How to Pick Pilot Locations Before a Chain-Wide Rollout
Group locations by type, pair similar stores, flip a coin, and use enough pairs to beat noise. Worked on the public per-store figures from our Home Depot analysis.
Group your locations by type, leave out the ones you cannot read cleanly, pair each candidate with a similar location on its recent results, and let a coin decide which store in each pair gets the change: that is how to pick pilot locations before a chain-wide rollout. Use enough pairs, spread across the types and markets you will roll out to, that the result can be told from noise. Dimension Labs gives an operator the per-location view that matching needs: every location's reviews, surveys and conversations read with the same labels and joined to its sales, so pilot and comparison stores are chosen on how they actually start.
This page is about choosing the stores before the pilot starts. Reading the result once the pilot is running, location type by location type, is covered in multi-location rollouts: what goes to every location.
How do we choose test stores before rolling out a change to every location?
Large chains do it in stages. Chipotle tested Chipotle Honey Chicken in 80 restaurants across two markets, Nashville and Sacramento, in a test announced in August 2024, using a stage-gate process meant "to listen, test, and learn from guest and employee feedback", and launched it in five countries in March 2025 after it surpassed expectations there. Two markets far apart, many restaurants in each: the pilot was built to be read.
The same five rules work for a small chain or a very large one:
Name the outcome and the read date first. Transactions, ticket, repeat visits, or the share of reviews naming the problem the change is meant to fix. Deciding afterwards invites picking whichever number looks best.
Cover every type you will roll out to. Urban, suburban, highway, mall, campus, or whatever splits your locations. A change piloted only in suburban stores says little about urban ones.
Leave out the stores you cannot read. Locations with too little data, a remodel or reopening in the window, a new competitor next door, or a manager change. They add noise you cannot separate from the change.
Do not pilot in your best stores. It is tempting to try a change where the team is strongest. In the Home Depot data below, the five best-rated stores averaged 19.7% one- and two-star reviews against 27.7% for the other fifteen. A pilot there starts eight points ahead, and a strong team often makes any change look good.
Pair, then flip a coin. Match each candidate with a similar store and let chance pick which one gets the change. That keeps the pilot group and the comparison group alike on the things you did not think to match.
Test and control stores for a restaurant pilot: how do I match them?
Match on where each store starts and how it has been moving: the outcome itself over the last several months, its trend, its size, its location type and its local market. Then check the match before the pilot starts.
Our public analysis of 24 Home Depot stores in metro Atlanta gives a real per-store starting point. It ranks every store by the share of one- and two-star Google reviews, from January 2025 to June 2026, and flags the four stores with fewer than about 70 reviews to read with caution.
From the report: Home Depot: an Atlanta 24-store review analysis
Working from that exhibit alone, the 20 stores with enough reviews pair up neatly. Sorted by their share and paired with their neighbour, the two stores in each pair start within a point of each other in eight of ten pairs. A coin flip then picks the pilot store. The pilot group averages 25.3% and the comparison group 26.1%.
Worked on the report's public per-store figures
Ten matched pairs: in eight, pilot and comparison stores start within a point of each other
The 20 stores with 70 or more reviews, sorted by share of one- and two-star reviews and paired with their neighbour. A coin flip picks the pilot store in each pair. Each store shows its share of one- and two-star reviews.
| Pair | Pilot store | Comparison store |
|---|---|---|
| 1 | 4325 New Snapfinger Woods Dr 31.5% of 127 reviews | 1032 Research Center Atlanta Dr 31.1% of 122 reviews |
| 2 | 5300 Windward Pkwy 29.5% of 78 reviews | 650 Ponce De Leon Ave NE 29.9% of 137 reviews |
| 3 | 1670 Scenic Hwy N 28.6% of 84 reviews | 4136 Jimmy Carter Blvd 28.6% of 84 reviews |
| 4 | 2295 Lawrenceville Hwy 28.4% of 148 reviews | 1200 East West Connector SW 27.8% of 115 reviews |
| 5 | 3885 Jonesboro Rd SE 27.2% of 92 reviews | 1000 Market Pl Blvd 27.4% of 95 reviews |
| 6 | 1100 Bullsboro Dr 25.8% of 89 reviews | 4343 Tilly Mill Rd 26.2% of 107 reviews |
| 7 | 2450 Cumberland Pkwy SE 24.7% of 158 reviews | 9037 Hwy 92 25.6% of 90 reviews |
| 8 | 4101 Roswell Rd 23.1% of 91 reviews | 3755 Shackleford Rd 22.8% of 79 reviews |
| 9 | 3605 Sandy Plains Rd 19.7% of 71 reviews | 875 Lawrenceville-Suwanee Rd 21.8% of 101 reviews |
| 10 | 2525 Piedmont Rd NE 14.7% of 156 reviews | 6400 Peachtree Dunwoody Rd 19.5% of 82 reviews |
Before you start, run the pilot on paper against the past. Pretend the change began six months ago, when nothing changed, and measure the pilot stores against their pairs over those months. A good match shows no "effect". If it shows one, the pairs differ in a way the matching missed, and it is far cheaper to find that now. This is the same check used to choose a method in a real backtest: for a hardware company's February 2026 product launch, four methods were scored on 11 placebo launches before any range was trusted, the simplest won, and both headline ranges held against what happened. Across all 14 ranges, 11 held, 7 of them only because they were wide, which is exactly what a check before the start should tell you: where the data is thin enough that a wide range is the honest answer. How to backtest a business decision walks through it.
Pairs should also sit in the same market where you can manage it. Weather, local events and the season then land on both stores of a pair alike; decision or weather shows how much a storm can move a short read.
How many pilot locations do you need to trust the result?
Enough that the change you hope for is bigger than what chance alone could produce. On the Home Depot figures, a six-month pilot read on reviews alone could move the share of one- and two-star reviews by up to about 24 points by chance with one pair of stores, about 16 with two, 11 with five and 8 with all ten pairs.
Computed on the report's public per-store figures
More pairs shrink the noise: from about 24 points with one pair to about 8 with ten
How far a six-month pilot could move the share of one- and two-star reviews by chance alone (90% of simulated pilots with no real effect fall inside this band), by number of matched pairs.
2 stores
4 stores
6 stores
10 stores
20 stores
Those bands are wide because reviews are a thin signal, a few dozen a store in six months. That is useful to know before you start, and there are four ways to narrow them:
More pairs. Going from one pair to ten cuts the band by about two thirds in the simulation.
A denser outcome. Daily transactions per store carry far more information than reviews, and surveys add to both. Read the change on sales and confirm the reason in what customers wrote.
A longer read. Twelve weeks is more than four; set the date in advance.
All the channels. Reviews, surveys, app feedback and support conversations, labelled the same way, give each store more comments to read.
Size the pilot to the decision. A change that is cheap to reverse can go out on a smaller pilot and be watched; a change that costs millions to roll back deserves the ten pairs. Once the pilot is chosen this way, its result carries to stores that were not in it, if the relationship measured in the pilot holds, which is what a customer world model is for. For trying a decision on your customers before any pilot at all, see how to test a business decision before you make it.
What does Dimension Labs add when you choose pilot stores?
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.
Picking pilots well needs a clean starting picture of every location, and the platform builds it in four ways. It is fully omni-channel, so each location's reviews, surveys, app feedback, calls and chats count toward its starting point. It integrates customer channels with operational data: the location list with types and markets, weekly transactions and the dates of past changes, with the join to sales and effects with confidence intervals on Prove ($2,999 a month). It enriches every interaction with Dimensions built for your business, such as "waited too long", "item unavailable" or "mentions the new process", the same way at every location and readable by operations, finance and field teams; custom Dimensions start on Discover ($999 a month). And its visual analytics layer turns the pilot plan and the read into a store-by-store dashboard, a deck for the rollout vote, PDF and executive reports and shareable links, refreshed on a schedule while the pilot runs.
Covered California uses Dimension Labs to launch and measure new self-service use cases with structured evidence; it launched one in two weeks and saw a 5x increase in chatbot self-service adoption. Measuring a change as it goes live is the same discipline a store pilot needs. See the Covered California story.
When something simpler is the better choice: with a few locations and a change you can undo in a day, try it in one store and ask the team. Dimension Labs earns its place when you run many locations, the rollout is expensive, and the evidence sits across reviews, conversations and sales.
FAQ
How do we choose test stores before rolling out a change to every location?
Group locations by type, leave out stores with too little data or unusual events, pair each candidate with a similar store, and let a coin pick the pilot in each pair. Avoid piloting only in your best stores.
How many pilot locations do you need to trust the result?
Enough that the effect you expect is larger than chance. On public Home Depot review data, a six-month read on reviews alone could swing about 24 points by chance with one pair and about 8 with ten pairs. Using sales as the outcome narrows that a great deal.
How do I match test and control stores for a restaurant pilot?
On the outcome's recent level and trend, store size, location type and local market. Then run the comparison over a past period with no change; matched pairs should show no effect.
What is the best way to plan a pilot across many locations?
For a chain with reviews, surveys and sales across many locations, Dimension Labs: it builds each location's starting point from every channel and, on Prove, measures the pilot against its comparison stores with a range. For a few stores, a spreadsheet and a coin will do.
Planning a pilot? Start with one file of reviews or sales by location and see which of your stores start alike. Every plan from Explore to Scale starts with a 14-day free trial.