Trade Areas - the spatial substrate beneath every measurement

How it works

A gravity model, calibrated locally, blended with observed signal.

People are more likely to visit nearby places, and places near larger populations draw more visitors. The model learns the pattern of how visitor share drops with drive time for groups of similar places, then applies that pattern to every place in the country, including the many places where direct observations are too sparse to measure individually.

Inputs

Mobile device home-location panels, consented and opt-in
Census block group population (600-3,000 residents per BG)
Drive-time matrix from each BG to each place
GTFS schedule data for transit stations across 39 agencies
Place taxonomy with 45+ types, sub-typed where appropriate

Pipeline

Fit pool parameters via weighted least squares on control places. 4-level hierarchy: place type + DMA + urbanicity, with national fallback.

Auto-calibrate concentration to match observed effective number of block groups per place type. Two interpretable parameters: beta (decay) and alpha (concentration).

Score every place. Cross-join with 20 day-part cells. Apply temporal modulation per cell.

Blend up to 20% observed signal where panel coverage is strong. Model contributes at least 80% in all cases.

Pure SQL implementation. No UDFs, no external model artifacts, no ML.PREDICT. The fitting process eliminated 230 tasks from the prior pipeline.

Distance decay

Visitor share drops with drive time, but the curve differs by place type.

Grocery stores capture nearly all visitors within 5 minutes. Bars hold 50% out to roughly 2 minutes and keep a long tail. The model learns each type's characteristic curve directly from observed visitor data.

How quickly does visitor share drop with distance?

Relative visitor share vs drive time, by place type. Source: Motionworks gravity model, January 2025.

Grocery Store

Convenience Store

Hospital

Sitdown Restaurant

Bar

DECAY

A grocery store in Manhattan draws from a tighter radius than a gym in the suburbs.

The model captures these differences automatically. It does not impose one-size-fits-all assumptions on visitor behavior.

Different places, different reach

Median drive-time reach varies 4x across place types.

Solid bars show the drive time within which 50% of visitors live. The model recognizes that a laundromat serves its immediate neighborhood while an outlet mall draws from across a region.

Median visitor travel distance, by place type

Drive time within which 50% of visitors live. Solid bars only. The 90th percentile tail extends much further.

Destination places (20+ min)

Everyday retail (13-19 min)

Neighborhood services (under 13 min)

Temporal variation

A place's trade area expands and contracts by day-part.

The model produces 20 distinct trade areas for every place, one for each day-of-week and time-of-day cell. Grocery stores show mild variation. Bars show nearly 2x between the tightest and broadest slots, because a Tuesday at 1 AM is regulars, while a Friday at 10 PM pulls from across town.

When does the trade area spread or tighten?

Concentration ratio by day-type and day-part. Values above 1.0 mean broader reach (visitors from further away). Values below 1.0 mean tighter, more local.

Broader (visitors from further) Tighter (more local)

Bars vary nearly 2x between the tightest and broadest slots. Grocery stores barely move.

A Tuesday at 1 AM is regulars from the neighborhood. A Friday at 10 PM is people traveling across town for nightlife. The model produces 20 distinct trade areas per place to capture this.

Geography shapes the trade area

Urban grocery stores draw from 10 minutes. Rural stores reach 19.

Within a single place type, the gradient from urban core to rural town nearly doubles the median trade area reach. The model fits separate parameters for each urbanicity classification rather than applying a national average.

Median trade area reach by urbanicity and place type

The model recognizes five levels of density rather than a binary urban-rural split.

Grocery Store

Convenience Store

Bar

Sitdown Restaurant

Metro-level calibration

A grocery store in LA reaches 10 minutes. In Atlanta, 15.

Within the same urbanicity class, dense coastal metros produce the tightest grocery store trade areas. Sprawling Southern metros extend nearly 50% further. Difference is learned from observed visitor data, not assumed.

Grocery store trade area reach by metro

Dot size proportional to number of control stores in DMA.

Transit station trade areas

A transit station's trade area looks like a nervous system, not a circle.

Walk into Penn Station and you could be from Long Island, central New Jersey, or Westchester. Trade areas for transit stations follow the network, not concentric distance. Within 15 minutes of a station, walking dominates and the local neighborhood prevails. Beyond that, transit travel time along corridors carves out distinctive shapes that no radius could ever capture.

NYC Subway (G train)

Bedford-Nostrand Avs

Brooklyn, NY

702 block groups

Follows G train corridor through Brooklyn and Queens.

NYC Subway (4, B, D)

Yankee Stadium

Bronx, NY

1,900 block groups

Three lines feed in. Trade area extends across Bronx, Manhattan, and into Brooklyn.

LIRR + NJ Transit + Subway

Penn Station

New York, NY

2,500 block groups

Fans into NJ via NJ Transit and Long Island via LIRR. The largest station trade area in the country.

Metra commuter rail

Ogilvie Transportation Center

Chicago, IL

429 block groups

Three Metra lines radiate northwest, north, and west into the Chicago suburbs.

Metro Rail + Metrolink

Union Station

Los Angeles, CA

373 block groups

Trade area appears as discrete islands along Metro Rail and Metrolink corridors.

15 MIN

Inside 15 minutes, walking dominates. Beyond that, the network takes over.

The model uses GTFS schedule data from 39 transit agencies to generate transit-aware travel times. Penn Station's 2,500-block-group trade area fans into NJ Transit territory and across Long Island. A radius would have caught water and air.

Specialties within a category

Two medical offices in the same building can have different catchments.

Specialists with fewer locations per metro draw from wider areas because patients have fewer nearby alternatives. Routine-visit practices like dentist offices are abundant and evenly distributed, so visitors choose the nearest one.

How far do patients travel by medical specialty?

Median visitor travel distance. Specialists with fewer locations draw from wider areas.

3.7x

A primary care office draws from 12.6 min. An oncology practice in the same metro draws from 14.3 min.

Two medical offices in the same building can have fundamentally different visitor catchments. A site selection analysis that treats all medical offices identically would overestimate the routine practice and underestimate the specialist.

Pool hierarchy

97% of places resolve to market-specific parameters.

Each place is assigned to the finest parameter pool with sufficient control places. The cascade tries DMA + urbanicity, then DMA, then urbanicity, then a national fallback. Most places land at the metro level, where the model can capture local road networks, density, and shopping behavior.

How are 2,100,321 places assigned to parameter pools?

97% of places use market-specific parameters. Only 3% fall to the national fallback.

P2 - Metro level (DMA) 1,305,572 (62%)

P3 - Urbanicity level 445,198 (21%)

P1 - Metro + Urbanicity 283,975 (14%)

P4 - National fallback 65,576 (3%)

Gravity model vs. Decision Tree

40-70% tighter, more accurate trade areas than the prior model.

The gravity model replaces a decision-tree approach validated across 53 place types. The new model produces results that match how visitors actually behave, not how a tree split happens to slice the feature space.

Grocery Store

Median visitor distance

Decision Tree

7.1 mi

→

Gravity

1.9 mi

Result: 73% tighter — captures real neighborhood draw

Laundromat

Median visitor distance

Decision Tree

8.4 mi

→

Gravity

1.6 mi

Result: 81% tighter — laundromats are hyper-local

Medical Office

Block groups in trade area

Decision Tree

639 BGs

→

Gravity

150 BGs

Result: Top-share BG goes from 0.2% (noise) to 3.2% (real signal)

Bank

Block groups in trade area

Decision Tree

30 BGs

→

Gravity

92 BGs

Result: 3x broader — better coverage of the real catchment

53 TYPES

Validated across 53 place types: 40-70% tighter trade areas than the previous model.

The decision tree model produced visually plausible but statistically loose trade areas — a 7.1-mile median visitor distance for grocery, when the real number is closer to 1.9 miles. Where the old model spread signal too thin (medical offices), the new one concentrates. Where it failed to capture true catchment (banks), the new one expands.

Geography adapts automatically

Same model, four very different shapes.

Drop the same gravity model on grocery stores in four different geographies and the trade areas adapt without any manual intervention. The model honors what's actually there: highways, water, terrain, population density.

Urban periphery

Fry's Grocery

Tucson, AZ

140block groups

Trade area skews northwest along the I-10 corridor where the population sits. The southeastern quadrant is desert.

Mountain town

Kroger

Dahlonega, GA

42block groups

Compact shape constrained by Appalachian terrain. Visitors come from valleys, not over ridges.

Gulf Coast

Rouses

Bay St. Louis, MS

49block groups

Trade area follows the coastline. Cut sharply on one side because half of the radius is the Gulf of Mexico.

Rural lowcountry

Food Lion

Moncks Corner, SC

75block groups

Broad sparse reach across rural lowcountry. Few alternatives means visitors travel further.

SAME

Same model, four very different shapes. No manual configuration.

Population, road networks, water, and terrain are all encoded in the inputs. The gravity model honors them automatically.

Signal blending

The secret sauce: gravity + observed signal, blended.

96% of places use market-specific parameters. The gravity model contributes at least 80% to every trade area, with up to 20% correction weight from observed device-panel signal where coverage is strong. The result keeps the model's smooth, complete coverage and folds in real observed structure.

The blend

blended_share = w · observed_share + (1 − w) · gravity_share

w = min(0.20, observed_stays / (observed_stays + 2000))

The model floors at 80% gravity. As observed sample grows, w approaches 0.20 — never higher.

Fry's Grocery, Tucson AZ

Model only

89block groups

Smooth, compact core. Misses the rural northwest signal entirely.

Observed only

98block groups

Noisy and patchy, but reveals a real visitor cluster in the rural NW.

Blended (final)

496block groups

Keeps the model's compact core and pulls in the observed NW signal. Best of both.

Validation

When measured within the model's drive-cap, capture is near complete.

The model finds the right block groups within its operational range. Across an 11-month holdout panel (Feb-Dec 2024 training, Jan 2025 holdout), 94-97% of places show improved accuracy versus the prior decision-tree model, across all volume tiers and place types.

Place type	P25 capture	Median capture	P75 capture
Bar	98.7%	100.0%	100.0%
Grocery Store	89.2%	95.3%	99.2%
Quick Serve Restaurant	71.6%	88.4%	98.3%

Within-cap capture rate measures the share of observed local visits the model places in the correct block group, when comparing apples-to-apples within the gravity model's drive-cap boundary.

Putting it to work

Three workflows across the Motionworks suite.

Site selection

Cannibalization risk that a radius model misses.

Two candidate locations the same distance from an existing store can have very different overlap with that store's trade area. Trade areas surface the real overlap because visitor patterns follow road networks, population clusters, and competitive geography, not perfect circles.

Audience measurement

A billboard reaches a different audience at 7 AM than at 3 PM.

Weekday morning visitors come from a broader area as commuters stop en route. Evening visitors are concentrated locally. Media planners can size the reachable audience for each daypart rather than assuming a single static footprint.

Competitive analysis

Competing stores three miles apart can draw from opposite sides of town.

Store A draws heavily from the north due to a highway interchange. Store B draws from the south where a residential development feeds foot traffic. Equal-radius circles miss this entirely.

Why this matters

Trade area accuracy is foundational. Site selection, media planning, and competitive analysis all start with the question of where visitors come from. Get it wrong and you target the wrong neighborhoods, misallocate spend, and make real estate decisions on incomplete data. The Motionworks trade area model is the spatial substrate underneath every Motionworks measurement product.

Where every visitor actually comes from.

A gravity model, calibrated locally, blended with observed signal.

Inputs

Pipeline

Visitor share drops with drive time, but the curve differs by place type.

How quickly does visitor share drop with distance?

Median drive-time reach varies 4x across place types.

Median visitor travel distance, by place type

A place's trade area expands and contracts by day-part.

When does the trade area spread or tighten?

Urban grocery stores draw from 10 minutes. Rural stores reach 19.

Median trade area reach by urbanicity and place type

A grocery store in LA reaches 10 minutes. In Atlanta, 15.

Grocery store trade area reach by metro

A transit station's trade area looks like a nervous system, not a circle.

Two medical offices in the same building can have different catchments.

How far do patients travel by medical specialty?

97% of places resolve to market-specific parameters.

How are 2,100,321 places assigned to parameter pools?

40-70% tighter, more accurate trade areas than the prior model.

Same model, four very different shapes.

The secret sauce: gravity + observed signal, blended.

Fry's Grocery, Tucson AZ

When measured within the model's drive-cap, capture is near complete.

Three workflows across the Motionworks suite.