VoxTrends at a 400-store European retail chain

The challenge

The chain's customer experience team was drowning in fragmented feedback. Social media mentions were tracked by the marketing agency in a spreadsheet updated weekly. Google Reviews were monitored store by store — manually — by regional managers. The call centre logged tickets in Zendesk but nobody cross-referenced them with social sentiment. NPS surveys ran quarterly and the results arrived two months after the fieldwork. Mystery shopper reports lived in PDF attachments that nobody aggregated. The CEO knew customer sentiment was declining. What she could not answer: in which stores, on which topics, driven by which channels, and whether the trend was getting better or worse. The data existed. The intelligence did not.

The solution

We deployed VoxTrends as a unified Voice of Customer platform covering every feedback channel the chain generates — 12 in total, across 6 countries and 14 languages. The technical deployment had five workstreams running in parallel. First, we wired the social media ingestion layer: Meta Graph API for Facebook and Instagram, X API v2 for Twitter, headless Playwright scraping for TikTok and local platforms (Heureka.cz, Zboží.cz). Second, we connected the review platforms: Google Maps API for location-bound reviews (matched to each of the 400 stores by Place ID), Trustpilot API, and custom scrapers for regional review sites. Third, the call centre integration: Genesys Cloud streaming transcription (voice-to-text) feeding directly into VoxTrends, with per-speaker diarisation and automatic PII redaction. Fourth, the structured feedback channels: Zendesk tickets, Qualtrics NPS export, mystery shopper PDF parsing via a document-understanding agent. Fifth, internal channels: regional-manager field notes and employee anonymous feedback from an internal survey tool. On the analysis side, every incoming event was normalised, language- detected (14 languages, auto-translated to English for LLM processing), and passed through the three-stage NLP + LLM pipeline: classical sentiment scoring, LLM topic extraction against a 47-topic taxonomy co-designed with the CX team, and LLM summarisation into natural-language insights with source citations. The output: eight dashboards (store heatmap, topic clustering, channel comparison, alert feed, competitive benchmark, trend detection, sentiment timeline, weekly executive summary) and a natural-language query interface where the CX director can ask "Which stores in Poland had a produce-freshness complaint spike this week?" and get a cited answer in seconds.

Why the spreadsheet era was over

When we arrived, the chain’s Voice of Customer infrastructure was a patchwork: one agency tracking social, one intern doing Google Reviews, a call centre on Zendesk, quarterly NPS on Qualtrics, and a binder of mystery shopper PDFs that nobody had time to read. Every channel had its own dashboard, its own taxonomy, its own cadence. Nothing was cross-referenced. The CEO had a gut feeling that certain stores were slipping — but no single place to look.

The core problem was not data availability. The data was abundant. The problem was synthesis: turning 12 fragmented streams into one answer per store, per week, with evidence.

What we built — the 12-channel pipeline

Social media (5 platforms)

• Facebook + Instagram: Meta Graph API for all brand page posts, comments, mentions, and DMs (with consent). Polling every 5 minutes for the brand account, hourly for competitor benchmarking.
• X (Twitter): X API v2 filtered stream for brand mentions, hashtags and competitor mentions. Near-real-time.
• TikTok: Headless Playwright scraping of brand-tagged content and comments. TikTok does not offer a search API for third-party monitoring, so we render the search page in Chromium, scroll programmatically, extract post metadata and comments, and rate-limit to one request per 4 seconds per IP via a residential proxy pool.
• Local platforms: Heureka.cz and Zboží.cz product reviews, matched to the chain’s SKU catalogue.

Review platforms (3 sources)

• Google Maps Reviews: Place ID-bound ingestion for all 400 stores. Every new review is geo-matched, sentiment-scored, and topic-tagged within 15 minutes of publication.
• Trustpilot: API integration for the main brand page and country-specific pages. Star rating + verbatim.
• Regional review sites: Custom scrapers for country-specific platforms (e.g. Ceneo.pl for Poland, Heureka for Czechia).

Call centre (voice-to-text)

• Genesys Cloud streaming integration. Every inbound call is transcribed in real time, diarised per speaker (agent vs. customer), and PII-redacted (names, card numbers, addresses stripped before LLM processing).
• Topic tagging happens on the full call transcript, not just the agent’s disposition code — which is often wrong or generic.
• Average 4,200 calls/day across the chain; VoxTrends processes every one.

Structured feedback

• Zendesk tickets: Every inbound ticket and internal comment, ingested via Zendesk API. Ticket categories are cross-mapped to the VoxTrends topic taxonomy.
• NPS / CSAT surveys: Qualtrics export (API), both structured scores and free-text verbatim.
• Mystery shopper reports: PDF parsing via a document- understanding agent. Reports are semi-structured (checklist + narrative), so the agent extracts both the checklist scores and the narrative findings.

Internal channels

• Regional manager field notes: Short-form mobile submissions from store visits, fed via an internal app → webhook → VoxTrends.
• Employee anonymous feedback: An internal survey tool (quarterly) with free-text comments about store conditions.

The LLM analysis layer

Stage 1: Classical NLP

Every voice event gets a fast sentiment score (positive / negative / neutral / mixed) via a fine-tuned DistilBERT classifier. Fast, cheap, and good enough for the aggregation layer. Runs on every event in under 50ms.

Stage 2: LLM topic extraction

A Claude-based pipeline reads the full verbatim (or the full call transcript, or the full mystery shopper narrative) and assigns it to one or more topics from a 47-topic taxonomy. The taxonomy was co-designed with the chain’s CX team over two weeks — each topic has a definition, three positive examples and three negative examples, so the LLM has unambiguous grounding.

Example output:

Topic: Queue length at checkout Sentiment: Negative (strong) Store: Ostrava — Poruba Verbatim: “Waited 25 minutes at the only open register. Three staff members visible but none opening another till.” Source: Google Review, 2026-03-14

Stage 3: LLM summarisation

Every Monday morning, VoxTrends generates a natural-language weekly brief per region and per country: the top 5 issues, the top 5 improving topics, the top 3 at-risk stores, and a recommended action list — all with source citations. The brief is exported as PDF and sent to the CX director and regional managers automatically.

The dashboards

Eight dashboards deployed, all customisable:

1. Sentiment Over Time — daily/weekly sentiment by store, topic, channel. The CX director uses this as her “weather report”.
2. Topic Clustering — bubble chart: volume × sentiment per topic. Click any bubble → drill into verbatim.
3. Store Heatmap — geographic map of all 400 stores, coloured by composite score. The most-used dashboard in the organisation.
4. Channel Comparison — which channels carry the loudest negative signal? Useful for resource allocation.
5. Alert Dashboard — real-time critical signals. A 1-star review spike, a viral social post, a call-centre escalation cluster. Slack notifications to the relevant regional manager.
6. Competitive Benchmark — sentiment, topic mix and share of voice vs. 4 named competitors, scraped from the same public sources.
7. Trend Detection — anomaly detection on topic frequency. Flagged an emerging “expired produce” complaint cluster in the Czech stores 3 weeks before it appeared in the NPS results.
8. Executive Summary — the weekly LLM-generated brief. The CEO reads this and nothing else.

Results

What the CX director said

“We used to argue about which stores had problems. Now we know. The store heatmap ended the arguments — it is the single most looked-at screen in our Monday meeting. VoxTrends did not just give us data. It gave us a shared truth.”

Services deployed

• VoxTrends (full platform: scraping, NLP, LLM, dashboards)
• Custom AI Agents (call-centre transcription agent, mystery-shopper PDF agent)
• Knowledge Engineering (47-topic taxonomy, geo-resolution ontology)
• AI Governance (PII redaction, GDPR retention policies, audit log)