intent | Loyalty balance card |
| Responsible unit | Department Routing group |
| Suggestions | Customer voice section, suggestion categories |
In the review list, the "AI sentiment" filter shows only reviews classified as negative, and the "high urgency" filter lists what needs attention today. Critical reviews reach the responsible manager by email through critical review alerts.
What it means for a regional manager
Picture a manager responsible for the Izmir region of a fashion chain. Opening the dashboard in the morning and filtering to their region, they see that the share of negative reviews has risen in two stores, that the most frequent topic in those stores is "checkout and payment", and that the responsible unit is mostly "store staff". Three critical-risk reviews have already landed in their inbox. That picture decides what they look at during the weekly store visit.
The same labels work for senior management. The impact matrix in Topic Analysis shows which topics are both frequently discussed and negative. A low-volume topic that is almost entirely negative (returns, for example) is invisible if you only look at volume; classification brings that signal to the surface.
Next steps
Classification is the foundation that alerts, reports and summaries in BranchVoice are built on; none of them is meaningful without accurate labels. After you import your store list and connect your platforms, historical reviews are processed with the same labels. To see how your own reviews come out, request a demo.