5.0 from 12 reviews or 4.5 from 3,000? How to read a rating

The small-numbers trap

A rating is an average, and an average of a few numbers is fragile. Take the 5.0 from 12 reviews: one new 1-star review drops it to 4.69 (61 stars over 13 reviews). The 4.5 from 3,000 absorbs the same review and still shows 4.5. One unhappy visitor moves the small place by three tenths of a star and leaves the big one unchanged on screen.

Small samples are also where the extremes live. Twelve reviews can all come from the opening week, from the owner's friends, or from the one kind of customer the place suits best. That doesn't make the place bad. New and small places all start this way. It only means the number hasn't been tested yet.

This is common, not rare. Among the 9,778 places on our lists (snapshot of 13 October 2026), 14% have fewer than 100 reviews. In nightlife it is 26%, and in cafés 17%. Even on lists built from well-known places, about one rating in seven rests on a small sample.

Why stars pile up at 5 and 1

If everyone who visited a place rated it, you would expect most scores to bunch in the middle, with fewer at the extremes. Online ratings rarely look like that. Studying Amazon products, Nan Hu and colleagues found that almost all of them had an asymmetric, J-shaped distribution: lots of 5-star reviews, some 1-star reviews and few in between.

Their explanation is two filters. First, the people who choose to buy (or, for a traveller, to go) are mostly the ones who expected to like it, which pushes the average up. Second, among them, the delighted and the furious are the ones who bother to write, to "brag or moan", while the merely satisfied stay quiet. In a controlled experiment where every participant rated a product, the ratings formed a roughly normal, single-peaked curve, and the mean was significantly lower than the same product's Amazon average.

The lesson: an average describes the people who chose to post, not everyone who went. A 4.6 made of 5s and 1s is a polarising place, and a 4.6 made of 4s and 5s is a consistently good one. The average alone can't tell them apart. The star bars can.

Ratings creep upward over time

A second distortion works more slowly. Apostolos Filippas, John Horton and Joseph Golden looked at five online marketplaces and found average ratings rising in every one. In the marketplace they studied most closely, an online labour market, the share of workers getting a perfect 5-star rating grew from 33% to 85% in six years. The same words also earned more stars over time: clients who called the work "terrible" gave it 1.4 stars on average in 2008 and 2.4 stars in 2015.

Their evidence points to raters being reluctant to give negative feedback in public. When the same marketplace also asked clients for private feedback that the worker would never see, they were far more candid: many who were privately dissatisfied still gave a public 5 stars. The authors suggest raters don't want to harm the person they rate, out of kindness or for fear of retaliation. That study is about freelancers, not restaurants, so we don't claim its numbers apply to maps and travel sites.

Our snapshot is consistent with this. Among the places on our lists, the median rating is 4.5, 59% are rated 4.5 or higher and 19% are rated 4.8 or higher. The lists are built from Google's top search results for each city and leave out places with fewer than 20 reviews, so they lean towards established, popular spots. Within them, a 4.5 is the middle of the pack, not a standout.

What the number on Google actually is

Google says it plainly: a place's review score is the average of all ratings published on Google for that place, and after a new review it can take up to two weeks for the score to update. It is a plain average. Nothing in that description weights ratings by how many there are, how recent they are or who wrote them.

That makes the number easy to understand and easy to misread. Sort a list by the plain average and the top fills up with places that have a handful of perfect scores, for the reasons above. Evan Miller's often-cited essay "How Not To Sort By Average Rating" makes the same point with a simple example: by average alone, an item with two positive votes and no negative ones beats an item with 100 positive and one negative.

How we adjust: the Bayesian score

Our lists use a standard fix: before trusting a place's own average, we blend it with a neutral starting point. We treat every place as if it began with 300 imaginary reviews at 4.0 stars, then add its real reviews: score = (reviews × rating + 300 × 4.0) ÷ (reviews + 300). With few reviews the score stays close to 4.0. With thousands, it lands almost exactly on the place's own rating. Above 300 reviews, the place's own rating carries more weight than the starting point.

Here is what that does to five example listings. The listings are made up, and the scores are the exact results of the formula, rounded to two decimals. Listing A is there to show the arithmetic: our lists leave out places with fewer than 20 reviews altogether.

Reading the result, and what "Best" adds

The 5.0 from 12 reviews drops to last. The formula doesn't say the place is bad, only that 12 reviews can't outweigh 300 imaginary 4.0 ones. The 4.8 from 200 keeps much of its lead, because 200 reviews is real evidence. The anchor of 4.0 is our choice: it sits below the 4.5 median of our lists, so for most places on them a thin record is pulled down, never up.

The "Best" sort starts from that score and adds two smaller terms. Reach is log10(reviews + 1) ÷ 5: about 0.22 for 12 reviews, 0.46 for 200, 0.70 for 3,000 and 0.78 for 8,000, so each tenfold increase in reviews adds only 0.2. Momentum adds up to 0.15 for places gathering new reviews quickly, and reaches that maximum at around 400 new reviews a month.

Before momentum, the five listings score A 4.26, B 4.43, C 4.78, D 5.15 and E 5.07. The big, well-liked places now lead, and even the full momentum bonus would take A only to 4.41, still below C. That is a judgement call: "Best" favours places that many people have been happy with. It is one option, not the only one.

Our city lists open sorted by "Most reviewed", because that is the one order you can check yourself: the review count is printed on every card. "Best", "Highest rated", "Hidden gems" and "Trending" are one tap away, and each answers a different question.

When few reviews are still worth a try: hidden gems

A strong rating from a modest number of reviews is often exactly what you want: a newer restaurant, a small bar, a café away from the main sights. Our "Hidden gems" sort is built for that. It only considers places with 40 to 2,500 reviews, enough to rule out a circle of friends and few enough to be off the main tourist track. It scores them as rating × 2 − log10(reviews + 1), so the rating counts heavily and fame counts against.

A 30-second checklist

With a hidden gem you trade a less tested number for a less crowded place, so the reviews themselves matter more. With 40 or 200 reviews you can read most of them. Run this on any listing, on our site or anywhere else:

• Read the review count before the stars. Under about 100 reviews, treat the rating as a first impression; under 20, as an anecdote.
• Sort by newest and skim the last few. Are they recent, and do they agree with the average? Places change owners, chefs and opening hours.
• Look at the star bars, not just the average. Mostly 4s and 5s means consistent. A wall of 5s with a cluster of 1s means it depends on the day, or on who is reviewing.
• Read two or three 3-star reviews. They tend to be the most specific about what works and what doesn't.
• Compare with similar places nearby. On our lists the median shop or mall is rated 4.3 and the median restaurant 4.6, so the same score means different things.
• Check that the praise is for what you want, such as the food, the view or the room, and not for something else the place also does.
• Confirm the opening hours for the day you plan to go.