Before you collect a million reviews to compute an average rating, check whether the platform already did it. Naver did — and it also computed something you could not have: a tally of what visitors mention.
What the aggregates contain
- Average rating, where the vertical has one.
- The distribution across ten star bands, not just the mean.
- Review and reviewer counts, plus an image-review count.
- Naver's theme analysis with counts — taste, price, atmosphere, friendliness and so on, each with how many visitors mentioned it.
- Blog review count, a Korea-specific reputation channel.
One detail request per business, billed at $0.0008, and it covers the business's entire review history.
Why the distribution beats the mean
A 4.2 average can be a business that everyone finds decent, or one that half its visitors love and a quarter find unacceptable. Those are completely different businesses and the mean cannot tell them apart.
The ten-band distribution can. In practice the shape is the more actionable signal: a bimodal distribution on a restaurant usually means inconsistency, which is a different problem from being mediocre and needs a different response.
The theme analysis, with counts
This is the part you cannot reproduce cheaply. Naver publishes what visitors mention and how often — 친절해요 63 is sixty-three visitors saying the staff were kind, not a tag somebody attached.
Deriving that yourself means collecting every review and running topic extraction over Korean text, which is a model, a validation set and a maintenance burden. Naver's version arrives as a number.
The cost argument
| Approach | Cost per business | Covers |
|---|---|---|
| Reputation aggregates | $0.0008 | The entire review history |
| 200 reviews, to compute your own | $0.09 | The 200 you paid for |
Roughly a hundred times cheaper, and more complete. For a thousand businesses that is 80 cents against ninety dollars — and the ninety-dollar version has a smaller denominator.
Where the rating is missing entirely
For clinics, academies and pharmacies Naver computes no visitor review score, so the average is null. The reviewer and image-review counts are still there, which means engagement is measurable even where a score is not.
Do not substitute zero. A table where every pharmacy in Korea is rated 0.0 is worse than a table with an honest gap.
When you still need the reviews
Three cases where aggregates are not enough:
- You need the words. Complaint taxonomies, feature requests, competitor mentions — none of that survives aggregation.
- You need a time series. Aggregates are current; reviews carry
visited_at, so only reviews show a business declining. - You need receipt-verified only. The aggregate covers all reviews; filtering to verified ones requires the rows.
The efficient shape is usually both: aggregates across the whole market to find the businesses worth looking at, then reviews for the shortlist.
