Competitor Review Analysis: Compare Stores Topic by Topic
Competitor review analysis that stops at the star average cannot say what the rival does better. Compare stores topic by topic, under collection rules writte...
Every month, someone on the regional team of a fashion chain opens Google Maps. They look up their own store, then the rival chain's store across the street. Two star averages and two review counts go into a spreadsheet. The rival is ahead again. In the operations meeting, the regional director asks the obvious question: ahead on what? The fitting rooms, the queue at the till, the staff, the exchanges? Nobody can say, because the spreadsheet only holds the averages. Competitor review analysis that stops at the star rating answers who is ahead and never why.
Key Takeaways
- Online product ratings tend to follow a J-shaped distribution, mostly 5-star, some 1-star and few in between, as Hu, Zhang and Pavlou showed in Communications of the ACM in 2009. An average hides which complaints sit at the bottom.
- Aspect-based sentiment analysis links each part of a review to a topic and a sentiment. In a 2024 comparison, the strongest method reported reached 90.33 percent accuracy on a restaurant benchmark and 86.21 percent on a laptop one.
- Academic frameworks already benchmark competitors topic by topic from reviews. One study used 26,934 reviews of 26 hotels in Bangkok.
- A fair comparison between stores depends on rules written before collection. Same source and period, a fixed topic list, a minimum of mentions, and a check for fake engagement.
What competitor review analysis means for a physical store
Competitor review analysis is the comparison of public reviews of your stores with those of competing stores. It shows where each one does better or worse, in the customer's own words.
A topic, also called an aspect, is one part of the experience a review talks about, such as queues, staff, prices, exchanges or the fitting rooms.
Aspect-based sentiment analysis is a text analysis method that finds the topics mentioned in a review and the sentiment attached to each one. A 2024 comparative study on arXiv describes it as identifying specific aspects within the text and the sentiment associated with each aspect. The same review can praise the staff and complain about the queue.
For a chain with physical stores, the review pages of its own store and the one next door are the freshest customer feedback anyone can read. They are public, they cover competitors, and they arrive without a questionnaire. The problem is what happens to them after they arrive.
How the comparison is made today
PERSONAL EXPERIENCE In retail and services clients I have worked with, the routine looks the same whatever the sector. Nobody would call it a method. It is what fits in an afternoon.
- Someone looks up each store and each competitor on the map, one by one.
- They type the star average and the review count into a spreadsheet, one row per store and month.
- They read the most recent reviews of the worst store and copy two or three complaints.
- Those quotes go into the monthly operations deck, next to the averages.
- When a competitor pulls ahead, someone visits the store to see what is different.
That is the word clients use when they describe it: they still do it in a spreadsheet. Nobody is proud of it, and everyone knows it misses most of what customers wrote.
Why the star average cannot say "better at what"
Star averages have two problems, and both are well documented. The first is shape. Hu, Zhang and Pavlou showed in Communications of the ACM in 2009 that product reviews tend to a J-shaped distribution. Most ratings are 5 stars, some are 1 star, and hardly any sit in between. An average of that shape is a mix of a happy majority and an angry minority. It cannot tell you what the angry minority wrote about.
The second is mixing. One star average blends every topic into one number. A half-star gap between two stores can come from queues at one and staff at the other. It can come from a single bad month, or from one store asking every customer for a review at the till. The spreadsheet records the gap and erases the reasons.
A J-shaped distribution is a pattern of ratings piled up at the top score, with a smaller pile at the bottom score and very few in the middle.
What it costs, and who answers for it
PERSONAL EXPERIENCE The cost shows up as the wrong fix. A chain sees the rival ahead and repaints the store, while customers were writing about the wait for the fitting rooms on Saturdays. Or it adds staff on the wrong days, because nobody split the complaints by topic and time.
The people who answer for it are regional operations and whoever owns market intelligence. They present the comparison every month. They are asked why the gap is still there after the investment. And they still build the comparison by hand, copying numbers from one screen into another, the same fragmented work we see across research operations.
Benchmarking research names the same gap. A 2025 study of hotel benchmarking from online reviews notes that traditional approaches rely on customer surveys and financial or operational metrics. Its authors describe them as time-consuming and limited in capturing the service attributes that come from actual customer experience.
Topic by topic: how the comparison gets built
The same Sogang University study, by Juram Kim and Chiehyeon Lim, built a framework in four parts. They tested it on 26,934 reviews from 26 hotels in Bangkok. Topic modeling found the service attributes in the review texts. Sentiment analysis measured how each hotel performed on each one. Clustering identified which hotels were real competitors. An importance and performance analysis then pointed each hotel to the attributes worth fixing first.
A 2024 study from the Hong Kong Polytechnic University, published in the International Journal of Hospitality Management, did the same for restaurants with 38,479 reviews. It measured attribute performance with a language model and followed how the importance of each attribute changed over time.
The method underneath works well on standard benchmarks, and it still errs. In the 2024 comparative study, the strongest method reported, LSA with DeBERTa, reached 90.33 percent accuracy on the Restaurant14 benchmark and 86.21 percent on Laptop14. The best method the authors ran themselves, FAST LSA, reached 87.6 and 82.6 percent. Even at the top, 10 to 14 labels in every 100 were wrong.
Even the best topic labels miss some
A bar chart of accuracy, in percent, for aspect-based sentiment analysis on two public benchmarks, from Jayakody and colleagues, Aspect-Based Sentiment Analysis Techniques: A Comparative Study, arXiv 2407.02834, 2024. Panel one, restaurant reviews on the Restaurant14 benchmark: LSA with DeBERTa, as reported, 90.33 percent; FAST LSA, the best method the authors ran themselves, 87.6 percent. Panel two, laptop reviews on the Laptop14 benchmark: LSA with DeBERTa, as reported, 86.21 percent; FAST LSA, 82.6 percent. The bars start at zero, so the gap to 100 is the share of labels that were wrong. A closing line says that between 10 and 14 labels in every 100 were wrong even for the strongest method, so the excerpts behind every percentage should stay visible. Every number belongs to the study named and none of them is a Cassi.ai measurement.
A topic percentage is a count of labels, and some labels are wrong. The excerpts are how a reader catches them.
That error rate is why every percentage on a topic needs its excerpts next to it. A reader who can open the reviews behind the share of queue complaints can catch a wrong label. A reader who only sees the number cannot.
| Method | What it shows | What it hides | Effort per month |
|---|---|---|---|
| Star average and count | Who is ahead | Why, by topic and by period | Minutes, by hand |
| Reading recent reviews | A few vivid complaints | How common each complaint is | Hours, and only for a few stores |
| Topic by topic comparison | Where each store wins or loses, with excerpts | Customers who never write a review | Set up once, then repeated |
The rows come from Hu, Zhang and Pavlou, the Bangkok hotel study and the aspect-based comparison. The effort column is our reading.
Write the collection rules before you compare
UNIQUE INSIGHT A topic comparison between two stores is only as fair as the rules that collected the reviews. Write them before the first collection, and apply them to every store in the same way.
- Define the competitor set: stores of a similar type in the same neighborhood or mall. The Bangkok study used clustering for this, because the store across the street is not always the real rival.
- Use the same source and the same period for every store, such as the last twelve months.
- Count reviews with text apart from star-only ratings, since only text can be read for topics.
- Fix the topic list before reading, and use the same list for every store.
- Set a minimum number of mentions before a topic gets a percentage.
- Flag unusual patterns before comparing, such as a burst of reviews in one week.
- Keep the excerpts behind every percentage one click away.
- Never name a winner where the difference sits inside the noise.
Five checks before two stores are compared
A decision flow in five checks, written before the first collection and applied to every store the same way, each with an exit to the right. Check one, a real competitor: a similar store type in the same neighborhood or mall. If not, it is not a rival and is left out of this comparison. Check two, same source and same period: every store read from the same place over the same months. If not, the stores are different samples and the reviews are collected again before comparing. Check three, enough mentions per topic: the topic passes the minimum written in advance. If not, the dashboard shows the count and no percentage. Check four, no unusual pattern: no burst of reviews and no sign of reviews asked only of happy customers. If a pattern is found, the store is flagged and the note sits next to the comparison. Check five, a difference outside the noise: the gap between the stores holds beyond the error. If not, it is a tie and no winner is named. A closing band says that when all five pass, each topic is shown for both stores with the reviews behind every percentage one click away. The figure contains no measured numbers.
Most unfair comparisons fail at check two. A store that asks for reviews at the till is a different sample from the one next door.
Rule six has a legal basis. Google's Maps user-generated content policy says contributions should reflect a genuine experience. It removes paid reviews, reviews posted from several accounts by one person, and content with unusual volumes or patterns meant to manipulate a rating. It also forbids posting on a competitor's page to undermine it, and it bars merchants from selectively asking only happy customers. In the United States, the FTC rule on consumer reviews and testimonials makes fake reviews and paid reviews with a required sentiment unfair or deceptive practices. It covers review suppression too.
Neither rule cleans the data for you. One store that asks every customer for a review at the till will collect more reviews than the store next door, and they will be warmer. Comparing the two without a note is comparing two different samples.
Rule eight is the discipline a brand tracker needs before it calls a movement real, which we covered in our article on brand tracker significance. A gap of a few points on a topic with few mentions is a tie, whatever the colors on the chart say.
What the comparison cannot tell you
UNIQUE INSIGHT Reviews are written by customers who chose to write. They are not a sample of everyone who walked in, and the J-shape is a reminder that the motivated extremes write most. A topic comparison tells you what the writers of each store complain about and praise. It does not tell you how many customers feel that way.
It also says nothing about sales. A store can have worse reviews on queues because it is busier. When the decision is about investment, the topic comparison points the question, and a survey or a store visit answers it. We made the same argument about building a dashboard from the decision in market research dashboards.
Where Cassi.ai comes in
cassiai.com is a software engineering company specialized in the pains of market research, innovation and insights, founded by Cassiano Albuquerque after twenty years in the field. We understand research first, and we automate research projects with or without language models.
Review Pulse is a Cassi.ai product that compares a chain's stores from public data, with hourly foot traffic, Google reputation and review topics, store by store.
Review Pulse is our dashboard for comparing physical stores across brands from public data. It puts each brand's stores side by side on hourly foot traffic, reputation and what customers write. Every review with text gets one or more topics. Every percentage shows the count of mentions behind it, and topics with fewer than 10 mentions get no percentage at all.
The collection runs without a language model, on explicit rules that can be checked. Each store is identified by its listing, the city is confirmed against the postal code, and the rating is computed from the star counts. AI comes in only at the end, over data already checked. The dashboard never shows a reviewer's name. It never compares stores with different audiences without saying so, and it never turns a tie into a ranking. The coding principle behind it is the one in the evidence chain behind qualitative coding: every claim one click from its source.
FAQ
What is competitor review analysis?
Comparing the public reviews of your stores with those of competing stores to find where each one does better or worse, in customers' own words. Done topic by topic, it shows what the rival does better, which an average cannot.
How do you compare Google reviews between two stores?
Use the same source and period for both, and read only reviews with text for topics. Apply the same topic list to both, and show the excerpts behind every percentage. A hotel benchmarking study built this kind of comparison from 26,934 reviews.
What is aspect-based sentiment analysis?
A method that finds the topics a review mentions and the sentiment attached to each one. In a 2024 comparison, the strongest method reported reached 86 to 90 percent accuracy on two benchmarks, so excerpts should stay visible.
Why is the average star rating misleading?
Because online ratings tend to a J-shaped distribution, mostly 5 stars and some 1 star, as Hu, Zhang and Pavlou showed for product reviews. An average blends every topic and every period into one number.
Should fake reviews be removed before comparing stores?
Flag unusual patterns before comparing. Google's policy removes fake engagement, and the FTC rule bans fake and paid reviews in the United States.
Before the next monthly comparison
UNIQUE INSIGHT Two star averages in a spreadsheet tell the regional team that the store next door is ahead. Two lists of topics, with the reviews behind them, tell them on what. The second costs more to set up once, and it is the only one that points to something a store manager can change on Monday.
Does your team still compare stores by copying averages from the map into a spreadsheet? Pick one store and its nearest rival, and write the eight rules first. We can set up the brands, cities and topics with your team and run the comparison from there.
About the author
Cassiano Albuquerque is the founder of cassiai.com. He has spent twenty years in market research, was ESOMAR's representative in Brazil and speaks at IIEX, ESOMAR and ABEP. He now builds AI systems for research teams. To talk about comparing your stores with the competition topic by topic, contact him on LinkedIn or see the cassiai.com portfolio.
Published by CassiAI. Read the full article at https://www.cassiai.com/blog/competitor-review-benchmarking.