The Secret of Polarizing Restaurants

When you read restaurant reviews, you sometimes see “the best meal of my life” and “never coming back” attached to the very same place. In the star-rating distribution, 5-star and 1-star reviews pour in at the same time, while 3-star ratings are almost nowhere to be found. For diners, it can be baffling.

In Korean, these places are often described as having strong “love-it-or-hate-it” appeal. But are such restaurants really places to avoid? We conducted a full analysis of reviewer-by-reviewer star-rating standard deviation across 81,679 restaurants in Seoul and Gyeonggi, using data to examine what these opinion-dividing places really are.

For the analysis, we used weighted standard deviation (σ) with reviewer credibility applied as a weight. It is a consensus indicator calculated by giving greater weight to Gold reviewers (50+ reviews, average rating 2.5–4.2). Restaurants with σ above 1.35 are classified as “polarizing” places where differences in opinion between reviewers are clearly pronounced.

* Weighted standard deviation (σ): Unlike ordinary standard deviation, it does not treat every reviewer’s rating equally. It gives greater weight to reviewers with extensive review experience and generally balanced judgment (high credibility), then calculates the spread. Reviewers with extremely high or low average ratings receive lower weight, making this a metric designed to capture only “real differences of opinion” without noise.

Of the 81,679 restaurants analyzed, 15,935 places had σ above 1.35. That means 19.5% of the total, or 1 in 5 restaurants, falls into the polarizing category. By contrast, there were 34,853 consensus restaurants with σ below 1.0 (42.7%), while 18,281 places (22.4%) fell into the mildly divided range of 1.0–1.35.

One notable finding is that as σ rises, the weighted positive rate (WPR) declines consistently. The average WPR in the σ 0.0–0.5 range is 99.0%, but it drops to 54.9% in the 1.35–1.5 range and all the way to 50.2% above 2.0. The more opinions diverge, the harder it naturally becomes to classify a place as a great restaurant.

* Weighted Positive Rate (WPR): The share obtained by dividing the weighted sum of reviewers who gave 4 stars or more by the total weighted sum. Unlike a simple positive-rate metric, it gives more importance to positive ratings from experienced reviewers with stronger discernment. A score of 75% or higher is classified as “great restaurant,” 50% or higher as “good,” 30% or higher as “average,” and below that as “caution.”

Even so, the average WPR remained above 50% in the σ 1.35+ range. In other words, being polarizing does not mean a restaurant is bad. Divided opinion and poor quality turned out to be two separate issues.

To see the nature of polarization more clearly, we extracted only restaurants that had been classified as “great restaurants” and compared the star-rating distributions of Gold reviewers. The difference between consensus great restaurants (σ < 1.0) and polarizing great restaurants (σ > 1.35) was striking.

In consensus great restaurants, the share of 5-star ratings from Gold reviewers was 52.6%; in polarizing great restaurants, it was 49.5%. The share of top scores was therefore almost the same. But the share of 1-star ratings was 19.8% for polarizing great restaurants, compared with just 0.4% for consensus great restaurants — about a 50-fold difference.

Even more notable is the absence of a middle ground. In polarizing great restaurants, the share of 3-star ratings from Gold reviewers was only 4.7%. Compared with 11.4% for consensus great restaurants, that is less than half. The essence of polarization is “good” or “bad,” with almost nothing in between.

Which kinds of food divide opinion the most? An analysis of polarization rates by category found that Chinese food (31.0%) ranked highest. It was followed by seafood/sashimi (30.4%), gopchang/makchang (30.4%), and meat/grilled meat (29.8%). What these categories share is plenty of room for personal preference to come into play — freshness of ingredients, intensity of wok hei or char, and strength of seasoning.

By contrast, cafés (22.2%) and pastry/bakeries (22.0%) had relatively lower polarization rates. In categories where atmosphere and visual appeal influence evaluations more than flavor variation, opinions tend to diverge less.

The tags attached to polarizing restaurants were also distinctive. Among 39 tags extracted through LLM-based semantic analysis, the tags that appeared significantly more often at polarizing restaurants were “noisy” (2.5% vs 1.6%), “company dinners / groups” (10.0% vs 7.7%), and “spacious / group seating” (39.3% vs 32.2%).

By contrast, the tags that appeared markedly less often at polarizing restaurants were “hip / aesthetic” (15.1% vs 24.6%), “Instagrammable” (7.0% vs 11.6%), “cozy” (5.0% vs 9.5%), and “quiet” (6.8% vs 10.5%).

This pattern matches the findings of Stanford Graduate School of Business with remarkable precision. Polarizing great restaurants are not the kinds of spaces people visit for Instagram. They are not especially quiet or cozy, either. Instead, they are noisy, spacious, and suited to groups. What divides opinion is their strong character, not a lack of flavor.

When opinions split to extremes, it is easy to suspect manipulation. It is natural to wonder, “Did a competitor launch a 1-star attack?” But the data told a different story.

We conducted a full profile analysis of 5,181 reviewers who gave 1-star ratings to polarizing great restaurants. 39.4% were Gold reviewers. These were people with 50+ reviews and average ratings between 2.5 and 4.2 — experienced reviewers who usually leave balanced evaluations. By contrast, the share of one-review accounts suspected of attack behavior (accounts created only to leave a single 1-star review) was just 9.0%.

The contrast becomes even clearer when compared with consensus great restaurants. Among reviewers who gave 1 star to consensus great restaurants, only 2.9% were Gold reviewers, while the share of one-review accounts (suspected attack behavior) was 14.9% — actually higher than for polarizing great restaurants. The side that sees more 1-star attacks is not polarizing restaurants, but consensus great restaurants.

Reviewers who gave 1 star to polarizing great restaurants had an average of 130.5 reviews, with a median of 33. These were not people maliciously targeting one specific restaurant; they were experienced consumers who regularly write reviews. At polarizing great restaurants, 1-star reviews are not attacks — they are sincere opinions.

We also compared the share of non-discerning reviews (so-called “praise bots”). The figures were almost identical: 21.5% for consensus great restaurants and 22.5% for polarizing great restaurants. We did not observe any pattern suggesting that manipulated reviews cluster more heavily around polarizing restaurants.

A difference did emerge in the rating gap between Kakao ratings and weighted analysis scores. Consensus great restaurants had Kakao 4.44 and Score 4.66, for a gap of -0.21. Polarizing great restaurants had Kakao 4.06 and Score 4.46, for a gap of -0.40. In both groups, Kakao ratings were lower than the weighted scores — in other words, they fall into a range where Kakao ratings undervalue actual quality.

The key point is that the Kakao rating for polarizing great restaurants is 4.06. That is 0.38 points lower than the 4.44 average for consensus great restaurants. If you choose restaurants based only on Kakao ratings, polarizing great restaurants barely come into view. Hidden in the 4.0–4.2 range, they are effectively invisible if you rely on Kakao ratings alone.

Semantic analysis also confirmed the authenticity of this polarization. The share of restaurants where “rave praise” and “not recommended” appeared at the same time was 8.7% for polarizing great restaurants, versus 4.8% for consensus great restaurants — 1.8 times higher. For the same menu, some diners are delighted while others are disappointed.

Let us look at the typical profile of a polarizing great restaurant as it appears in the data. These are places with at least 10 Gold reviewers, σ above 1.5, and still classified as great restaurants (WPR 75%+).

Wa*** is a bakery. It has 20 Gold reviewers, σ 1.53, and WPR 79%. In the reviews, glowing praise such as “ridiculously crispy” coexists with criticism like “why is this ajusshi so rude?” It is a typical pattern in which the food is well regarded, but opinions split over service.

Sel*** is a burger specialist. Reviews such as “top three burgers, easily” appear alongside comments like “the bun is a bit too much”. Mo*** is a fine-dining restaurant, where “the small bites were moving in their care and creativity” intersects with “rather monotonous compared with other tasting courses”.

In semantic analysis across all three restaurants, the most frequently cited drawback was “unfriendly service” (123 mentions combined). The essence of polarization is not bad food, but character so strong that opinions split over service and atmosphere. The food itself is already validated; it is the surrounding elements that divide diners.

Judging by Kakao ratings alone, these restaurants sit around 4.0–4.1 — not eye-catching high scores. But in weighted analysis based on Gold reviewers, they are classified as great restaurants. Their ratings are dragged down because opinions are divided, and consumers pass them by because of those lower ratings.

This is the conclusion from analyzing 15,935 polarizing restaurants: polarization was not a matter of quality, but a matter of character. It was not manipulation, and it was not bad food. Their character is simply so strong that for some people they become the meal of a lifetime, while for others they feel uncomfortable.

Restaurants that everyone agrees are good are safe choices. But truly surprising experiences rarely come from places everyone agrees on. Polarizing great restaurants may fail for some diners, but the data shows that when they succeed, the satisfaction can surpass even that of consensus great restaurants.