Culinary Variety Collapses at the Top of the Michelin Guide

What 19,460 restaurants show about specialization, and why the first version of this chart was wrong.

---

The Michelin Guide gets narrower as it gets more prestigious. Not in quality, in variety. The kitchens at the top of the guide draw on about half as many distinct cuisines as the kitchens in the middle, and almost all of that narrowing runs through a single label.

I built this as a dashboard for a visual analytics course, but the interesting part was not the charts. It was the objection I had to answer before anyone could take the charts seriously.

The obvious version of this chart is wrong

The dataset is a Kaggle snapshot of Michelin Guide listings, 19,460 restaurants across 53 countries, with each restaurant carrying an award level and a cuisine label. The straightforward move is to count distinct cuisines at each award level and plot the result.

That chart would be garbage, and it would look convincing.

Michelin lists 11,876 restaurants at the plain Selected level and 161 at three stars. If you count distinct cuisines in each group, the bigger group wins automatically. Draw ten thousand cards from a deck and you will see more suits than if you draw ten. The chart would show variety falling as stars increase, the reader would nod, and the entire pattern would be an artifact of counting.

Anyone with a stats background spots this in about four seconds. And my intended reader, a program director at a culinary institute, is exactly the kind of person who would.

Rarefaction

The fix comes from ecology, where the same problem shows up when biologists compare species counts across survey sites of different sizes. You sample every group down to the size of the smallest one, count, and repeat.

I sampled all five award tiers down to 161 restaurants, the size of the three-star tier, and re-counted across 2,000 random draws. Every tier is then measured at the same sample size, so any difference left over is a real difference in composition rather than a difference in how many rows you happened to have.

The pattern held, and it got more interesting. Variety does not fall steadily with prestige. It peaks at Bib Gourmand, Michelin's affordable-excellence award, at 67.5 distinct cuisines per 161 restaurants, then falls to 32.0 at three stars. Plain Selected restaurants sit at 55.6, below Bib Gourmand.

So the story is not "fancier means narrower." It is a hump, with the widest culinary range sitting in the value tier rather than at either end.

The second view, and what it is not

I plotted a second measure: the share of each tier held by its five most common cuisines. That runs 31.0% at Bib Gourmand up to 67.7% at three stars.

It is tempting to present two measures moving in opposite directions as independent corroboration. It is not. Both are functions of the same cuisine distribution, and they are mechanically inversely related, so any concentration pushes one down and the other up by construction. I included the second chart because it is a scale a reader can check without knowing what rarefaction is, not because it proves anything the first chart did not.

That distinction matters more than it sounds. Presenting a related measure as an independent test is a common way to make an argument look more supported than it is.

The mechanism

The third chart names the cause. One label, "Creative," goes from 0.7% of Bib Gourmand restaurants to 39.1% of three-star restaurants. That is 38.4 of the 43.5 points the residual category gives up between those two tiers.

At the top of the guide, roughly two in five kitchens are filed under a single style.

What I would do differently

Three things I did not close, and would want to before treating this as anything more than a finding.

No uncertainty on the page. I ran 2,000 draws, which means I have a distribution, and I reported point estimates anyway. Two stars sits at 32.8 and three stars at 32.0, which the chart presents as an ordering when they are almost certainly indistinguishable. The second chart even flips their order. Error bars would have cost nothing and would have been the more direct answer to the sample-size objection than the method caption alone.

Country is confounded with prestige. Three-star restaurants cluster heavily in France and Japan, and those two scenes attract exactly the labels that expand at the top. The concentration could be national rather than tied to prestige. The dataset has a country field, so this is a filter and a re-count, and I should have run it.

"Creative" is Michelin's word, not a kitchen's. It functions largely as a residual bucket for high-end tasting-menu restaurants that do not fit a national category. If that is what is happening, measured variety collapses partly by definitional absorption, and both of my measures inherit it since both count labels. For a reader advising students on specialization, this is the difference between a real signal and a classification quirk. I flagged it rather than resolved it.

The dashboard

Built in Tableau, three views on one page, sized for print and later given a phone layout. Direct labels on every bar, zero baselines visible, method stated beside the chart it applies to.

Dashboard (optimized for mobile view, excluding second view)

Dashboard (optimized for desktop view)

Data: ngshiheng, Michelin Guide restaurants, Kaggle, CC BY-NC 4.0, snapshot retrieved September 2, 2026. Underlying listings and award designations are the property of Michelin.