Metric Formulas and Calibration

Every TechEco DZ metric (0-10 or 0-100), with its exact formula: review signals, digital maturity, popularity, economic geography, tech-hiring potential and data confidence.

Review signals

company.review_volume · landmark.review_volume · company.rating_quality · landmark.rating_quality

Review volume. The raw Google review count is transformed, then robustly normalized within the city, where p05/p95 are the 5th/95th percentiles observed among the city's rated entities. If p95 equals p05 (no spread), every rated entity scores 5.

raw = lo g (1 + review_count)

score = 10 \cdot clamp (\frac{raw - p _{05}}{p _{95} - p _{05}}, 0, 1)

Rating quality. The raw star rating (0 to 5) is first shrunk toward the city mean by sample strength: a handful of 5-star reviews can't outrank thousands of 4.6-star ones.

posterior = \frac{n \cdot rating + 20 \cdot city_mean}{n + 20}

The result is then scaled to 0-10 through a logistic function centered on the mean:

score = \frac{10}{1 + e ^{- (posterior - mean) / scale}}

scale = max (0.15, σ_{ratings})

Missing reviews. An entity with zero reviews gets the neutral average of both scores above across the city's rated entities, flagged imputed = true with reduced confidence (0.35 instead of 1.0). An imputed estimate is never presented as an actual Google rating.

Digital maturity

company.digital_maturity

An additive 0-10 score, counted only when the declared URL is a genuinely owned domain (not a social-media page) and was successfully crawled:

digital_maturity = min (10, if crawled successfully 2.5 + 1.0 + 1.0 + 1.5 + 0.5 + 1.0 + 1.0 + min (1.5, 0.5 \cdot social_profiles))

+2.5 reachable, verified owned domain
+1.0 successful crawl
+1.0 an "about" or "services" page was found
+1.5 a "careers" page was found
+0.5 a "contact" page was found
+1.0 served over HTTPS
+1.0 crawled within the last 90 days

The social-profile bonus applies independent of the website and caps out at 3 validated profiles.

Popularity

company.public_prominence · landmark.intrinsic

The score shown everywhere as "Popularity" is a weighted average of the metrics above, specific to the entity type:

public_prominence = 0.35 \cdot review_volume + 0.25 \cdot rating_quality + 0.40 \cdot digital_maturity

intrinsic = 0.45 \cdot review_volume + 0.25 \cdot rating_quality + 0.30 \cdot economic_relevance

Because every component is still computed without reviews (see "Review signals" above), popularity is always computed for every published entity. A low score means genuinely low popularity, not missing data.

Economic geography

landmark.economic_relevance · landmark.ecosystem_relevance · company.economic_location

Economic relevance (landmarks). A fixed weight by landmark type, not a computed metric: factories and industrial zones = 10; shopping malls and business districts = 9; transit hubs, commercial hubs and markets = 8; universities and hospitals = 7; hotels, government buildings, busy roads/squares, ports, upscale residential = 6; beach resorts and airports = 5; any other type = 4.

Ecosystem relevance (landmarks). The count of companies within 3km of the landmark, robustly normalized (same p05-p95 method as review volume) across every landmark in the city. A relative measure of local economic density.

Economic location (companies). For each company, the distance-weighted sum of every landmark's popularity in the city:

raw = ℓ \in landmarks \sum popularity (ℓ) \cdot e^{- distance_km (ℓ) / 3}

A landmark 3km away contributes about 37% of its popularity, one 6km away about 13%. The result is then robustly normalized (p05-p95) across the city's companies.

Tech-hiring potential

company.tech_opportunity · company.top_role_potential · tech_calibration (0-100)

This potential is produced in two distinct, versioned stages, never from a single language-model guess.

Stage 1, assessment. A structured LangChain agent assesses all 16 controlled specialties and declares, per specialty, a model score (0-100) and a self-declared evidence level (none, weak, moderate, strong or direct). That self-declared level is then capped by a deterministic check (specialty terms searched for in the company's name, Maps category, sector/subsector and observed job postings): with no match, the level is capped to "weak" (or "moderate" for a SoftwareIT-sector company). Each specialty's final score:

final = round (prior + w_{level} \cdot (model_score - prior))

w_{level} = ⎩ ⎨ ⎧ 0 0.25 0.50 0.75 1.00 none weak moderate strong direct

An unsupported score can therefore only move partway from the sector baseline.

Stage 2, calibration. Every specialty is then re-checked against evidence actually stored for the company (job postings, crawled pages, social profiles, Maps category, community reviews — see below), never against the model's own claims. The best matching source sets a reliability and a score cap:

r_{level} = ⎩ ⎨ ⎧ 0.00 0.10 0.40 0.70 0.90 none weak moderate strong direct

cap_{level} = ⎩ ⎨ ⎧ 495974100100 none weak moderate strong direct

score = min (cap_{level}, round (prior + r_{level} \cdot q_{source} \cdot (model_score - prior)))

A specialty with no grounding at all can therefore never exceed 49/100.

Community-review signal. A signed-in user leaving a review can tag the specialties they believe the company is hiring for. That tag alone proves nothing — it only becomes evidence once 2 separate reviews name the same specialty (a "weak" cap, 59/100), or 5 separate reviews (a "moderate" cap, 74/100) — never beyond that, no matter how many reviews pile up. When a review includes a comment, a language model first verifies the comment text actually corroborates the tagged specialty (rejecting vague or contradicting comments, e.g. mentioning a layoff); a tag with no comment is counted as-is. Real, scraped evidence (a job posting, a careers page) always outranks this community signal.

Company-level score (shown as a percentage, never presented as a vacancy), where top1-3 are the three highest calibrated scores:

opportunity = 0.55 \cdot top_{1} + 0.25 \cdot top_{2} + 0.12 \cdot top_{3} + 0.08 \cdot breadth

breadth = min (100, 25 \cdot n_{grounded \geq 60})

company.tech_opportunity (out of 10) is this same opportunity divided by 10; company.top_role_potential is the highest-ranked specialty's calibrated score.

Hiring activity

company.hiring_activity · company.career_infrastructure

Hiring activity. A single stale posting decays toward 0; several recent postings push the score toward 10. Capped at 10.

score = 10 (1 - e^{- postings /2}) \cdot freshness

freshness = e^{- days_since_last_seen /90}

Career infrastructure. +6 if a distinct "careers" page was found on the site, +4 more if at least one active posting was observed. The maximum (10) requires both.

Data confidence

company.data_confidence · company.entity_quality_risk

Data confidence, an internal quality signal, not a popularity score:

confidence = 10 (0.25 reviews_present + 0.25 website_crawled + 0.25 verification_confidence + 0.15 has_coordinates + 0.10 tech_evidence_weight)

where reviews_present is 1 if reviews are not imputed (else 0.35), and tech_evidence_weight reuses the reliability scale above for the company's best tech-evidence level.

Entity quality risk, an internal QA-only signal, never shown publicly. If two or more published companies share the exact same rounded coordinates (7 decimal places, sub-meter precision, a strong signal of a duplicate or miscoded location):

risk = min (10, 3 + count_sharing_that_point)

Otherwise the risk is 0.