What the score
actually measures.
A dated model of public information about a company. It does not measure workmanship, safety or who will do the best job in your home.
84 of 1719 published branch records meet the scoring minimum; 80 meet the additional ranking floor.
Nineteen inputs. Five fixed weights.
| Pillar | Weight | Score inputs |
|---|---|---|
| Reputation | 35% | bayesian rating, review volume, low star share |
| Presence | 20% | gallery, before after, process, pricing, financing, pagespeed, https mobile |
| Credibility | 20% | verified association, registration, warranty, years, designer |
| Scale | 10% | locations, showroom, state review percentile |
| Momentum | 15% | change types |
Minimum evidence before a number
Each non-Momentum pillar needs at least three observed score inputs. A composite needs at least two qualified pillars. A supported not-applicable input is removed from the possible-input count, but never lowers the three-input minimum. At least 10 Google reviews and no unresolved conflicting evidence are required for rank. Refacing has a separate peer group; adjacent suppliers and unresolved service-fit records are not scored.
Momentum is unknown for every first-edition record. It remains in the possible-input and nominal-weight coverage denominators, contributes no numerical value, and is never replaced with zero. A future history method needs its own comparable observation window; an initial crawl does not establish 180 days of history.
Unknown, zero and contradiction
A failed fetch, missing claim, unmatched registry identity or conflicting statement is unknown. Explicit supported absence or a measured zero can be an observed value. Company-stated association membership is not directory verification. State registration needs an exact official identity match and status evidence.
How the arithmetic works
Each pillar is the weighted mean of its observed normalized inputs, after the three-input gate. Reputation uses 60% rating, 25% volume and 15% low-star share. Presence, Credibility and Scale weight their inputs equally. The composite is the weighted mean of qualified pillars using the fixed weights above; unavailable pillars are excluded from its numerical denominator.
Input coverage is observed inputs divided by applicable possible inputs. Qualified-weight coverage is the sum of qualified nominal pillar weights divided by 100. Both appear beside the score. For example, qualifying Reputation and Presence covers 55% of the fixed weight; a high score with that coverage is still based on limited information.
Normalization choices
- Reputation: Bayesian rating = (25 × metro/category median + review count × rating) ÷ (25 + review count), then divided by five and multiplied by 100. Review volume = 100 × min(1, log(1+n) ÷ log(1+peer p95)). Low-star input = 100 × (1 − one/two-star share), only when all five rating bins sum to the same dated listing review count.
- Presence: gallery = distinct supported project URLs ÷ 10, capped at 100%; before/after, described design→selection→build process, published pricing and financing are 100 for observed yes and zero for explicit supported no. PageSpeed uses the actual mobile score. HTTPS and a mobile viewport are one combined input, worth 50 points each and available only when both were checked.
- Credibility: exact verified association/designation and matched registration are 100 for supported yes and zero for supported no. Workmanship warranty is years ÷ 5, capped at 100%; stated lifetime is 100. Claimed tenure is years ÷ 20, capped at 100%. An in-house designer is an observed yes/no input.
- Scale: explicitly supported locations ÷ 5, capped at 100%; observed showroom yes/no; and review-volume percentile within the frozen eligible state/category pool. Shared websites or phones never establish a location count or combine branch reviews.
These caps and transformations are modeling choices, not validated predictors of construction outcomes. Contact paths, domain age and unverified dealer programs remain descriptive; they are not extra score inputs.
Frozen peers, ties and denominators
Reference values use the published eligible full-remodeling or refacing pool with a resolved metro and at least 10 reviews. The median and p95 use R7 linear interpolation at position (N−1)p. An empty or invalid reference denominator leaves the dependent input unknown. State percentiles use 100 × (number below + half the number equal) ÷ N.
Ranks compare the one-decimal published composite and use competition ties: 1, 1, 3. Percentiles use the same midrank convention. The labeled metro, state and published-pool denominators contain qualified observed records only. They do not mean every company in that geography. Table order may remain alphabetical while the peer-pool rank is shown separately.
Sources and geography
Individual reports publish raw field values, source links, observation dates, input counts, reference medians and peer-pool denominators. Different Google place IDs retain separate review counts. New physical metro assignments use listing coordinates against Census 2024 CBSA boundaries; a metro is broader than its central city. A service-area search association is labeled and cannot override conflicting physical geography.
Method history
The October 29 records and method are preserved as a fixed edition.
October 29, 2026: the initial 21-key information-coverage preview was preserved. Method v1 explicitly adopts 19 score inputs: HTTPS/mobile is combined, and contact remains descriptive. The per-pillar three-input safeguard is retained; first-edition Momentum is unknown. Earlier review-only KC rankings are not imported into this method.
Claims, payment and commercial relationships never change inclusion, scores or rank. Public observations · Coverage and missingness