Social listening and online reviews: what public conversations reveal and where the data stop

1. The problem: thousands of unsolicited opinions

One morning a restaurant chain reads a string of angry comments about a late delivery, spreading fast on a social network. The marketing director asks two things: how serious is this, and does it reflect what customers think in general? The first question has an answer in the online data. The second usually does not. The difference between them is the key to using social listening properly, that is, the systematic monitoring and analysis of public online conversations.

Online conversations about companies, products, public services and social issues are a vast body of material, generated spontaneously, with no questionnaire and no interviewer. Reviews on shopping platforms, app stores and online maps, comments on social networks, forum threads and comments under news articles all say something real about people's experience. The temptation is to read them as a free survey. They are something else: a valuable source with its own rules of interpretation.

2. What social listening is

Social listening works in three stages. The first is collection: a tool or a team gathers public posts and comments from the chosen sources for a given period, based on a keyword query. The second is organisation: removing duplicates and irrelevant messages, classifying by source, theme and tone. The third is analysis: what is being discussed, how much, in what tone, how it changes over time and how it compares with competitors.

It is important to distinguish monitoring from research. Monitoring tracks mentions continuously so the organisation can react quickly, like an alert system. Research uses the same data to answer a specific question, over a defined period, with a carefully coded sample of messages and documented conclusions. Both are legitimate, but the second demands the rigour of any data analysis and belongs with our monitoring and testing and data analysis services.

Online reviews are a special case. Unlike scattered comments, they are tied to a product or a location, usually carry a numerical rating and a text, and sometimes indicate that the author actually bought the product. They are the most structured form of unsolicited feedback and, for many local businesses, the most plentiful.

3. What it can and cannot tell you

Social listening data suit a few specific questions:

• Early warnings: a product fault, a breakdown or a new grievance often surfaces online before it reaches official complaints.
• Theme discovery: what people talk about, in what words, which aspects of the experience they raise unprompted. This is excellent material for building a questionnaire or an interview guide.
• Reactions to events: a campaign, a price change or a crisis, with volume and tone tracked day by day.
• Competitor comparison: which themes come up for each brand, which praise and which criticism are specific to it.
• Detailed diagnosis from reviews: which aspects of the service recur in the negative reviews of a location or a product.

They are not suited to estimates about a population: what share of customers are satisfied, how well known the brand is, what purchase intentions exist. Those need a sample that represents the population, in other words a customer satisfaction study or a brand awareness and image study, carried out through quantitative research.

4. Why online data are not representative

The underlying problem is self-selection. Nobody was chosen to write; those who write have a reason to. Online review ratings tend to bunch at the extremes: very satisfied and very dissatisfied customers write far more often than those who had an ordinary experience. The average rating therefore describes the balance between two motivated groups rather than the experience of the typical customer.

Other distortions add to this. Each platform's users have their own age, location and language profile, and a large part of the population never posts at all. Conversations in private groups and messaging apps are out of reach, though they may be the most candid. Some messages come from automated accounts or coordinated campaigns, and some reviews are fake, positive or negative. Access to data depends on platform rules, which change and can leave entire sources out. None of this makes the data useless, but it does mean they describe the online conversation, not the population.

5. How a social listening study is designed

A good study begins, like any research, with the question and the decision it has to support. Four design decisions follow, each documented.

Sources: which social networks, forums, news sites, review platforms and app stores are included, and what is missing. The list follows where the target audience actually talks, which is checked in an initial exploration as part of desk research.

The query: keywords, synonyms, common misspellings and product names, with operators that exclude irrelevant mentions. The query is tested: a share of the results is read by hand to check that the messages found are relevant, and additional searches check that important messages are not being missed. A brand whose name is also an everyday word needs a much more careful query.

Language: in Moldova, online conversations take place in Romanian and Russian, sometimes mixed within one message, written in both Latin and Cyrillic script and in transliteration. The query and the coding must cover all these variants, which takes a multilingual team, not just a machine translation tool.

The period: a baseline period before the event being studied and one after it, so that change can be separated from ordinary fluctuation.

6. Analysis: themes, tone and manual checks

Automated tools classify messages by tone, positive, negative or neutral, and by theme. This is useful for large volumes but often wrong: irony, sarcasm, negation, local expressions and mixed-language messages are misread. The rule for serious analysis is that a random sample of messages is coded by hand by researchers, following a written codebook, and the result is compared with the automated classification. If the gap is large, the manual coding is reported and the automated classification is used only for trends.

The codebook is built from the research question and from reading an initial batch of messages: experience themes, problem types, product attributes. When two researchers code the same subsample independently, their level of agreement shows how clear the codes are. The output is a map of themes, with their relative frequency in the conversation, their evolution over time and representative examples, paraphrased.

For online reviews, analysis combines the rating with the text: which themes appear in 1-2 star reviews compared with 4-5 star ones, how they differ between locations, what changed after an intervention. The company's replies to reviews can be analysed too: how quickly they come and whether they resolve the issue.

7. How the data become decisions

Findings from online conversations lead to three kinds of decision. Fast operational decisions: fixing a repeatedly reported problem, responding to a crisis, contacting the customers who wrote. Research decisions: themes discovered online become items in a survey that measures how widespread they are among all customers. And communication decisions: the language customers use spontaneously helps in wording messages.

The most productive combination uses online data as an alert system and a source of hypotheses, and the survey as the measuring instrument. If negative reviews often mention waiting times, a satisfaction survey shows whether the problem affects many customers or just a few very vocal ones, and mystery shopping shows what actually happens in the outlets.

8. Limits, common errors and ethics

The common errors are predictable: percentages of messages presented as percentages of customers; an automated "net sentiment" reported without a manual check; volume spikes interpreted without checking whether they come from a few authors; untested queries that sweep in irrelevant messages or miss the main sources; and comparisons across periods in which the available sources changed.

Ethics need particular care. A post being public does not mean its author agreed to become a research subject or to be quoted in a report. Data are analysed in aggregate, user names and profiles do not appear in the results, and quotes are paraphrased so they cannot be traced through a simple search. Personal data are processed in line with Moldova's Law no. 133/2011 on personal data protection and the GDPR, the platforms' terms and the ICC/ESOMAR International Code. Researchers do not join conversations under a false identity and do not use the data collected to contact people for commercial purposes.

9. Social listening compared with other feedback sources

10. What to ask the agency

• The research question and the decision the analysis supports.
• The list of sources included and, explicitly, those that cannot be covered.
• The full query, in Romanian and Russian, with spelling variants, and the results of testing it.
• The period analysed and the baseline period.
• The codebook and the procedure for manual coding of a random sample.
• A comparison between automated sentiment classification and manual coding.
• Reporting of the number of unique authors, not just the number of messages.
• The rules for anonymisation and personal data protection.

Used properly, social listening is an organisation's permanent ear: it hears early what is being discussed, uncovers themes nobody thought to ask about and shows how the active online public reacts. Used as a substitute for a survey, it mistakes the voice of those who write for the opinion of everyone. For a system that combines both sources, see our monitoring and testing and data analysis and desk research services.

See also: Monitoring and testing Data analysis and desk research Multilingual team Customer satisfaction and NPS Brand awareness and image study Desk research: what you learn before fieldwork

The Perspicuus research team · 30 June 2025