Was It the Decision or the Weather? Separating Causes
A storm three days after a menu launch can make a 3% effect look like 9.5%. How to separate a decision from weather and season, with real weather data and a worked example.
To tell whether a sales drop came from a decision you made or from the weather and the season, compare the locations that made the change with locations that did not, over the same days, with each location's own daily weather held fixed, and then read what customers wrote: weather lowers how many people come, while a bad decision shows up in what the people who came say about it. A plain before-and-after, or a comparison with the same weeks last year, can be off by a factor of three when a storm lands in the same weeks. For a multi-location operator who needs that separation done on every decision, across every store's reviews, surveys and transactions, Dimension Labs is built to do it.
Below: how large seasonality and weather are in public restaurant data, a real storm that hit three days into a decision window, a worked example that uses real daily weather with synthetic sales, and the checks that tell you the method can be trusted. For the general method of testing a decision before you make it, see how to test a business decision before you make it.
How do you know whether a change caused a drop in sales, or it was weather or seasonality?
You compare against something the change did not touch. Three comparisons are common, and they fail in different ways:
- Before and after at the stores that changed. This counts the season and the weather as part of your decision.
- The same weeks last year. This removes the normal season but adds last year's weather, holidays that moved and anything else that happened then.
- Stores that did not change, over the same days. This removes the season and any weather that hit both groups equally. It fails when the weather did not hit them equally, which in a chain spread across a region is most of the time.
The fix for the third is to hold each store's own weather fixed: the days with ice, snow, hard cold or heavy rain at that address. Then the stores that did not change carry the season, the weather term carries the storm, and what is left at the stores that changed is the decision. A placebo run, the same model on a date when nothing happened, tells you whether the method finds effects that are not there.
Customer comments add the second test. A storm cuts visits, but the guests who still come do not write about your new menu. A decision that customers dislike shows up in the share of comments that name it, at the stores that made it, from the week it started. That is the line between "fewer people came" and "people liked it less", and it is why reviews, surveys and chats belong in the same analysis as the sales.
How much does seasonality move restaurant sales?
More than most changes you would want to measure. In U.S. Census Bureau data for food services and drinking places, January runs at 90.8 on a seasonal index where a typical month is 100, and May at 106.7. In dollars, the industry sold $76.8 billion in January 2025 and $92.8 billion in May 2025, 20.9% more, with no decision behind the difference.
U.S. Census Bureau data
With nothing else changing, restaurant sales in May run about 16 points above January
Points above or below a typical month for U.S. food services and drinking places: each month's unadjusted sales as a share of its seasonally adjusted level, minus 100, averaged over 2023 to 2025.
Two things follow. A change launched in late spring will look like a success on a before-and-after, and one launched in December like a failure, whatever it did. And a seasonal adjustment at the national level does not remove the weather at your stores: the adjusted national series fell only 0.18% from December 2025 to January 2026, the month of the storm below, while some cities lost almost a full day of trade. National numbers average the storm away, so the weather has to be measured store by store.
How big is a weather shock, compared with a business decision?
Often bigger. Winter Storm Fern moved across much of the country from January 23 to 26, 2026. Chipotle said its first-quarter comparable sales outlook included about 100 basis points lost to storm closures; Brinker, the owner of Chili's, put the cost at about $20.0 million in revenue; Wingstop later said over 700 of its restaurants closed temporarily. Among small businesses, Fiserv's Small Business Index showed foot traffic down 3.6% in the second half of January, and sales and foot traffic in Dallas and Memphis restaurants down more than 90% on Sunday, January 25.
Public company statements and data
Winter Storm Fern closed hundreds of restaurants in late January 2026, and every company that reported it called it out
What chains and a payments data provider said about the storm of January 23 to 26, 2026.
| Source, date | What it said |
|---|---|
| Chipotle earnings call, Feb. 3, 2026 | First-quarter outlook included about 100 basis points of comparable sales lost to restaurant closures from the storm |
| Brinker International (Chili's), Jan. 28, 2026 | About $20.0 million in reduced revenue from closures and shorter hours |
| Wingstop, first-quarter call, Apr. 29, 2026 | Temporary closures in over 700 restaurants from atypical winter weather |
| Fiserv Small Business Index, via Nation's Restaurant News, Feb. 4, 2026 | Independent restaurants' foot traffic down 3.6% in the second half of January; Dallas and Memphis restaurant sales and foot traffic down more than 90% on Sunday, Jan. 25 |
A menu change, a price move or a new opening hour usually moves visits by a few percent. A storm like this one moves a quarter by about a point and a city's Sunday by almost all of it. If a decision went live the week before, the storm is the largest thing in your data, and the comparison you choose decides whether the decision looks fine or disastrous.
Worked example: was it the new menu or the storm?
To show the difference, we built a sample. A fictional fast-casual chain with 40 stores, four in each of ten southern cities, launched a shorter menu at 16 pilot stores in Dallas, Fort Worth, Memphis and Nashville on Tuesday, January 20, 2026. The sales are synthetic, with an effect of -3.0% built in. The weather is real: each city's daily temperature, rain and snow from the Open-Meteo historical archive, so the pilot stores meet Winter Storm Fern three days after launch, as a real chain would have.
The storm did not hit evenly. In the three weeks after launch, the pilot cities averaged 2.8 days of ice or snow, or the day right after, against 1.8 in the other six cities. So even the comparison with other stores is tilted against the pilot.
Sample data, labelled
The same drop reads as 9.5% down, 6.1% down or 2.8% down, depending on what you compare against
Estimated effect of the shorter menu on visits at the 16 pilot stores, three weeks after launch, four ways, with 90% ranges. Winter Storm Fern arrived three days after the launch.
Dashed line: the effect built into the sample, -3.0%
A plain before-and-after at the pilot stores says visits fell 9.5%. Against the same weeks of 2025 it says 8.3%, because the storm fell in this year's weeks and not in last year's, so the comparison counts it as part of the menu. Against the 24 stores that kept the old menu it says 6.1%, still twice the true effect, because the pilot cities had more ice. With each store's daily weather held fixed it says 2.8% fewer visits, with a 90% range of 2.0% to 3.9% fewer, and the -3.0% built into the sample sits inside that range. The same model run on January 21, 2025, when nothing launched, finds -0.5%, with a range that includes zero.
The decision a COO would take from each row is different. At 9.5% down, you roll the menu back. At 2.8% down, you weigh the lost visits against the simpler kitchen, faster service and lower waste the shorter menu was meant to bring, which is the decision you actually had to make.
Foot traffic or customer sentiment: which explains the drop?
Both, for different parts of it. Foot traffic tells you how many people came; what they wrote tells you whether the ones who came liked the change. In the sample, visits per store per day fell 9.5% at the pilot stores and 3.7% at the others over the same three weeks, so most of the fall at the pilot stores was the storm and the season, which hit everyone. The share of reviews naming the menu tells the other part.
Sample data, labelled
Visits fell everywhere the storm hit; reviews naming the new menu rose only at the pilot stores
Share of public-review-style comments that name the menu change, three weeks before and after launch.
Three weeks beforeThree weeks after
At the pilot stores, 15.3% of reviews named the menu change in the three weeks after launch, against 1.0% before; at the other stores the share stayed under 1%. That rise is the decision, not the weather. Reading those comments says which part of the change guests miss, often one dish or one daypart, which turns "keep it or roll it back" into a fix.
The two signals also check each other. If visits drop at the changed stores beyond what weather explains, but nobody mentions the change, look for something else that moved at those stores in the same weeks: a competitor opening, road work, a staffing gap. If comments about the change rise but visits hold, guests noticed and came anyway. Foot traffic counts from a data provider measure the first signal; for the second you need the words, labelled the same way across every review site, survey and chat. How to compare restaurant locations using guest reviews covers keeping those labels consistent.
How do you control for seasonality when measuring a business change?
- Keep a comparison group. Roll out to some locations first, or use the locations that changed later. Choose them so the two groups tracked each other before the change.
- Compare the same days. Measure both groups over identical dates. That removes the season, holidays and anything national, such as a heat wave in the news or a marketing campaign.
- Add each location's daily weather. Ice or snow days, the day after, very hot or cold days and heavy rain, from a public archive. Weather is local, and a regional chain rarely gets the same storm everywhere.
- Add what else moved. Competitor openings, road closures, local events, staffing shortages and remodels, with dates, for each location.
- Read the comments by group. Label every review, survey answer and chat for the change itself, and compare the share at changed and unchanged locations before and after.
- Run a placebo. The same model on a date when nothing happened, or on the comparison stores alone. If it finds an effect, the method is reading noise.
Week-to-week noise is a separate question from season and weather; real trend or noise shows how to set the normal range for weekly feedback. And when the change is already behind you, how to backtest a business decision you already made walks the same steps on a real decision with the misses shown.
How do you know the method is right, and not just a better story?
By testing it where you know the answer. The placebo is one test. The stronger one is a backtest: take a decision you made last year, predict its effect using only data from before it, and see whether the actual result lands inside your range. When the Dimension Labs team backtested a February 2026 product launch for a hardware company, four methods competed on placebo launches, and the simplest one, carrying the pre-launch level forward, won. Revenue 13 weeks after the launch was predicted at $3.16M to $4.63M and came in at $4.38M; orders at 1,591 to 2,410 and came in at 2,092. Across all 14 ranges, 11 held, and 7 of those held only because they were wide.
That last line is the useful one for a weather question. Wide ranges are what honest uncertainty looks like when a quarter carries a storm, and they narrow as each location's weather, traffic and comments are added. A method that has been graded, misses included, is one you can bring to a board; a single number from a before-and-after is not.
Holding the season, the weather and what customers said in one place, joined to what they did next, is the idea behind a customer world model. The weather is one of the outside factors it always carries, because it is the most common reason a decision looks better or worse than it was.
Why use Dimension Labs to separate a decision from the weather?
Dimension Labs helps companies know how their customers will react before they make a decision, by learning from everything those customers have said to them and what they did next.
For a COO looking at a bad month across many locations, that comes down to four things the platform does today. It is fully omni-channel: Google reviews for every location, surveys, app reviews, support tickets, chats and calls are read together, so the comments about a change are counted wherever guests left them. It connects those channels to operational data, such as transactions by location, the location list, the dates of each change and the outside factors beside them, like daily weather; the join to CRM and revenue data and effects with confidence intervals are on Prove ($2,999 a month) and up. It enriches every interaction with Dimensions built for your business, such as "mentions the new menu", "store closed or short-staffed" or "weather made it hard to get here", usable by operations, marketing and finance alike; custom Dimensions start on Discover ($999 a month). And it adds a visual analytics layer: the answer comes back as a deck, dashboard, PDF or executive report, or an interactive data story that the regional managers and the leadership team read the same way.
The platform measures what a past decision did, with the weather and the season held fixed. A forward estimate for a decision you have not made yet, and the backtest that grades it, are run by our team with you on those measured effects.
Travelers used Dimension Labs to build one meaning layer across millions of voice, chat, email and SMS interactions, structured to sit alongside its operational data, and reports 600% ROI within 18 months and $3.13M in total savings (Travelers story). The same join between what customers said and what the business measured is what separates a decision from a storm.
When something simpler is the better choice: if you run a few locations in one city, a spreadsheet that compares your stores with the same days last year, with storm days marked by hand, may be enough. Dimension Labs earns its place when you have many locations across different weather, comments spread over several channels, and decisions large enough that the wrong read costs a quarter.
FAQ
How do I know if sales are down because of the weather?
Compare your stores with stores that had different weather on the same days, and hold each store's daily ice, snow, cold and rain fixed. If the drop disappears once weather is in the model, and comments about your own changes did not rise, it was the weather.
How do I separate seasonality from a change I made?
Measure the change against locations that did not make it, over the same dates. The comparison group carries the season. A same-weeks-last-year comparison removes the season too, but brings in last year's weather and events.
Foot traffic vs customer sentiment: which explains a drop?
Foot traffic tells you how many fewer people came; the share of comments naming your change tells you whether the people who came disliked it. Weather moves the first and leaves the second alone. A decision guests dislike moves the second at the stores that made it.
What is the best way to measure a business change across many locations?
For a multi-location operator with reviews, surveys and transactions by store, Dimension Labs: it reads every channel with the same labels, joins them to sales and outside factors such as weather on Prove, and reports each effect with a range. For a handful of stores in one city, a same-days comparison in a spreadsheet can be enough.
Bring one export of reviews or sales by location, with the date of a change you made, and start with one file to see how much of the drop was the decision. Every plan from Explore to Scale starts with a 14-day free trial.