Gathering 30,000 reviews, but no one reads them
Review data collection is successful, but often faces challenges in utilization. Even though reviews from various channels are automatically gathered, there is no one to read the thousands to tens of thousands of accumulated texts. Eventually, the responsible person stops at skimming through a few pages of recent reviews occasionally.
Reviews are feedback (VOC) that customers invest their money and time in. The problem lies in the quantity. If the amount is too much for a person to read, the task of reading should be delegated to machines, while humans handle interpretation and judgment. This article outlines how to attach AI analysis to collected review data.
Where are they scattered: Review channel map
VOC is not gathered in one place. Depending on the product and industry, different channels need to be observed.
| Channel | Contained VOC | Main viewers |
|---|---|---|
| E-commerce reviews (Coupang, Naver, etc.) | Product quality, delivery, satisfaction relative to price | Brand, manufacturer, MD |
| App Store, Play Store | App quality, update responses, bugs | Service, app operation |
| Google Maps, Places | Store experience, customer service, hygiene | Offline store operation |
| Overseas platforms (Amazon, etc.) | Country-specific responses, comparison with local competitors | Global business |
| SNS, communities, blogs | Voluntary mentions, virality, spread of complaints | Marketing, PR |
As the number of channels increases, manual work quickly reaches its limits. Especially with overseas reviews overlapping language barriers, collected data can easily become unread by anyone.
The structure is simple: Adding analysis columns to each review
AI analysis does not require a separate complex system. It involves adding analysis result columns to each collected review record.
AI analysis data example
{
"channel": "앱스토어(일본)",
"review_body": "アップデート後、起動が遅くなった気がします...",
"translated": "업데이트 후에 실행이 느려진 것 같습니다...",
"sentiment": "Negative",
"keywords": ["업데이트", "실행 속도"],
"categories": [
{"category": "성능", "subcategory": "앱 구동", "type": "Negative"}
]
}
With translation, sentiment, keyword, and category attached to the original review (raw data), the recipient can start filtering and aggregating instead of reading the text.
AI analysis items
| Analysis Item | Task | Utilization |
|---|---|---|
| Sentiment analysis | Discrimination of positive/negative/neutral | Tracking negative ratio by channel and period, detecting spikes |
| Category classification | Tagging topics such as quality, delivery, price, CS, etc. | Identifying areas with high complaints, prioritizing improvements |
| Keyword extraction | Deriving frequently mentioned expressions | Early detection of new issues |
| Translation | Translating multilingual reviews into Korean | Comparing VOC by country on one screen |
Accuracy depends on operations, not the model
Even with the same AI model, analysis quality varies among teams. Since the model itself is now accessible to anyone, quality differences arise from three points outside the model. Hashscraper has refined VOC analysis across various industries through operations.
1. Quality of input data
The analysis starts with the input data. Even with a good model, analysis will be distorted if there are mixed duplicate reviews or mismatches in channel-product matching. We ensure that refined data is guaranteed from the collection stage to the analysis stage within a single pipeline.
2. Classification system tailored to the industry
If categories are too broad ("quality complaints"), there will be no actionable insights, and if they are too narrow, classification will be unstable. Just as "backlog" in cosmetics and "backlog" in apps represent different issues, industry terms should also be considered. Therefore, we design the classification system tailored to the business units of the client's improvement organization before starting the analysis. Starting from a classification system operated across various industries provides a faster and more stable foundation than starting from scratch.
3. Inspection and correction loop
We do not deliver AI classification results as they are. Initial results are reviewed by humans through samples to adjust criteria, and if new complaint types emerge during operation, categories are added and the system is refined. Criteria for handling complex reviews like "fast delivery but disappointing product" as positive delivery + negative quality signals are also determined through this process.
The difference between one-time analysis scripts and operational analysis pipelines lies here. Anyone can create a plausible graph, but analysis that can be trusted for decision-making requires these three elements.
With analysis, you can answer these questions
- Has the ratio of negative reviews increased this month compared to last month? In which channels?
- After launching a new product, what new categories of complaints have emerged?
- How do responses to the same product vary by country?
- In which areas are we weaker compared to competitor reviews?
These are questions that cannot be answered by "reading" each review individually and require aggregation and comparison. While each review is anecdotal, once analysis columns are added and aggregated, they become indicators.
Summary
The bottleneck of VOC analysis is not in collection but in reading. By adding sentiment, classification, and translation columns to collected reviews, people can shift from reading to interpreting and making judgments.
Hashscraper provides review collection and AI analysis as a single pipeline. It is possible to simply add analysis to existing data being collected, while maintaining the collection system as it is. We are responsible for crawler operation, maintenance, and monitoring.
Recommended reads
- From crawling data to decision-making
- How does a global consulting firm read reactions to a smartphone worldwide? - Case study of analyzing reviews and VOC from 7 countries
- How do global brands manage ratings and review responses from dozens of stores worldwide? - Case study of integrated monitoring of Google Maps reputation
- Analyzing reactions to own and competitor products through e-commerce review crawling
Start now
When was the last time you read through all the accumulated reviews? From collection to AI analysis, we will help you create a structure where reviews become indicators.
