Every retail chain has questions of its own. A fashion brand cares about the fit of the new collection, a grocery chain about checkout queues, a cosmetics brand about whether testers are available. No ready-made topic list covers all of them. This post explains how analysis groups and topic taxonomy work in BranchVoice, how to define your own topic, and how to apply it to reviews that were analysed before the topic existed.
What the topic taxonomy is
The taxonomy is the list of topic questions the AI answers while reading each review. During AI review classification, these topics are asked alongside the core questions such as sentiment and risk, and the topics found in a review are stored as its labels.
In BranchVoice, topics are organised into groups. Two system groups come ready in every company:
- Core Sector Parameters: default topics for your sector; in fashion retail that means staff attitude, fit and sizing, fitting rooms, returns and exchanges, stock, store layout and so on.
- Department Routing: which unit a review concerns (store staff, store hygiene, e-commerce logistics, e-commerce product, customer service). No separate question is asked for this group; it is derived from the core analysis.
Alongside these you add your own groups, such as "Collection and Product Experience" or "Campaign and Price Perception". System groups cannot be deleted but can be deactivated; custom groups are entirely under your control.
Defining analysis groups and topics
Defining a topic means deciding three things:
1. Question type. There are three:
| Type | What it returns | Example |
|---|---|---|
| Yes / no | Whether the topic is present, as a probability | "Does the customer complain about the fitting room?" |
| Scale | One of the levels you define (2 to 10 levels) | Campaign perception: very negative … very positive |
| Choice | One of the options you define | Reason for complaint: size missing / colour missing / model missing |
For choice topics, "no such situation" and "other" options are always added automatically; the AI is never forced into one of your options.
2. Instruction. The question the AI applies to the review. Instructions are written in English and must make sense on their own; the title and description appear in the interface in your language.
3. Channel. Which reviews the topic is asked on: physical store, e-commerce or both. Asking a "delivery time" topic on Google Maps reviews is wasted cost; the channel filter prevents it.
Before saving, a preview runs the topic against the last five reviews so you can see what the AI answers for each one and refine the instruction. The number of custom topics per company is capped (fifty), a reminder that every topic is one more question on every review and that cost grows with the number of topics.
Applying a new topic to older reviews
When you save a new topic it is only asked on reviews that arrive afterwards; older reviews do not change by themselves. That is what backfill is for. After saving, the interface shows how many reviews have not yet been asked the topic and offers to backfill them. In this run the AI is asked only that topic's question; the review's sentiment, risk and other labels are not recalculated.
If you later change a topic's instruction, backfilling in "all" mode re-evaluates every review the topic applies to with the new question. Deactivating a topic stops the question and the cost; past labels are not deleted.
A full re-analysis (including all core questions) is a separate operation limited by your plan's monthly allowance; in day-to-day use, backfill is almost always enough.
How groups appear on the dashboard
Each group is a tab on the Topic Analysis page, with system and custom groups on the same row in the order you set. Inside a group tab, each topic is shown with a card suited to its type: yes/no topics with detection rate and negative share, scale topics with average level and distribution, choice topics with option shares. Below that come the topic trend, topic-by-platform and topic-by-store heatmaps, and example reviews.
The impact matrix on the Overview can be filtered by group to show only your topics on the volume and negativity axes; the dashboard metrics guide explains how to read those charts. A topic or group with "show on dashboard" switched off is still asked during analysis but takes up no screen space.
There is one more kind of group: lead groups. Topics in these groups are asked only on prospective-customer posts captured from social media, and the topic weights determine the lead score; see capturing leads from social media for details.
A worked example for a fashion chain
A fashion chain that wants to track its new season collection creates a "Collection and Product Experience" group. It adds three topics: "Fit and sizing" (scale, five levels), "Fabric and durability complaint" (yes/no) and "Product type not liked" (choice: underwear, socks, loungewear, other). All three are asked only on the physical store and e-commerce channels; the chain previews them, saves them and backfills the last ninety days.
Two weeks later, Topic Analysis shows that fit complaints cluster around one product type and that two regions are affected more than the others. That is a concrete finding for the product team, delivered in the weekly report, and it could not have been seen with the ready-made topic list.
Next steps
Analysis groups and topic taxonomy turn BranchVoice from a generic review tool into an analysis layer that asks your chain's own questions. System groups work from day one; custom topics can be defined and previewed in minutes and backfilled onto older reviews. To try your own questions, request a demo.