How to make customer feedback count in product development

Companies have more customer feedback than ever: surveys, app store reviews, support conversations, user tests and interviews. It reaches product teams in pieces, and never as a complete picture. The score makes it into the monthly report, while the comments behind it go unread.

Product teams decide with the data in front of them. Feedback counts in those decisions when three things come together: every comment is counted, the words behind the numbers are one step away, and both arrive in the tools people already use. This post shows what it takes to get there.

Where customer feedback comes from

The sources look very different. There are short scores, one-line reviews, support conversations, session data and recordings of user tests and interviews. One setup handles them all.

Surveys. A score and a short comment, often thousands a day. Together they tell you how customers rate an experience and what drives the score.
App store reviews. A star rating and a few lines, written in public. Useful for seeing how users react to the app, often right after a release.
Customer service. Conversations about a specific problem, from a short chat to a long email thread. This is where you learn where customers get stuck and what they need help with.
Captured user interactions. Sessions recorded with tools like Sentry. A record of what users actually did, for example right before they got stuck.
User tests and interviews. Recordings of what customers say and do while they use the product. They explain behaviour in depth.
Everything else. Social media posts, public reviews, community forums, sales conversations. Anything customers write or say can be added as a source.

The building blocks behind them stay the same. Only the connector at the start changes.

Combine quantitative and qualitative analysis

You get the most out of feedback when you can look at it in two ways. The numbers show what changed and by how much: which topics grow, which complaints spike after a release, how NPS moves. The words show why: what customers actually wrote or said.

Many setups give you one or the other. A dashboard of scores tells you that something went wrong, but not what. A handful of quotes is convincing, but you can't tell how typical they are. Often you need both, and a quick way to move between them.

Classification turns comments into numbers. Every comment is sorted into the company's own categories, the topics the team actually works on. Then feedback can be counted, tracked over time and read alongside scores such as NPS and CSAT.
The knowledge repository explains the numbers. It combines search with a wiki. Search finds the original quotes and recordings across all feedback. The wiki condenses them into linked pages on topics and findings, updated as new feedback comes in, and every statement points back to its sources.

Deliver where people already look

Companies already have places where people look at data: a warehouse such as BigQuery, dashboards in Looker, recurring reports.

That is where we want feedback to arrive. Instead of asking people to open one more tool, we deliver the data to the places they already look. There it sits next to the other data that product decisions are based on.

Warehouse and dashboards. Classified feedback lands in your warehouse, for example BigQuery, and appears in the dashboards your team already uses, such as Looker.
AI assistants. People connect an agent they already use, such as Claude Desktop, through MCP. They ask questions about the feedback in plain language and get answers with numbers and quotes.
The podojo app. The same data is also available in our own app, for people who prefer a classic interface to asking an assistant.
A daily digest by email or Slack. It does two jobs. People see right away what came in and which categories went up. It also works as a check on the pipeline: when the digest arrives with fresh data, everything before it has run.

How it works

The building blocks

Every source runs through the same five building blocks. The results arrive in the infrastructure your team already uses.

It takes two kinds of expertise

None of this is a purely technical project. It works when two roles work closely together.

Technical expertise. This is the part we bring. We build and run the setup: the connectors, the removal of personal data, the classification and its evals, and the integration into the company's infrastructure.
Subject matter expertise. An expert inside the company who knows the data and the business, owns the categories and checks whether the results make sense.

Proven in practice

We have run this setup for mobile.de since early 2024. It covers website surveys, in-house survey data, app store reviews, user tests and interviews: thousands of responses every week. Classification reaches around 90% accuracy. Classified results flow back into mobile.de's own storage and from there into BigQuery and Looker. Teams across mobile.de use it in the app, through MCP from their AI assistant, and in dashboards.

With Amnesty International Germany, we used the same building blocks to analyse several hundred thousand social media posts.

Where to start

You don't need all of this at once. Pick the source closest to a product decision your team is about to make, agree on the categories with your expert, and deliver the results to one place people already look. Every further source then reuses the same building blocks.