How to analyse customer reviews: a step-by-step method

What analysing reviews actually means

If you want to know how to analyze customer reviews in a way that leads to decisions, the goal is simple to state: turn thousands of free-text comments into a few numbers you can compare, without losing what customers actually said. Which topics do customers raise, how do they feel about each, where and when is that changing, and why?

The method below works whether you have two hundred reviews in a spreadsheet or two hundred thousand in a database. The tools change with scale; the steps do not.

Step 1: gather every review

• Collect from every source that matters: Google Maps for each store, app stores for your app, and any other review site your customers use.
• Keep the fields you will need: source, location, date of the review, rating and the full text.
• Map each listing to the store it belongs to, so results roll up correctly.
• Collect regularly, not once. Analysis is most useful as a trend.

Step 2: clean and prepare

• Remove exact duplicates, which inflate counts.
• Keep reviews in their original language; translate for reading, but classify the original where your method allows.
• Set aside texts too short to carry a topic, such as "good" or "ok". Count them, but do not force them into a category.
• Note rating-only reviews separately. They count towards the rating but tell you nothing about why.

Step 3: split each review into mentions

A review is rarely about one thing. "Beautiful designs, helpful staff, but billing took forever" has three mentions: two positive, one negative. Analysing at the level of the mention, not the review, is what keeps both the praise and the complaint. For a small set you can do this by hand; at scale it needs a model that identifies each topic and the words that support it.

Step 4: classify into a fixed set of themes

Every mention needs to land in the same set of categories, or nothing can be compared. A three-level taxonomy works well: a theme (Billing), a sub-theme (Waiting time) and a specific issue (Only one counter open at weekends). Build it from your own reviews, not from an org chart, and keep it stable over time.

Step 5: measure sentiment per mention

Give each mention its own sentiment: positive, negative or neutral. Then, for each topic, calculate the share of mentions that are negative. That single number, negative share, is easier to explain and compare than an average sentiment score.

Step 6: find what matters

• Rank topics by negative volume, negative share and change since the previous period.
• Break the top issues down by location, source and month to see where they concentrate.
• Flag findings built on few reviews instead of ranking them.
• Read a sample of reviews behind each finding before drawing conclusions.

Step 7: act, and check

Give each rising issue to the person who can change its cause, with the evidence attached. Record the baseline and a date to look again. When the date comes, compare the issue's negative share with stores where nothing changed.

Doing it by hand

With a few hundred reviews, a spreadsheet works. Put one review per row with its date, location and rating, add columns for up to three topics and their sentiment, and tag each review yourself. Use a fixed list of topics in a drop-down so spelling does not split categories. Count with a pivot table. It is slow, but it teaches you what your customers talk about, which makes any later automation better.

Common mistakes

• Starting with a word cloud. Frequent words are not topics, and they lose sentiment entirely.
• Classifying whole reviews to one category, which drops half of every mixed review.
• Changing the categories every month, which breaks every trend you are trying to track.
• Reporting without the reviews. Store teams trust a finding they can read.
• Treating the rating as the outcome. Measure the issue you are trying to fix.

Doing this with Falcon

Falcon does this for your Google Maps and Google Play reviews. Every review is classified into themes, sub-themes and issues from your own taxonomy, with the sentiment of each mention; specific issues are found by clustering what customers wrote, each with a short summary; and every count opens the reviews behind it, with the supporting quote highlighted.