Case Study · Voice of Customer

A voice-of-customer radar that runs itself

A European automotive brand gets a weekly read on what customers say in public: comments from video platforms, Reddit and car forums, each one scored for sentiment and tagged by topic by LLM judges. The whole chain runs unattended every Monday, and an integrity gate blocks anything it cannot prove.

→Roughly 18,000 public comments scored and topic-tagged as of 10 August 2026.
→One cron job, weekly runs, cents per week, hard cost caps.
→Every published number traceable to a database row; every comment to a permalink.

Book a discovery call Get new case studies by email

Published 13 August 2026 by Creative Data Engineers. Corpus figures are as of 10 August 2026.

The short version

→A European automotive brand gets a weekly read on what customers say in public: roughly 18,000 comments from video platforms, Reddit and car forums as of 10 August 2026, each one scored for sentiment and tagged by topic.
→Two LLM judges do the reading: a 4-tier sentiment scale (negative, neutral, positive, recommended) and a fixed 13-topic taxonomy. An answer the parser cannot read stays unlabeled; the system never invents a label.
→The whole chain runs unattended every Monday on one cron job, for cents per week. Since late June it has run on its own, including the weeks where the correct action was to publish nothing.
→An integrity gate stands between the pipeline and the published dashboard. If any number on the page cannot be traced back to a database row, nothing deploys and the live dashboard stays untouched.
→The findings moved decisions: the most negative topic was a brand-trust question rather than any product attribute, and the brand's own channels drew harsher comments than third-party reviews.

The problem

Everyone has the comments, nobody reads them

When a brand launches a new product across several markets, the reaction shows up in public within hours. Under launch videos. In Reddit threads. On the car forums where owners talk to each other with no marketing department in the room.

That feedback is free, specific, and brutally honest. And in most companies it goes unread, because the volume is unmanageable: this brand's corpus spans 19 months across several markets. The classic answer is a one-off listening study: an agency reads a sample, builds a slide, and the picture is stale before the next quarterly meeting.

We wanted a radar: always on, cheap enough to leave running, and honest enough to trust without re-checking it by hand.

The system

Fetch, judge, gate, publish

The pipeline runs every Monday morning from a single cron entry and does its work in a fixed order.

→Collect. New public comments arrive from the video platforms (the brand's own channels and third-party review channels), Reddit, and car forums. Every row carries its permalink.
→Judge sentiment. An LLM judge scores each new comment on a 4-tier scale: negative, neutral, positive, or recommended. "Recommended" is reserved for an explicit endorsement, which is why it is rare. The judge runs at temperature 0, returns structured JSON with a confidence value, and only ever sees comments that have no label yet. If its answer cannot be parsed, the comment stays unlabeled. No label is ever guessed.
→Tag topics. A second judge assigns each comment up to several topics from a fixed 13-topic list (design, price and value, range and charging, software, reliability and service, brand trust, and so on). The list is closed: the judge cannot invent a new topic.
→Rebuild and gate. The dashboard is a single static page with the data inlined; rebuilding it is deterministic. Before anything goes live, an integrity test suite checks the build. Only a green gate commits and deploys.
→Report. Every run ends with an email: what failed and where, or how many new comments landed, or the honest "ran fine, nothing new this week."

Cost control is built in, and boring on purpose. Each judging step prices its workload before running and aborts above a hard cap. A weekly run costs cents.

The integrity gate

The dashboard refuses to publish what it cannot prove

→Every published comment must belong to a registered source and carry a permalink. A new source cannot appear silently.
→No number may be hardcoded in the page markup. Every figure is bound to the underlying data, and the gate hunts for stray percentages that are not.
→The gate re-computes a sample of the published numbers directly against the source database before every deploy. A skipped check does not count as a passed check.
→A week with no new data publishes nothing. The live dashboard is never touched by an empty run.

The gate keeps the dashboard honest. For a report a leadership team acts on, honesty is the property everything else depends on.

The findings

What the radar surfaced

As of 10 August 2026 the corpus stands at roughly 18,000 comments, every one classified. The useful part is where the negativity concentrates.

→The loudest pain was a brand-trust question. The most negative topic in the corpus was about whether the product is authentic to the brand's identity. That is a question no spec sheet answers, and nobody would have ranked it first by gut feeling.
→The brand's own channels drew harsher comments than third-party reviews. Comment sections under the brand's own videos ran measurably more negative than the same conversation under independent reviewers. The assumed direction was the opposite.
→Persistent topics point at the next content work. Because the radar runs weekly, movement is visible while it happens: a topic that keeps appearing while its negativity falls means the answers are landing, and a topic that persists without improving is where the next page, video or FAQ belongs.

A sentiment snapshot tells you where you stand. The weekly rhythm tells you whether anything you did worked.

The service

What the service looks like

Voice of Customer monitoring is part of our AI Search Visibility System, and it can also run alone. What you receive:

→Your scan. Sentiment split by topic, market and source, refreshed weekly, on a password-protected dashboard.
→The comments behind every number. Each with its permalink. No summary you have to take on faith.
→Recommendations. Where negativity concentrates, which topics are rising, and what to do about it.
→Content topics. The questions and themes your customers raise persistently that nothing on your site answers yet. These feed directly into content and AI-search visibility work, with demand evidence behind them.

FAQ

Common questions

What sources can a voice-of-customer radar read?

Public comment surfaces: video platform comments, Reddit, forums, review sites. Anything with public text and a permalink. Private communities and logged-in-only content stay out.

Does the LLM ever make up results?

The design assumes it will try. Labels come from a closed vocabulary, unparseable answers stay unlabeled, every published comment must trace to a source permalink, and an integrity gate re-checks published numbers against the database before anything deploys.

What does it cost to run?

At this brand's volume, cents per week for the scoring plus fixed infrastructure. Hard cost caps abort any run that would price above the limit.

Can I get this for my brand?

Yes. Voice of Customer monitoring is part of the AI Search Visibility System and is also available on its own. Book a call and bring the channels you suspect are talking about you.

Want to know what your customers say when you are not in the room?

Voice of Customer monitoring runs as part of the AI Search Visibility System, or on its own. Weekly scans, traceable numbers, recommendations you can act on.

Book a discovery call See the full service More case studies