How a match is computed
Overview
Section titled “Overview”LookAlike scores every continental US census tract by how well it matches the pattern of tracts where a brand already has stores. For each brand it fits a presence model that contrasts the brand’s store tracts with a national background of tracts. The method is reproducible; this page walks through every step, from raw data to the 0–100 fit percentile you see on the map.
The data: 18 scored features in 5 themes
Section titled “The data: 18 scored features in 5 themes”Every tract is described by 18 scored features, organized into 5 themes. Three more shares — race and ethnicity — are shown as population context only and are not model inputs.
1. Income, wealth & housing cost (5 features)
Section titled “1. Income, wealth & housing cost (5 features)”- Median household income ($)
- Per-capita income ($)
- Median home value ($)
- Median gross rent ($/month)
- Poverty rate (%)
2. Education (1 feature)
Section titled “2. Education (1 feature)”- Bachelor’s degree or higher (%)
3. Density & transportation (4 features)
Section titled “3. Density & transportation (4 features)”- Population density (people/km²)
- Households without a vehicle (%)
- Walk/bike/transit commute (%)
- Renter-occupied housing (%)
4. Age & household type (3 features)
Section titled “4. Age & household type (3 features)”- Median age (years)
- Average household size (people)
- Households with children (%)
5. Retail & dining context (5 features)
Section titled “5. Retail & dining context (5 features)”From Overture Maps Places, aggregated per km²:
- All points of interest
- Grocery stores
- Restaurants
- Cafés & coffee
- Shopping & apparel
These features are drawn from the US Census American Community Survey (ACS 2020–2024, 5-year estimates), Census TIGER/Line 2024 tract boundaries, and Overture Maps Places. See data sources for full provenance.
Mapping stores to tracts
Section titled “Mapping stores to tracts”A brand’s store locations are mapped to the tracts that contain them. The method uses one observation per occupied tract, so co-located facilities (like a Costco fuel station and warehouse at the same location) count only once. A tract with ten of the brand’s stores counts the same as a tract with one.
Scaling features
Section titled “Scaling features”Each feature is prepared the same way for every tract.
- Log-scale skewed features. Income, per-capita income, home value, rent, population density, households without a vehicle, walk/bike/transit commute, poverty, and the five retail densities are transformed with
log1pfirst. On a log scale, the gap between a rural and a suburban tract counts as much as the gap between a suburban and a dense urban tract. - Clip to the national 1st–99th percentile. This prevents a handful of extreme tracts from distorting the scale. Each feature uses its own p1 and p99 bounds.
- Standardize against the national distribution. Each clipped value becomes a standardized score (z), centered and scaled on all scoreable tracts.
The presence model
Section titled “The presence model”For each brand, LookAlike fits an L2-regularized logistic regression — the presence model. It contrasts the brand’s store tracts (one observation per tract) with a fixed background of every fourth scoreable tract.
Each of the 18 features enters with a linear term and a squared term. The squared term lets a feature have a preferred middle range, so the model can prefer moderate density, not only “more” or “less.” The fit learns one coefficient for each term, plus an intercept.
Two limits keep the score about resemblance:
- Squared terms cannot be positive. A positive squared term would raise the score at both extremes of a feature. With this limit, a squared term can only lower the score as a tract moves away from the brand’s preferred range.
- Each feature is capped at the brand’s store range. Before fitting and scoring, each standardized feature is clipped to the range that covers the middle 95% of the brand’s store tracts (the 2.5th to 97.5th percentile). A value beyond anything the stores show earns no extra credit.
Release 2026-10-07-02 introduced both limits. In tests on all 99 brands, held-out store tracts scored at the same median percentile as before (92.0 against 91.9). Top candidates that matched on a value outside the brand’s usual range fell from 44% to 29%.
The model’s output for a tract is its linear predictor — the weighted sum of the tract’s standardized features and their squares.
Fit percentile
Section titled “Fit percentile”The linear predictor is turned into a 0–100 fit percentile, where 100 = the tract that best matches the brand’s pattern and 0 = the worst:
rank = percentile rank of the model score (descending; ties get average rank)fit = 100 × (1 − (rank − 0.5) / n_tracts)For example:
- The tract with the highest model score ranks 1st and scores ~100.
- A tract with the median score ranks ~41,000th (out of ~83,000) and scores ~50.
- Ties get average ranks, so tracts with identical scores receive the same percentile.
Fit bands
Section titled “Fit bands”The percentile answers “how does this tract rank against every other tract, for this brand?” The fit band answers a different question: “how does the tract’s score compare with the scores of the brand’s own store tracts?” The labels are Strong, Moderate, and Weak. These are within-brand descriptions, not an absolute scale shared across brands.
The band compares the tract’s score, under the default theme weights, against the brand’s own store-tract scores:
- Strong — the tract scores at least as high as 75% of the brand’s store tracts.
- Moderate — the tract scores at least as high as the lowest 5% of them.
- Weak — lower than that.
- No band — when the score or the reference is missing (a blank rather than a guess).
For a custom upload, the browser computes the reference from your own uploaded locations. Brands have different feature scales and footprint distributions, so their bands do not establish equivalent absolute fit. The band does not measure viability or predict performance.
Validation: footprint coherence
Section titled “Validation: footprint coherence”To check whether the fit score captures anything, the pipeline uses out-of-sample validation. It refits the model five times. Each time it holds out the stores in one fifth of the counties, then records where those held-out store tracts rank.
The folds are blocked by county: whole counties are held out together, rather than individual stores being split at random. A held-out store’s same-county neighbors have nearly identical demographics; if they stayed in training, the model would partly recognize the held-out tract because it had already seen its neighbors. County blocking removes that leakage.
The reported coherence is the median held-out percentile. The labels are:
- Distinctive — 93 or above. Stores sit in a recognizable kind of tract; the fit score is a meaningful signal.
- Moderate — 85 or above.
- Diffuse — below 85. Stores sit in very different kinds of tract; the fit score is a weak signal for this brand.
Coherence across the 99 curated brands
Section titled “Coherence across the 99 curated brands”Coherence ranges from 75.5 (Circle K) to 99.2 (Sephora), with a median of 92.0. Ubiquitous chains (Circle K, Walgreens, 7-Eleven, McDonald’s, CVS, Dunkin’) score lowest, because they operate almost everywhere. Specialty and mall chains (Sephora, Dick’s Sporting Goods, Sweetgreen, Shake Shack, Barnes & Noble, REI) score highest.
In a county-blocked comparison on 21 brands, the presence model ranked held-out store tracts at a median of about the 93rd percentile, against about the 72nd for the earlier distance-to-average method.
The app shows each brand’s coherence so you know how much weight to place on the fit score for that brand. A low coherence does not mean the tool is broken; it reflects that the brand operates across many neighborhood types.
Explainability: per-feature contributions
Section titled “Explainability: per-feature contributions”When you click a tract, the app computes which features drive the score. Each feature’s contribution is its linear term plus its squared term, measured against an average US tract (z = 0). The contributions add up to the score. The app reports:
- “Why it matches” — the features with the largest positive contributions.
- “What works against it” — the features with the most negative contributions.
Each feature shows the tract’s value and the brand’s typical value — the median of the brand’s store tracts. These match the Python pipeline; the browser model is validated by web/src/lib/consistency.test.ts and web/src/lib/contract.test.ts.
How theme weights work (tuning scenario assumptions)
Section titled “How theme weights work (tuning scenario assumptions)”The Factors panel has a slider for each theme’s weight. At 1× the score is the fitted model. Other weights scale each theme’s contribution before the contributions are summed. A weight of 0 turns off the theme entirely; its features never appear in explanations. A weight above 1 amplifies that theme.
Preset scenarios adjust all five theme weights as shortcuts:
- Balanced — the fitted model.
- Affluence-led — emphasizes income and education.
- Urbanism-led — emphasizes density and transportation.
- Household-led — emphasizes age and household type.
Each theme also has an advanced per-feature layer: you can weight a single feature within a theme. With every per-feature weight left at its default of 1, the scores are exactly the theme-weighted scores above, so the advanced layer only changes results when you move a per-feature slider.
Why race and ethnicity are not model inputs
Section titled “Why race and ethnicity are not model inputs”Race and ethnicity shares (White non-Hispanic, Hispanic/Latino, Asian non-Hispanic) are shown as population context only. They are not inputs to the fit score.
In a county-blocked test, removing race reduced held-out accuracy by less than one percentile point, but it changed about a third of each brand’s top 1,000 candidates. Ranking neighborhoods by racial resemblance to existing stores can encode discriminatory targeting. The old “exclude race & ethnicity” control is gone, because race is no longer scored.
Custom uploads: the scoring stays the same
Section titled “Custom uploads: the scoring stays the same”When you upload your own store data, the browser fits the same presence model:
- Each store is assigned to the tract polygon that contains it (read from the release file
tracts/boundaries.binwith range requests), or to the nearest tract boundary within 1 km. Stores outside the continental US coverage are left out and counted in a notice. - The browser fits a presence model on your store tracts against the same background.
- If you provide a performance column (revenue, sales, etc.), higher performers get more weight in the fit.
- If you provide a status column (closed, failed, low), the browser fits a second model on the failure tracts and ranks by
winners − (penalty strength) × failures. - The result is a 0–100 fit percentile, like the built-in brands, with a fit band computed from your own locations.
You need at least 3 non-failed stores. Winners are shown as colored dots; failures as red-ringed dots. The labels are your own ground truth, not a sales model. A custom upload can also be compared against a built-in brand; the compare view uses the upload’s plain fit percentiles, not the avoid-failures score.
Browser scoring: client-side re-implementation
Section titled “Browser scoring: client-side re-implementation”The scoring is re-implemented in JavaScript (web/src/lib/model.ts, web/src/lib/scoring.ts) so that:
- The detail panel shows per-feature contributions consistent with the map.
- Custom uploads fit the model and score in the browser. In the authenticated product, the source file, normalized rows, metric definition, tract assignments, and lineage are persisted privately so teams can return to a reproducible version.
The browser reproduces the Python pipeline’s results for all built-in brands. web/src/lib/contract.test.ts checks this in CI against a fixture exported from the pipeline. web/src/lib/consistency.test.ts checks every brand against a full release directory named by LOOKALIKE_RELEASE_DIR.
Summary: what the score means
Section titled “Summary: what the score means”A tract’s fit percentile reflects how well its observable census demographics and commercial context match the pattern of tracts where the brand already has stores. High scores mean “looks like the tracts where this brand already is.” They do not predict whether a store would be profitable, generate foot traffic, or succeed. Use the fit score as a triage tool — a way to narrow a search from tens of thousands of tracts to hundreds of candidates — and combine it with local knowledge, real-estate expertise, and on-the-ground scouting.