Twice a year America runs a free experiment on every restaurant in the country.

SECTION 01The experiment nobody runs

Sunset is a terrible thing to study, because it never moves alone. It drifts with the season, and so does the weather, the holidays, the tourists and the mood. Any sensible comparison of long evenings against short ones is really a comparison of July against January, and tells you nothing.

Except twice a year, when the clock jumps an hour and the sun stays exactly where it was.

On the second Sunday in March, at two in the morning, every clock in the country moves forward sixty minutes. The following Monday is the same weekday as the Monday before it, one week later, in the same season, at the same restaurants, with the same people going to them. One thing has changed: the sun now sets an hour later by the clock.

That is a controlled experiment. It runs on every restaurant in America, it runs twice a year, it is free, and as far as we can tell nobody reads the results.

So we read them.

SECTION 02What we actually measured

Nobody publishes restaurant covers. New York publishes something almost as good and far larger: every yellow taxi trip, with the dropoff timestamped to the second.[1]

So the measurement is evening taxi dropoffs in eleven Manhattan zones that are mostly restaurants and bars: the East Village, both halves of Greenwich Village, the West Village, Meatpacking, SoHo, Little Italy, the Lower East Side, TriBeCa and both halves of Clinton. Monday to Thursday only, because Friday and the weekend answer to something other than daylight. Four hours either side of dinner, four years of it.

That is 28 monthly files and 103,072,623 trips to read through, which sounds like a lot and is about four minutes of work, because the files sit on the open internet and nothing has to be downloaded to query them.

And on the other side, sunset. Not fetched, computed: the National Oceanic and Atmospheric Administration publishes the solar equations in full,[2] so a short piece of code gives sunset for any point on earth on any date. We checked ours against the US Naval Observatory for seven places and dates and it agreed every time, to the minute.[3]

Then one question. When the clock hands the evening an extra hour of light, does dinner move?

SECTION 03The first number, which means nothing

March 2024. The week before the clock change against the week after.

Fourteen minutes later, and the mix swings hard from the five-to-seven sitting toward the eight-to-ten one. It looks like an answer. We then ran the autumn change and got a smaller move in the opposite direction, which looked like confirmation.

It is not an answer, and it was not confirmation. Neither number means anything yet, because nothing here tells us how much two ordinary weeks differ from each other for no reason at all.

SECTION 04Twenty ordinary weeks

So we took twenty pairs of adjacent Monday-to-Thursday weeks across four years, chosen in April, May, June, September and October, none of them containing a clock change, and measured exactly the same thing.

Two ordinary consecutive weeks in Manhattan can move this measure by twenty-five percent with nothing whatsoever happening. A conference, a cold snap, a marathon, a week of rain.

Which means a single result inside that band is not a result. It is a week.

This is the part of the work that has no charm and decides everything. It is also the part that got skipped in our own first pass, and skipping it would have had us publish something that was half true.

SECTION 05The asymmetry

Here is every clock change against that yardstick.

Read the last column, not the first.

The two most recent spring changes moved the evening further than any of the twenty ordinary weeks, in a direction the ordinary weeks do not go. That is a real effect and it is a large one.

Every autumn change sits on top of the noise. The one in 2022 goes the wrong way entirely.

An hour of new evening light moves dinner later. An hour of lost evening light does not move it earlier.

Which is strange until you think about what the clock is actually competing with. In March, extra light opens an option and people take it: the patio works, the walk home is pleasant, there is no reason to hurry. In November the light goes away, but the thing that produced the eight o'clock habit in the first place, the end of the working day, has not moved an inch. So the habit holds and the dark just happens around it.

Gaining an hour is an invitation. Losing one is only weather.

The two earlier springs show almost nothing, and we are not going to pretend to know exactly why. Those two years also carry materially lower evening volume in these zones than 2024 and 2025 do.

SECTION 06The time zone is ninety-one minutes wide

While the sunset code was open we pointed it at something else, and it produced the fact in this piece we like most.

A time zone is a political object about a thousand miles across. The sun is not consulted. So on the same evening, on the same clock, all on Eastern time:

Marquette and Miami are one hour and thirty-one minutes apart, on the same day, reading the same clock. Take the zone's two actual edges, Marquette and Eastport, and they run 88 minutes apart in June and 75 apart in December.

A restaurant group with a house rule that the patio closes at nine is running a completely different business in those two towns, and the rule looks identical on paper in both.

There is a smaller oddity hiding in that table. Miami, the furthest south, has the earliest solstice sunset on the list. It is also further west than New York and much nearer the equator, so its evenings are steadier all year: New York has usable light after seven on 189 days of the year, Miami on 215.

SECTION 07The week the evening moves twelve minutes

The other thing the sun does not do is move at a constant rate.

In New York, over the week ending September 24, sunset moves 11.9 minutes. Over the week ending July 2, it moves 0.1 minutes. The same seven days, a hundred times the effect.

So there are weeks in spring and autumn when the evening a business sells shifts by most of a quarter hour, week on week, for six weeks running. And there are weeks in June when it is frozen. Nobody feels either one, because neither is visible from inside a single day.

SECTION 08What any of this is worth

Directly, and we would rather say it plainly: not very much on its own. A fourteen minute shift in arrival times is a real thing to know if you write schedules or set a kitchen rota, and it is worth exactly one conversation.

What it is worth is as a demonstration of the shape.

Nobody in this story had a data problem. The taxi records are free and enormous. The sunset equations have been public for decades and the answers are checkable against the Navy. The clock change happens on a published schedule. Every single input was sitting there, and the question had simply never been put to them.

And the thing that made it an answer rather than an anecdote was twenty boring weeks of control, which is not clever and takes most of the time.

SECTION 09The point

Every business is sitting inside a handful of natural experiments it never noticed.

A road closed for eleven weeks. A competitor shut for a refit. A holiday that fell on a Thursday this year and a Monday last. A price that changed on one product and not its neighbour. Each of those is a moment when something moved for a reason that had nothing to do with you, which is the cheapest causal evidence there is, because somebody else already paid for the change.

The experiments have already been run. The results are already in the building. Nobody has ever gone and read them.