How Do I Test a Business Decision on My Customers Before I Make It?
How a multi-location business can test a decision on its customers before making it, using past changes as natural experiments, with a worked 5% price increase on sample data.
You can know how your customers will react to a decision before you make it, instead of finding out a quarter later. To test a business decision on your customers before you make it, find the times your customers already met a version of it: a past price change in one region, a location that changed its hours, a store that closed. Measure what those customers did next against comparable locations that did not change, with weather, season, foot traffic and competitor activity held fixed, and turn the result into a range for each customer segment. Check the method on a decision you have already made, then decide and set a date to review it. We built Dimension Labs for the multi-location operator doing this: it connects reviews, surveys, tickets and chats to your transaction and location data, enriches every conversation into Dimensions, and measures cause and effect with controls on the Prove plan.
The worked example below uses synthetic sample data: a fictional fast-casual restaurant chain with 160 locations in four regions and two years of weekly data, weighing whether to raise prices 5% in the Midwest. Every sample number on this page comes from that data, and the Python is linked in the source notes.
How do I test a decision on customers before making it?
Seven steps. The first three are about framing, and most of the value is there.
1. Name the decision as a lever you can measure
"Raise prices" is not testable. "List prices up 5% on every item, in the 40 Midwest locations, in store and on delivery apps, from the first week of the quarter" is. Write down the size, where it applies, which channels and from when. The size matters because a 2% change and a 10% change do not have to behave alike, and your past moments will only cover some sizes.
2. Pick the outcome and the window before you look
Choose the number the decision is meant to move and the one it might hurt: transactions per location per week, revenue, visits per loyalty member, share of one- and two-star reviews. Pick a window, such as 13 weeks after the change. Choosing after you see the data makes it easy to find the answer you hoped for.
3. Find the moments your customers already met the lever
These are natural experiments: something changed for some customers and not for others, for reasons that had little to do with how those customers would react. In the sample, the South raised prices 4% in week 56 and the West raised them 6% in week 78, while the Midwest and Northeast held prices. A multi-location business usually has several of these on file without calling them experiments.
Where to find a past moment
Most decisions a multi-location operator weighs have already happened somewhere, on a smaller scale
The lever, the past moment that tests it, the comparison, and what to hold fixed.
| Decision you are weighing | Past moment to learn from | Compare with | Hold fixed |
|---|---|---|---|
| Raise prices 5% in one region | An earlier price change in some regions or a menu reprice | Locations that kept their prices | Weather, season, foot traffic, competitor openings |
| Close some locations | Past closures, temporary closures for remodels | Locations of the same type farther from the closed one | Season, trade-area changes, competitor openings |
| Change opening hours | Locations that already open earlier or changed hours | Locations with the old hours | Weather, local events, day of week |
| Add or raise a delivery fee | A fee change in one market or on one app | Markets and apps without the change | Weather (rain moves delivery), promotions |
| Cut a discount or loyalty perk | A past promotion ending, a perk removed in a test market | Members who never had it | Season, competitor promotions |
4. Hold fixed what else moved at the same time
A factor belongs in the comparison if it plausibly moves the outcome you care about. For a restaurant, hotel, gym or store, four almost always do: weather, season and the calendar, foot traffic (which separates "fewer people came" from "people liked it less") and competitor activity, such as a new store opening nearby. Leave out things that do not tie to an outcome, such as industry satisfaction benchmarks. That rule, and the four kinds of data it sorts, are set out in what is a customer world model.
In the sample, all three ways of reading the South's increase give a different answer. A plain before-and-after says walk-in transactions rose 3.8%, because the season was turning up when prices went up. Comparing with regions that did not change, but without the other data, says they fell 6.5%, because a foot traffic slump and new competitors hit the South in the same weeks. With weather, foot traffic and competitor openings held fixed, the estimate is a 4.0% drop (90% range 3.5% to 4.5%), which is the 4.0% drop the sample was built with.
Sample data, fictional restaurant chain
Before-and-after said walk-ins rose. With the confounders held fixed, they fell about 4%
Estimated change in walk-in guest transactions per location after the South raised prices 4%, three ways, with 90% ranges.
5. Estimate per segment, as a range
Customers do not react to a decision as one group. In the sample, loyalty members barely pulled back, walk-in guests pulled back about in proportion to the price, and delivery app orders pulled back the most. One average across them would hide the segment that decides the outcome. Report each estimate as a range, built by resampling locations and refitting, so the uncertainty travels with the number.
6. Check the method on a past decision first
Before you trust a prediction about the Midwest, see whether the same method would have predicted something you already know. Two checks:
A placebo date. Run the analysis on a week when nothing changed. In the sample, a fake 4% increase in the South at week 30 shows effects within about half a percent of zero for every segment, and every range includes zero.
A past decision held out. Use only the data from before the West's increase, learn from the South, predict the West, then compare. Walk-in and loyalty ranges held. Delivery orders fell 8.9%, outside the predicted 9.2% to 10.8%.
Sample data, fictional restaurant chain
Checked on a past decision, the method held for two segments and missed one
Using only the South's earlier 4% increase, we predicted what the West's 6% increase would do to transactions per location, then compared with what the West did.
This miss is the reason to run the check. It showed the ranges were too narrow when carried from one region to another, so every forward range on this page is widened by a factor of 1.43, the amount that would have made all three West ranges hold. Without the check, the delivery range would have gone into the decision overconfident; with it, you know how far to trust each number before you bet on it. The full version of this check, with a scorecard of predicted ranges against what happened, is in how to backtest a decision we already made.
7. Decide, then set a review date
Write down the decision, the range you expected, and the date you will compare it with what happened: for a price change, 8 and 13 weeks after. Keep comparison locations where you can. If the result lands outside the range, find out why before another decision rests on the same method.
How do I predict how customers will react to a price increase?
Use what your customers did after earlier price changes, not what they say they would do. In the sample, learning from both past increases and holding the confounders fixed, a 5% increase in the Midwest would most likely lower total Midwest revenue by 0.36%, with a 90% range from 0.13% to 0.56% lower, if the relationship we measured holds. Over 13 weeks that is about $54,061 less revenue ($20,264 to $83,976).
Sample data, fictional restaurant chain
A 5% increase in the Midwest would most likely lower revenue a little, because delivery pulls back hardest
Predicted change in revenue over 13 weeks by customer segment if the Midwest raises prices 5%, with 90% ranges, if the relationship measured on the two past changes holds.
Dashed line: no change in revenue
The reason is the mix. Delivery app orders are 41% of Midwest revenue in the sample, more than in the other regions by design of the sample, and they are the segment that pulls back hardest: about 7.4% fewer orders for a 5% higher price. Loyalty members would bring in 2.9% more revenue. The most useful output is a better-shaped decision: in this sample, a price increase on in-store and loyalty orders with delivery prices held is the version worth testing next.
Two more reads belong next to the numbers. What customers said about price in reviews and surveys after the earlier increases, counted by star band; how to tell whether customers think you are too expensive from their comments covers that. And whether the change reached repeat customers' behaviour, which needs feedback joined to loyalty or order data; see how to join customer feedback to revenue, orders and retention data.
Will closing locations hurt sales at the others?
Sometimes they help. Some customers of a closed location move to the nearest remaining one, some go to a competitor and some stop coming. To estimate the split before you close:
Find past closures, including temporary ones for remodels, flooding or staffing.
Group the remaining locations by distance from each closed one: the nearest one or two, a ring further out, and comparable locations far away as the comparison.
Compare the change in transactions at the near locations with the far ones over the same weeks, with season, weather and any competitor openings held fixed.
Where loyalty IDs or card tokens exist, follow the closed location's regulars and count how many show up at another of your locations within 90 days.
Read what those customers wrote afterwards. Reviews of the nearby location that mention the closed one, or complaints about longer waits, show whether the transfer came with a worse visit.
The result is a recapture range: of the revenue the closed location made, how much the nearby locations kept. Comparing locations fairly is the first step; how to compare restaurant locations using guest reviews shows how to group stores by location type.
When is a staged rollout or an A/B test the better choice?
When you can run one, a controlled test is stronger evidence than any natural experiment, because you choose who gets the change and nothing about the customers decided it. It is the better choice when:
The decision is easy to reverse, such as a menu price, a delivery fee or a promotion, so a few weeks of test cost little.
You have no past moment that resembles the lever: a new kind of change, or a size far outside anything you have tried.
You can pick the test locations or customers at random, or at least stagger the rollout so later regions serve as the comparison.
The decision is large enough that the cost of being wrong is bigger than the cost of waiting a quarter.
The two work together. The natural-experiment estimate tells you the likely size and which segment to watch, which sets how many locations and weeks the test needs. The test then confirms or corrects it. For decisions you cannot test, such as a closure or a whole-chain repositioning, the past-moments method is how you get evidence at all, and it is the method Dimension Labs runs on your own data. Measuring a change after it ships, with a comparison group, is covered in how to check whether a CX fix moved the number.
What does Dimension Labs do for an operator weighing a decision like this?
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 or CEO of a chain, that rests on four things the platform does today:
It is fully omni-channel: reviews for every location, surveys, support tickets, chats and calls read together, so the reaction to a past price change is counted wherever customers voiced it.
It connects those channels to operational data, such as transactions, loyalty records, location lists and the dates of past changes, with weather and other context beside them. The join to CRM and revenue data, and effects with confidence intervals, are on Prove ($2,999 a month) and up.
It enriches every conversation into Dimensions written for your business and industry, such as price complaints by what the price is about, wait times or order accuracy, so operations, finance and marketing read the same labels. Custom Dimensions start on Discover ($999 a month).
It turns the result into the deck, dashboard or executive report the decision needs, shared across the company.
The platform measures what past changes did. For the forward range and its review, steps 6 and 7 above, our team runs the analysis with you on those measured effects, and every range comes with the results of its method check, misses included, so you know how far to trust it.
Travelers measured its program the way a finance team would: 600% ROI within 18 months and $3.13M in total savings (Travelers story).
When something simpler is the better choice: if you can randomize the change and the outcome sits in one system, an A/B testing tool or your point-of-sale reports with a spreadsheet will answer it. Dimension Labs earns its place when the evidence is spread across many locations and channels and has to be joined to transactions, which is where a chain's biggest decisions usually sit.
If you have a past price change, closure or hours change on file, your customers have already shown you how they react. Dimension Labs reads that record so the next decision is made knowing the likely result, not discovered a quarter later.
FAQ
Can AI simulate how customers will react to a price change?
It can suggest a direction, but treat the size with care. A Nature study of 70 preregistered survey experiments found that GPT-4's predictions correlated strongly with the real effects but systematically overestimated their size (Ashokkumar, Hewitt, Ghezae and Willer, 2026). On price specifically, Gui and Toubia found that changing the price in a prompt also shifted other things that should have stayed fixed, and the simulated demand curves came out as an inverted U, while real people's sloped downward (arXiv, v3, November 2025). What your customers did after your own past price changes is better evidence; are synthetic customers accurate enough for pricing decisions? goes through the research in more detail.
How many past changes do I need?
One clean past change with enough comparison locations can give a usable range; two let you check one against the other, as in the sample. With none, run a small staged test first.
What is the best way to test a business decision before launch?
For a multi-location operator whose evidence sits in reviews, surveys, tickets and transactions across many locations, Dimension Labs, with the revenue join and effects with confidence intervals on Prove. For a reversible change you can randomize in one system, a controlled A/B test or staged rollout is the stronger and simpler choice.
How long after the decision should I review it?
Set the date when you decide, long enough for most customers to meet the change and for the comparison to settle: for a price change, 8 and 13 weeks is a reasonable pair. Use the same window you predicted for.
Know how your customers will react before your next price or hours change: bring one export of transactions or reviews by location with the dates of a past change, and start with one file. Every plan from Explore to Scale starts with a 14-day free trial.