Frequently asked questions

Does it predict sales or revenue?

No. This tool does not predict revenue, sales, foot traffic, or whether a store would succeed. It measures observable neighborhood resemblance only. It is a discovery and triage tool — a way to decide where to look, not a recommendation to open.

This scope is load-bearing. The tool makes the claim it can support (“these neighborhoods look like where you already operate”) and refuses the one it cannot (“these neighborhoods will perform”).

How accurate is OpenStreetMap store coverage?

OpenStreetMap coverage is good for large national chains but imperfect. A few stores may be missing or stale, especially for smaller or regional brands.

How this affects you:

Distinctive chains (Sweetgreen, H Mart, Whole Foods) recover their own neighborhoods to roughly the 83rd–90th percentile in out-of-sample validation, so the archetype is meaningful.
Ubiquitous chains (Starbucks) recover to roughly the 57th percentile — they operate in so many kinds of neighborhood that the archetype is diffuse and resemblance is a weak signal.
The app reports each brand’s archetype coherence score so you know how much weight to put on the results.

If OSM coverage matters for your brand, upload your own store list to re-profile against an exact footprint.

Why these features? Can you add or remove features?

The 21 features are organized into 6 thematic groups:

Income, wealth & housing cost (5 features)
Education (1 feature)
Density & transportation (4 features)
Age & household type (3 features)
Race & ethnicity (3 features)
Retail & dining context (5 features from Overture Maps)

All are visible in the detail panel and in the full method guide. The set is kept small and interpretable on purpose — the method is sensitive to which features are included, and different reasonable features would shift rankings somewhat. Trust the explanation over the exact rank.

The feature set itself is fixed, but you can reweight the 6 thematic groups (0–2× each) using the factor-weight sliders, open a group to weight an individual feature inside it, and see how results change under different scenario assumptions.

Can I upload my own store locations?

Yes. Custom upload lets you supply a CSV of lat,lon plus optional identifiers, performance, or status columns. The authenticated product saves an immutable version under a retail client, while the archetype and national re-scoring remain in the browser. You can weight toward better-performing stores and avoid lookalikes of failed locations.

What is a census tract?

A census tract is a geographic subdivision of the US defined by the Census Bureau for data collection. Roughly 83,000 tracts cover the continental United States, each with ~4 km² average area (ranging from rural to dense urban). The tool ranks tracts instead of smaller units (like parcels) because:

Census data (income, education, density, etc.) is published at tract level.
Tracts are coarse enough to be nationwide but fine enough for market discovery.

This is a market-discovery tool, not a site-level tool. For parcel-specific analysis, use custom upload to feed your exact store locations.

Why is American Community Survey data a couple years old?

The tool uses US Census ACS 2023 5-year estimates, which have a built-in lag because the Census aggregates multiple years of surveys. ACS estimates also carry sampling error, worst in small tracts.

This matters: demographic data shifts over time. If your brand’s target market is changing fast, the archetype will lag behind. The tool uses open data on purpose: every input is public, documented, and reproducible.

Can it run nationwide in the browser?

Yes. The tool scores all ~83,000 continental US census tracts live in the browser. A full re-score (e.g., when you adjust factor weights or upload custom stores) takes roughly ~30 milliseconds. No server required. Switching back to a brand you have already scored (with the same weights) is instant rather than a re-score, because each computed set of national scores is kept in an in-memory cache and reused when the inputs match.

This is possible because scoring is a lightweight standardized-distance calculation. The tract features, archetype, and scoring formula are all loaded as compact binary files (features.f32, sims.u8, etc.); the browser does the math without network calls.

Is my uploaded data private?

Yes, subject to the product’s documented access and retention controls. Source CSVs live in a separate private R2 bucket and are not reachable through /data/*. Normalized rows are client-scoped in D1, and every API request derives access from the server session. Analysts may write only explicitly granted clients; viewers are read-only. Deleting a dataset removes the private source, normalized records, and dependent analyses. Sensitive row values are not written to application logs.

How is this different from Placer, ESRI suitability scores, or Mapbox place scores?

This tool differs in three ways:

Honest positioning. It is a discovery and triage tool, not a sales or revenue predictor. It finds neighborhoods that look like your existing stores, not neighborhoods that will perform like your stores.
Transparent, explainable scoring. Every match is broken down into specific features (income, density, retail context, etc.). You can see exactly which features drove the rank and reweight them yourself via the factor-weight sliders to test different assumptions.
Runs on open data. The tool ships on Census demographic data and Overture Maps POI density. Every input is public, documented, and reproducible, so any result can be traced back to its sources.

Placer, ESRI, and similar tools excel at answering “what will this site do?” This tool answers “what neighborhoods look like the kind of place where I already operate?”

Why do percentile ranks differ across brands?

A similarity percentile is a within-brand rank. A 90th percentile tract for Trader Joe’s does not mean the same absolute level of resemblance as a 90th percentile tract for Starbucks.

Why: Each brand has its own archetype. Distinctive chains (Whole Foods) cluster tightly on certain features, so their tracts spread across a narrow band. Ubiquitous chains (Starbucks) operate in every neighborhood type, so their archetype is diffuse and high-similarity tracts are common. Percentiles reflect this: Whole Foods’ 90th percentile is more meaningful than Starbucks’ 90th percentile.

The app’s archetype coherence badge tells you how tightly each brand clusters (out-of-sample k-fold recovery score). Use that to calibrate: high coherence → trust the percentile more; low coherence → treat it as a weak signal.

How does custom upload weighting work?

Custom upload lets you add a performance column (revenue, sales, value, etc.) and a status column (closed, failed, low, etc.):

Weight by performance: The archetype uses percentile-rank weights by default, in the stated higher-is-better or lower-is-better direction. Raw currency values do not directly determine influence.
Avoid my failed locations: The tool builds a second “failure” archetype and scores tracts as:
- resemblance-to-winners − (avoidance strength) × resemblance-to-failures
- A slider controls the avoidance strength (0–100%, default 60%).

Both toggles are scenario assumptions, not learned importance. They are descriptive — the labels are your own ground truth, not a model-learned sales driver.

How are nearby chains calculated?

For any selected tract, the tool shows which of the 11 built-in brands (Trader Joe’s, Whole Foods, Sprouts, Chipotle, Starbucks, Costco, Dollar General, Tractor Supply Co., Sweetgreen, H Mart, Cracker Barrel) have a store within 6 km. This is a quick competitive-density read using existing OpenStreetMap store data, not a similarity score.

What happens if a brand has few stores?

Brands with very few stores may not produce a meaningful archetype. The tool is built for large national chains (after cleaning, Starbucks has 10,984 mapped store locations; Trader Joe’s has 603). For small regional brands, consider uploading your own store list to ensure coverage.

What does the “Strong / Moderate / Weak” fit band mean?

Alongside the similarity percentile, each neighborhood shows an absolute-fit band. The percentile is a within-brand rank, so a brand’s best tract always sits near the 100th percentile even if it is only a middling absolute match. The band fixes that: it compares the neighborhood’s raw distance to the archetype against the distances of the brand’s own store neighborhoods. Strong means it is as close as the brand’s typical store neighborhood; Moderate means it is within the range the brand already operates in; Weak means it is farther than that. Unlike the percentile, the band is comparable across brands. It is still a resemblance distance, not a viability score.

Can I analyze a brand that isn’t one of the eleven?

Yes, in two ways. You can add any United States chain from OpenStreetMap by typing its name (or picking one of the suggested chains, which attaches a Wikidata identifier for a more reliable match); the tool fetches that chain’s store points live from the Overpass API, dedupes them, and scores the country against them. Coverage depends on how completely the chain is mapped in OpenStreetMap, which the tool notes. Alternatively, upload your own store list for an exact footprint.

Can I save and switch between analyses?

Yes. Named scenarios save the current brand, weights, compare brand, focused metro, and filters, so a broker juggling several markets can switch between them. When you are signed in, scenarios, shortlists, and not-a-fit flags are stored on the server under the retail client you are working in, so they follow your account and are shared with colleagues who have access to that client. In the unauthenticated local development mode they fall back to that one browser’s local storage. The current view is also fully encoded in the URL, so a scenario is linkable as well.

Separately, any brand you add from a CSV upload or an OpenStreetMap fetch is saved in your browser and reappears in the brand picker under “Your brands.” Selecting a saved brand rebuilds its profile from the stored store points instantly, with no network refetch. Brand definitions themselves are still browser-local; a CSV you upload while signed in additionally becomes a versioned dataset stored under the retail client, which you can load again later without re-uploading. See uploading your own stores.

Yes. Your brand selection, filters, factor weights, compare brand, focused metro, pinned shortlist, and selected tract are encoded in the URL hash. A Copy link button puts the current view on your clipboard; share the URL directly and the recipient will see the same view.

What if I export a shortlist as CSV? Is there a disclaimer?

Yes. The CSV export includes a footer restating that the tool measures resemblance, not a forecast, and that high similarity does not predict revenue or store success. The export is formula-injection-safe.

See reports and shortlists for export details.