Operators already trading profitably in a market have made their site choices publicly visible. Compare the parcels they selected, the road access they paid for, the co-tenants nearby, and the daytime population within their catchment. That gives your own candidates a screen based on decisions already made. It takes longer than market sizing, but it is much easier to test.
Most expansion plans begin with a market estimate: for every market on the list, the team counts households, applies category spend, assumes a share, and produces a demand figure. That calculation helps gauge how large the market could become over time, and market opportunity analysis covers it properly. But the real estate committee is deciding something else: whether that market has a parcel worth signing for.
What a revealed-preference read actually measures
Revealed preference shows the operator's choice after commitment, rather than demand an analyst estimated in advance. After a location has traded through more than one lease cycle, the evidence includes signed rent, a funded build-out, and a catchment that lasted long enough for renewal. Those commitments reflect judgment that no demand model contains.
All of those facts can be observed from outside the business. The methods differ in what they give the committee. Top-down sizing returns a market-level figure, but that figure doesn't identify a place someone can walk. A revealed-preference screen points to corridors and parcels that a broker can walk on Thursday. Site decisions evaluated on the GrowthFactor platform commonly represent $1M to $20M of committed capital per location, so the discussion is always about one specific address.
The approach is borrowed, not new. Economics took up revealed preference through Paul Samuelson's 1938 note in Economica, which said a completed choice gives stronger evidence of preference than a stated intention. Retail applied that reasoning to geography early. Reilly's Law of Retail Gravitation (1931) and Huff's probabilistic trade areas (Land Economics, 1963) modeled observed shopping behavior, not demand reported in surveys. Applying a screen to an operating fleet brings the same logic into real estate rather than applying it to shoppers.
How to choose the reference fleet
A reference fleet is the group of operating locations on which you run a screen. The most common mistake is selecting that group by category. Two brands within one category often serve different trips, so they occupy different parcels for good reason. Define the fleet by the trip your customer is making, not by the shelf holding the product.
Coffee purchased during the morning commute needs inbound-side access, a fast turn, and a queue that clears. A sit-down coffee destination can be mid-block with parking behind it. Although both fall under the same category, a screen fitted across them averages two incompatible site logics and ends up recommending parcels suited to neither. The same distinction applies to a grab-and-go pharmacy visit versus a monthly stock-up, and to a quick-service lunch versus a weekend family dinner.
Three filters keep the set useful. Retain operators whose customers arrive the same way and during the same hours as yours. For locations, require enough operating history to have renewed at least once, because a first-year store has not yet been tested by anything. Finally, make the set broad enough to cover several different corridor types. If the entire fleet sits on one road class, the model has nothing to distinguish.
Which attributes transfer, and which ones do not
Site attributes belong to two categories, and the reliability of each differs. Physical and locational facts about a parcel carry over well between operators because a turn lane functions the same way regardless of who uses it. But anything that explains how well a particular store performs remains in that operator's books, and no amount of data spending changes that.
Most reverse-engineering fails quietly in the right-hand column. A team measures ten locations, sees the same road class and co-tenant pattern, and decides those features caused the performance. But that pattern might instead reflect what the operator could lease at the time, at the rent it could carry, and in markets its supply chain already reached. Those constraints shaped the footprint just as surely as any site criterion, yet none appears on a map.
Write down which hidden variables you're choosing to set aside, and include that list in the same memo as the ranking. Assumptions in a committee packet can be challenged; buried in a screen, they're treated as findings.
Turning observed sites into a screen you can rank with
Once the reference fleet is selected, the process is mechanical. The screen turns the attributes those locations share into criteria for scoring a parcel you haven't visited. Evaluate every reference site against the same variables, identify which attributes show the least variation across the group, and use those as requirements, not preferences.
Work through it in this order.
- Geocode the reference fleet and pull the same measurements for every location. Comparisons mean nothing unless every location is measured against the same factors: road class and nearest-road volume, turn access, number of entrances, co-tenant categories within the center, drive-time catchment, resident and daytime population inside that catchment, and relative visit timing. These measures must match exactly.
- Separate the tight variables from the loose ones. Tight clustering of a variable across the fleet means operators treated it as a limit they would not cross. When the results scatter, they considered that variable negotiable. That distinction is the exercise's actual result, and it matters more than any individual composite score.
- Write the tight variables as pass or fail criteria. A parcel is quickly ruled out if it falls below the minimum daytime population within the drive time, lacks required turn access, or has no co-tenant category that has to be present. In a first pass, speed provides most of the value.
- Rank what survives on the loose variables. Parcels meeting every constraint are then ranked by the attributes the fleet deemed negotiable, because those attributes contain the remaining upside.
- Test the screen against locations you already know. If you operate in a comparable market, put your own stores through the screen and confirm that its ranking places the strongest ones near the top. A screen unable to order a portfolio you already understand isn't ready to order an unfamiliar one.
Most teams omit that final step, even though it follows the same discipline as a vendor back-test. how to back-test a site score explains how to run it properly, what score compression looks like, and why a model that gives every parcel a grade within a narrow band isn't separating anything at all.
Where a revealed-preference read misleads
A surviving fleet can lead to three specific failure modes, and correcting each costs less than making the mistake. What you can observe today is a filtered sample. The corridors where that fleet sits had different pricing when it arrived. Visit data explains shape more clearly than it explains volume.
Survivorship. A location appears in your data only if it survived long enough to be measured. Abraham Wald set out the broader form of that argument for the Statistical Research Group in 1943: aircraft returning from missions should be armored in the areas showing no damage, since planes hit there were the ones that never returned to be counted. Retail fleets are filtered in much the same way, and quickly. Coresight Research tracked 12% more US store closures and 11% fewer openings in 2025 than in 2024. Most closures needed to make a reference set honest go unrecorded, while the sites operators considered and rejected are entirely unrecorded. Public closure reporting recovers part of the first group. Adding documented closed locations to the reference set changes the tight-variable list more often than teams expect. Before relying on a closure count, check which tracker produced it, since announcement-based trackers and layoff-filing trackers report materially different totals for the same year.
Timing. A parcel may be excellent, or the operator's position may simply be old. Entering before the corridor filled in meant signing a rent and facing competition that no longer exist. Give recent openings separate weight from long-tenured stores to distinguish those situations. A store opening into today's corridor is making today's bet.
Reading foot traffic as volume. Visit data from a device panel serves as a relative signal. It reliably compares the shape of two locations, including which hours account for the trips and how dwell differs. But as an absolute count for a single store in a small market, it is much less reliable. The categories where this matters most are covered in when foot traffic data predicts a site and when it does not.
Add another check to the set. A market can clear a revealed-preference screen and still be full: a dense fleet of successful operators is also evidence that demand is already served. Apply the screen alongside a market saturation analysis, and treat strong results on both as a warning rather than confirmation.
Where GrowthFactor fits in a revealed-preference read
The method uses GrowthFactor for its measurement and scoring layers, not for the competitor's numbers. For each address, the platform geocodes the location and returns nearby businesses, demographics, foot traffic, vehicle traffic and trade zones. You can collect the same measurements across a reference fleet rather than assembling them property by property from separate tools.
Scoring turns the tight and loose variables into a screening method. GrowthFactor Enterprise evaluates candidate sites against configurable criteria, with every input and weight visible. When your reference fleet is clustered on a variable, its weight reflects the supporting evidence, and you can explain that weight when asked why. Analog matching calibrates the results against locations in your own portfolio, while market planning moves the ranked output into a pipeline so parcels are tracked as deals rather than spreadsheet rows.
One constraint sets the limit on what any vendor can honestly promise: retailers do not pool or share store sales data, so no model can show what a competitor's location earns. Your own sales history is required to build a revenue model, which is GrowthFactor Labs work and generally needs 40 or more mature stores with revenue attached. Teams below that threshold can still run the screen. They can't assign a dollar value to the parcels it ranks. A screen that ranks without forecasting is honest work, as long as it's presented that way.
Frequently Asked Questions about choosing your next market
What does it mean to choose a market by revealed preference?
Revealed preference means reading the site decisions operators already committed to instead of the demand a model estimates. A store that has traded through several lease cycles represents real capital, a signed rent, and a catchment that held. Measuring the parcels those operators chose produces a screen you can test against the ground, rather than a market-level number you cannot.
How many comparable stores do you need before the pattern is usable?
There is no established minimum in the research literature, and the figures that circulate, commonly a minimum of three analogs, trace to vendor marketing rather than to a published standard. What matters is variation: a reference set where every location sits on the same road class and the same co-tenant mix teaches a model nothing, because there is nothing to separate. Most teams we work with get a usable read once the set spans several distinct corridor types and includes locations that have traded for more than a year.
Can you see a competitor's sales when you study their locations?
No. Unit revenue, rent, build-out cost, and labor spend stay inside the operator's books. What is observable from outside is the parcel and its context: road class, turn access, entrance count, co-tenants, drive-time catchment, and relative visit patterns. Any statement about how well a competitor's store performs is an inference, and it belongs in the memo as an assumption rather than an input.
How do you avoid survivorship bias when studying stores that are still open?
You cannot remove it, so account for it. A fleet visible today is the set that survived, and the closures that would make the sample honest are largely unrecorded. Two practical corrections help: add closed locations you can document from public closure reporting, and weight recent openings separately from long-tenured ones, since a store that opened before a corridor filled in faced a different market than a store opening into it now.
How does GrowthFactor compare to Placer.ai for studying competitor locations?
Placer.ai and GrowthFactor solve adjacent halves of the problem. Placer.ai sells observed foot traffic from a large mobile device panel and is the category benchmark for visit volume and trade-area patterns at a given property. GrowthFactor takes the physical and demographic attributes of a set of locations and scores candidate sites against them, with every input and weight visible and configurable, then carries those scored sites into a deal pipeline. Teams frequently run both, using panel data as one relative signal inside a scored screen.