1. The Committee Problem
Every expansion team knows this moment.
You've done the analysis. You've run the demographics. You've mapped the competition. The model produces a score — Site A: 74, Site B: 68, Site C: 81 — and you walk into committee ready to recommend Site C.
The CFO looks up from the deck: "Why 81?"
If your answer is "the algorithm said so," you've already lost.
The score has to survive the meeting. Not just generate numbers, but defend them. Not just predict performance, but explain why. Not just rank candidates, but tell a story that a real estate committee, investment committee, or franchise approval board will believe enough to commit capital and sign a multi-year lease.
This is the fundamental challenge of AI in site selection: the gap between what a model knows and what a team can act on.
A new unit can represent seven figures of capital and a multi-year lease commitment. Committees default to risk control — and unexplained outputs trigger risk aversion, regardless of the model's statistical accuracy.
A story that happens more often than it should
A team pilots an ML forecast that ranks a suburban site #1. The VP likes the score. The field team is skeptical — the parking looks tight and the access is awkward. In committee, the only defense is "the model learned this pattern from our historical stores." The CFO asks: "What happens if traffic drops 15%? What if the competitor across the street expands?" No one can answer. The recommendation is tabled for "more analysis." The site is leased by a competitor two weeks later.
The model didn't fail on accuracy — it failed on governance.
2. What "Explainable" Actually Means
A site score or forecast where decision-makers can trace the output back to (1) named components (reach, demand, competition, access, cannibalization), (2) documented data sources and vintages, and (3) a transparent method (weights, rules, or model explanations) — so the result can be challenged, audited, and defended in committee.
In plain terms: an explainable site score is one you can reproduce on a whiteboard (or trace step-by-step in a report): inputs → components → weights → total. If you can't show that chain, the score won't survive finance review.
Explainability vs. Justification
A committee doesn't need to understand the mathematics. It needs the decision to be:
- Reproducible: Could we rerun this next week and get the same answer with the same inputs?
- Auditable: What data did you use? When was it pulled? What were the thresholds?
- Attributable: Who signed off on the weights and criteria?
Two Flavors of Explainability
Glass-box models are intrinsically interpretable. The methodology itself is transparent — weighted scorecards, linear models, rule-based systems. You can write the logic on a whiteboard.
Black-box models with explanation layers use sophisticated ML (random forests, neural networks, gradient boosting) but add after-the-fact interpretability through techniques like SHAP or LIME (tools that show which inputs mattered most). The model itself is opaque, but explanations are generated after the fact.
Glass-box models explain how the score was calculated; explanation layers explain what the model found important. For investment committee presentations, the former is usually easier to defend.
3. The Hidden KPI: Decision Throughput
The metric that matters isn't model accuracy. It's decision throughput: how quickly your organization can evaluate, align, and approve (or kill) sites with confidence.
Every expansion team knows the pipeline pressure:
- Candidates accumulating faster than committee bandwidth
- Rework cycles when stakeholders question the analysis
- "Next meeting" delays that cost you the site
- Inconsistent criteria across evaluators
Explainability improves throughput because stakeholders can debate inputs and trade-offs, not argue about a number they can't interrogate.
When a CFO can see that Site C's score comes from strong reach (+28) offset by competitive density (-12), the conversation becomes: "Is that competition level acceptable for this market?" That's a strategic discussion. Compare it to: "Why does the model say 81?" That's an interrogation of the tool itself — and it stalls decisions.
Track: (1) Median days from site sourced → committee decision, (2) Committee cycles per site (how many "next meeting" loops), (3) % of sites approved with no rework request. If explainability cuts even one committee cycle, you can save weeks per site.
4. What Went Wrong With Black-Box AI
When machine learning first hit site selection, the pitch was seductive: feed the algorithm enough data — demographics, traffic patterns, mobile signals, sales history — and it would find patterns invisible to human intuition.
The early results were mixed.
The Trust Gap Emerged
Early adopters discovered that accuracy on paper doesn't translate to adoption in practice. When a model contradicts 30 years of market experience with no explanation, skepticism kills the recommendation regardless of its statistical validity.
The pattern repeated across the industry:
- Models would recommend sites that contradicted executive intuition
- No one could explain why the model favored Site A over Site B
- Recommendations got overruled or ignored
- The tool failed to reach sustained adoption
The Core Problem
Black-box outputs often fail governance unless translated into auditable drivers, assumptions, and trade-offs. The model may be technically correct, but if no one can explain why — and more importantly, what would have to change to flip the decision — then governance defaults to "no."
The Validation Problem
Site selection has a unique challenge: you don't know if a forecast was right for years. A store takes 12–24 months to reach maturity. By the time you can validate the model, you've already opened (or not opened) dozens of locations based on its recommendations.
This lag makes pre-commitment explainability critical — because the decision is irreversible long before the learning arrives.
5. A Brief History of Site Selection Models
6. How Explainable Site Scoring Works
Explainable site selection doesn't mean abandoning sophistication. It means structuring your analysis so that outputs can be traced back to inputs — so that every score tells a story.
The Anatomy of an Explainable Score
Consider a composite score of 74 for a candidate site:
| Component | Contribution | Why |
|---|---|---|
| Reach | +28 | 45,000 population within 10-min drive |
| Demand | +31 | Median income $87K, above target threshold |
| Competition | -12 | 4 direct competitors within trade area |
| Accessibility | +27 | Strong ingress/egress, high visibility |
| Total | 74 |
Now the conversation changes:
"Site A scores 74 because it reaches 45,000 people within 10 minutes, has above-target income demographics, but loses points for competitive density."
Stakeholders can engage meaningfully:
- "Is 4 competitors really that bad for this category?"
- "The income floor is $75K — what happens if we raise it to $85K?"
- "How does this compare to our top-performing stores?"
Component-Based vs. Monolithic
Each component has a clear definition, data source, and weight. Disagreements become discussions about specific trade-offs, not the entire methodology.
Where f is some opaque function learned from data. Even if accurate, you can't explain why any given site scored as it did without additional explanation layers.
See explainable scoring in action
Drop any address into Geod and get component-based scores with full methodology transparency.
Try Free - No Card Required→7. Glass-Box vs Black-Box Site Selection Models
| Dimension | Glass-Box | Black-Box |
|---|---|---|
| How it works | Visible formula, defined components | Learned function, opaque internals |
| CFO question | "These are the weights and inputs" | "The algorithm determined..." |
| Adjustability | Weights tunable by strategy | Requires retraining |
| Audit trail | Complete: inputs → components → score | Shows what mattered, not how it was calculated |
| Adoption rate | Generally high (teams use what they trust) | Variable (often requires translation layer) |
| Failure mode | Wrong weights (fixable, discussable) | Wrong model (opaque, hard to debug) |
Accuracy that doesn't get used is worthless. A slightly less accurate model that the whole team trusts will outperform a theoretically superior black-box that never reaches sustained adoption.
8. What the Data Shows
The key insight from Buxton's data: the lift came not just from the model's accuracy, but from the chain trusting the recommendations enough to follow them consistently. Analytics can drive lift when it's trusted and deployed consistently — an unused model, however accurate, produces zero lift.
9. Industry Approaches
Buxton: Transparent Outputs via Score Sheets
Buxton has long emphasized accessible outputs. Their site score model produces an index score (e.g., 110 meaning 10% above average store performance) with each contributing variable itemized on the score sheet. They note forecasting models typically require 51+ locations open for at least a year to build reliably.
SiteZeus: AI-Driven Forecasting + Explanation via Zeus.ai
SiteZeus describes AI-driven forecasting as a core capability, with Zeus.ai providing explanation layers that translate predictions into stakeholder-ready narratives. This "complex model + explanation layer" approach represents one path to having both sophistication and governance.
Placer.ai: Observability-First Explainability
Placer's workflows often emphasize observable visitation evidence and overlap metrics, which many teams find easier to defend than purely model-derived scores. "This site sees 20,000 monthly visits, matches our customer profile, and overlaps with our best store's trade area" — that story is built from observable facts, not learned functions.
10. The 5 Questions Every Committee Asks
If your analysis can answer these five questions clearly, it will hold up under scrutiny.
1. What's driving the score?
What they're really asking: "Show me the components so I can evaluate the trade-offs."
2. What would have to be true for this to be wrong?
What they're really asking: "How robust is this recommendation? What are the key assumptions?"
3. How does this compare to our winners and losers?
What they're really asking: "Does this site look like stores that succeeded or stores that failed?"
4. What changed since last year?
What they're really asking: "Is this data stale? Are there developments not captured?"
5. What's the downside risk to the existing network?
What they're really asking: "Are we going to cannibalize ourselves?"
11. Building Your Explainable Framework
Step 1: Define Your Components
| Question | Component | Data Sources |
|---|---|---|
| Who can you reach? | Reach / Trade Area | Drive-time isochrones, population (more on trade areas) |
| How much demand exists? | Demand | Income, households, density |
| How crowded is it? | Competition | POI counts, competitor mapping |
| Can customers get there? | Accessibility | Ingress/egress, visibility, parking |
| What about your network? | Cannibalization | Trade area overlap with existing stores |
Step 2: Make Weights Explicit
Weights encode strategy. Default weights might be: Reach (30%), Demand (30%), Competition (25%), Accessibility (15%). But these should be adjustable — ideally by users, not just data scientists.
Step 3: Show Your Work
Step 4: Document Methodology
Every analysis should include:
- Data sources (with vintage dates)
- Aggregation methodology
- Scoring formulas
- Weight justifications
- Snapshot timestamp
12. Minimum Viable Explainable Model
Not ready for a full framework? Start here. This can be implemented in a week.
4 Components: Reach (population within trade area), Demand (income/spending power), Competition (direct competitor count), Access (visibility, ingress/egress)
3 Trade Areas: 5-min drive (core), 10-min drive (primary), 15-min drive (secondary). Use peak-hour traffic, not free-flow.
1 Network Metric: Overlap % with your nearest existing store
1 Output: Score + one-paragraph narrative explaining the recommendation
That's it. You can expand later, but this baseline gives you something defensible in committee immediately.
13. What Not to Do
- Don't use different trade area definitions per analyst. Consistency enables comparison.
- Don't change weights per site to justify a favorite. Weights should reflect strategy, not outcomes.
- Don't present a composite score without the components first. Lead with the breakdown, then show the total.
- Don't cite traffic/demographics without a data vintage. "Population: 45,000" is incomplete. "Population: 45,000 (ACS 2019-2023)" is defensible.
- Don't skip cannibalization analysis for network operators. A "great" site that steals 40% of a nearby store's revenue isn't great — it's a wash.
- Don't use AI jargon with non-technical stakeholders. Say "the model" or "the analysis," not "the neural network" or "the gradient-boosted model."
14. Committee-Proof Site Selection Score Checklist
Use this checklist before presenting any site recommendation. If you can check every box, the score will withstand scrutiny.
Data & Sources
- All input data sources listed
- Data vintages documented (e.g., "ACS 2019-2023")
- Known data limitations acknowledged
Trade Area
- Trade area method specified (radius vs. drive-time)
- Time-of-day assumptions stated (if using traffic-aware isochrones)
- Geographic boundaries shown on map
Score Breakdown
- Component breakdown visible (not just total score)
- Weights documented and justified
- Each component traces to specific data inputs
Network Impact
- Cannibalization quantified (for multi-unit operators)
- Overlap with existing stores calculated
- Net new vs. transfer estimated
15. Best Practices
For Building Models
1. Start with the story — Before you open Excel, articulate the narrative you want to tell. "This site wins because..." should complete naturally from your methodology.
2. Fewer, meaningful variables — Resist the temptation to include everything. Each variable should have a clear relationship to performance.
3. Use consistent methodology — Apply the same model, weights, and data sources to every site. Consistency enables comparison.
For Presenting Results
1. Lead with the recommendation — Don't bury the conclusion. Start with "Site C is our top candidate" then unpack why.
2. Show components before composites — Let stakeholders see the parts before the whole.
3. Acknowledge trade-offs — No site is perfect. Transparency about weaknesses builds credibility.
For Organizational Adoption
1. Involve stakeholders in weight-setting — When executives help define what matters, they're invested in the output.
2. Validate against history — Run your model on past sites and compare to actual performance.
Summary
The evolution of site selection analytics follows a clear arc:
- Simple scorecards — transparent but unsophisticated
- Black-box AI — sophisticated but opaque
- Explainable AI — sophisticated and transparent
The third stage is where the industry is moving. The operators winning today aren't choosing between accuracy and explainability — they're demanding both.
When you build for explainability, you build for adoption. And adoption is what turns analytics from an experiment into an advantage.
References
- Buxton. "Understanding the 3 Primary Site Selection Model Types." buxtonco.com
- Deloitte / Harvard Business Review. (2025). "Workers Don't Trust AI. Here's How Companies Can Change That." hbr.org
- KPMG / University of Melbourne. (2025). "Trust, Attitudes and Use of Artificial Intelligence: A Global Study 2025." kpmg.com
- McKinsey & Company. (2024). "Building Trust in AI: The Role of Explainability." mckinsey.com
- SiteZeus. (2026). "SiteZeus Delivers Record Growth and Breakthrough AI Advances in Site Selection in 2025." prnewswire.com