Scattered sentiment delayed brand moves — BigQuery pipeline, RAG summaries & Looker

Overview

A hotel group’s brand repositioning effort kept stalling because sentiment signals were scattered across social feeds, review sites, and call transcripts. Teams assembled snapshots by hand, definitions drifted by channel, and meetings focused on reconciling anecdotes rather than acting on trends. Intelligex built a sentiment pipeline that ingests social posts, review content, and call transcripts into BigQuery, applies channel-specific normalization, and generates executive summaries with a guarded large language model (LLM) that cites sources. Strategy and Marketing aligned on a single view of sentiment and drivers, reacted sooner to emerging issues, and avoided misreads because every statement linked back to governed evidence.

Client Profile

• Industry: Hospitality (multi-brand hotel group)
• Company size (range): Enterprise with global footprint and regional operations
• Stage: Active brand refresh and repositioning across select flags and regions
• Department owner: Strategy, Analytics & Executive Leadership (Corporate Strategy / Brand & Marketing)
• Other stakeholders: Digital/CRM, Customer Care/Contact Center, Operations, Revenue Management, Legal & Privacy, IT/Data Platforms, PR/Communications

The Challenge

Brand conversations were happening everywhere: guests posted on social channels, wrote detailed reviews, and called customer care with specific complaints and compliments. Each channel had its own cadence, format, and noise. Social listening tools surfaced volume and engagement but lacked structured links to property and segment. Review exports were inconsistent by site and region. Call transcripts mixed sentiment with operational details and personally identifiable information. Teams produced separate decks with overlapping claims, and leadership struggled to see whether a narrative reflected a real, multi-channel signal or a noisy week on one platform.

Constraints compounded the problem. Privacy rules limited how transcripts could be shared, and legal review slowed distribution of sample quotes. Marketing used campaign taxonomies that Strategy did not share, while Operations tracked issues in separate systems. The company did not want to rip out existing social monitoring or contact center tools; it needed a unified, safe analytics layer and a way to summarize trends credibly, with links back to sources.

Why It Was Happening

Identity, taxonomies, and calendars were fragmented. Social data referenced handles and hashtags, reviews referenced properties with inconsistent naming, and transcripts referenced booking or loyalty IDs. Channel teams used different time windows and sentiment scales. Manual stitching in spreadsheets produced defensible but incompatible views.

Governance arrived late. Quotes and charts were circulated without citations or privacy checks, and “top issues” lists were derived from spot samples. Summaries were narrative-led rather than evidence-led, and the burden of proof fell to the next meeting. Without a single operating layer, misinterpretations spread and course-corrections lagged.

The Solution

We implemented a governed sentiment pipeline and executive brief workflow. Social posts, reviews, and call transcripts landed in BigQuery on a schedule. Channel-specific parsers normalized entities, property identifiers, and campaign tags. A classification layer assigned topics and sentiment with confidence and masked sensitive text. Executive summaries were generated by a retrieval-augmented LLM that pulled from the governed dataset, inserted inline citations, and applied safety filters. Strategy and Marketing reviewed drafts in a light approval step before distribution. Nothing was replatformed: existing social tools, review feeds, and contact center systems remained the sources; the orchestration unified ingestion, normalization, summarization, and governance.

• Centralized storage and modeling in BigQuery with channel- and brand-level schemas
• Privacy protection and sensitive data masking for transcripts via Cloud Data Loss Prevention (DLP)
• Optional transcription and diarization for recorded calls using Cloud Speech-to-Text where needed
• Topic and sentiment classifiers tuned by channel, with confidence scores and human-in-the-loop review for low-confidence cases
• Executive dashboards in Looker with drill-through to source snippets and property-level context
• Guarded LLM summaries using a retrieval-augmented pattern that cites specific posts, reviews, and transcript excerpts (RAG: Use your data)
• Safety and policy filters applied during generation to reduce hallucinations and screen content (Vertex AI Safety)
• Role-based access tiers so Strategy, Marketing, and Operations see the right depth; sensitive text masked outside Customer Care
• Approval workflow capturing Strategy and Communications sign-off with comments and citations
• Audit trail including dataset versions, model configs, prompts, citations, and approvers

Implementation

• Discovery: Mapped social and review sources by channel and region, contact center transcript availability, and existing campaign taxonomies. Collected representative examples of misreads and conflicting narratives. Reviewed privacy and legal constraints on sharing quotes and transcripts.
• Design: Defined entity resolution for properties and brands, channel-specific schemas, and shared calendars. Authored topic taxonomy and sentiment scales. Specified DLP masking rules and access tiers. Designed Looker dashboards and the executive summary template with inline citations and caveats. Planned the approval flow and audit fields.
• Build: Stood up ingestion jobs into BigQuery; implemented normalization for properties, brands, and campaigns; configured DLP masking and transcript handling; trained and tuned classifiers with confidence thresholds; published Looker dashboards; and built the LLM summarization workflow using retrieval against the governed dataset with safety filters and citation injection.
• Testing and QA: Replayed prior months across channels, compared classifier outputs to hand labels, and tuned thresholds to reduce noise. Validated entity resolution and calendar alignment. Tested summary generation for citation accuracy and safety filters. Dry-ran the approval flow with Strategy and Communications.
• Rollout: Launched dashboards in read-only mode while teams continued their legacy decks. After validation, enabled the LLM summary for pilot brands with required approvals. Expanded channel and region coverage progressively, retaining a manual override for sensitive communications with post-review documentation.
• Training and hand-off: Delivered quick guides for Strategy and Marketing on reading dashboards and approving summaries, for Customer Care on transcript hygiene and masking, and for Data teams on taxonomy stewardship. Established a cadence for taxonomy updates and periodic model reviews with a human-in-the-loop path for edge cases.

Results

Leadership opened a single view of sentiment by brand, property, and topic with drill-through to evidence. Executive briefs cited the exact posts, reviews, and transcript snippets behind claims, and safety filters kept examples appropriate for broad audiences. Strategy and Marketing aligned on what mattered each cycle—shifts in cleanliness perceptions, front-desk experience, loyalty benefit recognition—and discussed actions, not definitions.

Rework declined because normalization, masking, and taxonomy rules were encoded upstream. When a campaign launched or a service change rolled out, the team saw how sentiment shifted across channels without rebuilding trackers. Approvals and audit trails documented why a summary read a certain way, making follow-ups and PR coordination smoother.

What Changed for the Team

• Before: Channel teams produced separate decks with conflicting scales. After: BigQuery and Looker delivered a shared, drilled view with consistent taxonomies.
• Before: Quotes circulated without citations or privacy checks. After: Summaries included inline citations and masking governed by DLP policies.
• Before: Narrative led evidence. After: A retrieval-augmented LLM anchored summaries in sourced snippets with safety filters.
• Before: Sign-off happened after distribution. After: Strategy and Communications approved summaries inside the workflow with an audit trail.
• Before: Misreads lingered across cycles. After: Normalization, shared calendars, and topic definitions reduced noise and re-interpretation.

Key Takeaways

• Unify social, reviews, and call transcripts under a shared identity and taxonomy to ground brand conversations.
• Mask sensitive data at ingestion and control access by role; privacy should be built into the pipeline, not applied in decks.
• Use retrieval-augmented generation with citations to produce explainable executive summaries; safety filters matter for broad distribution.
• Embed approvals and audit trails so Strategy, Marketing, and Communications align before narratives reach leadership or media.
• Keep existing listening and contact center tools; layer ingestion, normalization, summarization, and governance around them.

FAQ

What tools did this integrate with?
We centralized data in BigQuery, applied masking with Cloud DLP, optionally transcribed calls using Cloud Speech-to-Text, delivered dashboards in Looker, and generated summaries with a retrieval-augmented approach that cites sources (RAG) with safety controls informed by Vertex AI Safety. Existing social listening and review tools remained the data sources.

How did you handle quality control and governance?
We encoded entity resolution for properties and brands, standardized calendars, and tuned topic and sentiment classifiers with confidence thresholds. DLP masked sensitive transcript content, and role-based access limited exposure to raw text. Summaries required inline citations and moved through a light approval with Strategy and Communications. An audit trail recorded dataset versions, model configs, prompts, citations, and approvers.

How did you roll this out without disruption?
Dashboards launched in parallel with existing decks. After validating normalization and classifier behavior, we enabled the LLM summary for pilot brands and required approvals before distribution. Channel sources and tools stayed the same; the new layer orchestrated ingestion, masking, summarization, and governance on top.

How were different channels normalized and compared?
Channel parsers mapped handle/property mentions to a shared property and brand dimension, aligned timestamps to a common calendar, and converted sentiment to a consistent scale. Topic classification used a shared taxonomy with channel-specific tuning. Reports showed cross-channel views and allowed drill-through to raw snippets for context.

How did you prevent hallucinations or unsafe summaries?
Summaries used retrieval-augmented generation, limiting the model to the governed dataset and requiring cited snippets for claims. Safety filters screened for inappropriate content, and examples were masked as needed. Low-confidence or policy-sensitive drafts routed to a human reviewer, and nothing was released without Strategy and Communications approval.