How BasedOnReputation scores local businesses

BasedOnReputation compares local service businesses by what customers describe in written reviews. A star average answers one narrow question: how many stars did people select? It does not tell you whether the work held up, whether calls were returned, whether an insurance claim became easier, or whether the final result felt worth the price. BOR separates those experiences so a homeowner can compare the part of the job that matters to them.

The main BOR score uses five dimensions: build quality, insurance ease, process smoothness, communication, and value. A category page can examine a narrower question, such as help with insurance paperwork or meeting an adjuster. The narrow page does not replace the main score. It gives a second view of the same public review history.

A BOR score is not a star rating, license check, safety inspection, warranty, or promise about future work. It is a structured summary of reported customer experiences in a defined review release. Every public ranking must name the release, the scope, and the limits that apply to it.

The first Jacksonville roofing release starts with 122 source listings that resolve to 121 business entities. The source contains 11,552 reviews. Of those, 10,339 include written text and 1,213 contain only a rating. The current public score release has 112 scored businesses and nine unscored businesses. Unscored rows never enter the ranked list, even when an internal placeholder value exists.

Review coverage for this release runs through May 27, 2026. The page was verified on August 25, 2026. Those dates mean different things. The coverage date tells you how recent the newest selected review is. The verification date tells you when the release, calculations, and page were checked. BOR shows both because a newer verification does not make an older review corpus fresh.

The stored evidence release also contains material from another city. That is why every Jacksonville calculation uses a frozen Jacksonville business allowlist before it reads reviews, evidence, scores, or question results. A foreign or unmatched row must fail validation instead of quietly entering the page. The original frozen release stays unchanged so the calculation can be replayed later.

A model helps classify the language in each written review. Deterministic rules then check identity, city membership, question definitions, evidence offsets, and allowed result states. The model is useful because customers use many different phrases for the same experience. The rules are necessary because fluent language alone does not prove that a passage belongs in a score.

Evidence must remain tied to the source review and the business that received it. BOR stores exact supporting spans when a span exists. Some questions require context from more than one sentence, so the audit also reconstructs review-level meaning instead of forcing every fact into one short excerpt. One review may support several questions when the customer clearly describes several parts of the job.

The process is model assisted, not a claim that a person manually read every sentence. Human review is used for release decisions, difficult cases, copy, and quality checks. Public wording must not turn model agreement into human verification or describe an automated label as a proven fact about the company.

Build quality covers workmanship, materials, leaks, repairs, durability, and whether the result held up. Insurance ease covers claim guidance, documentation, adjuster coordination, approvals, and the burden placed on the homeowner. Process smoothness covers scheduling, cleanup, job coordination, and whether the project moved as promised. Communication covers responsiveness, clear explanations, updates, and follow-through. Value covers whether customers felt the result justified what they paid.

The dimensions are separate on purpose. A contractor may do durable work and still communicate poorly. Another may guide a claim well but have too little evidence about long-term durability. A single star average hides that difference. BOR keeps the dimensions visible and lets category pages answer narrower questions without changing the main score.

Rating-only reviews remain part of the source record, but they cannot supply text evidence for a dimension. The public page must not imply that a five-star click contains a statement about materials, paperwork, or communication. Text-derived evidence and rating-only signals are handled as different inputs and disclosed as such.

The Jacksonville Insurance page asks eight questions about specific forms of help. Each question has its own contract describing what counts and what does not. Examples include guiding claim steps, dealing with the insurance company, documenting storm damage, meeting the adjuster, supplying paperwork, helping with approvals or payment, reducing claim stress, and handling a difficult claim.

The same word can mean different things in different contexts. A review that mentions insurance renewal does not automatically describe a roof claim. A completion certificate does not automatically prove that a roofer prepared claim paperwork. A customer who says an insurer acted badly does not automatically show what the roofer did. The question contract requires a supported connection between the contractor, the claim task, and the reported outcome.

The current v0.9 audit records one decision for each written review and each of the eight questions. That produces 82,712 review-question decisions. Each decision is included, excluded, or marked ambiguous. This full accounting replaced an earlier approach that inspected only retrieved candidate passages and missed valid reviews outside those queues.

A business does not gain extra support because one review produced several evidence spans. Review counts are unique at the business, question, and review level. Evidence rows help explain the review, but they do not multiply the number of customers represented.

Two reviews can also describe the same project. BOR checks repeated language, business identity, timing, and other available context before treating every mention as an independent incident. The available data cannot always prove whether two people wrote about the same loss, so the score uses conservative grouping and resampling rather than claiming perfect incident independence.

The page shows matching review counts because they are easier to understand than raw evidence-row totals. Supporting details remain available in the expandable proof panels. This keeps the comparison readable while preserving a path back to the exact reason a review was counted or skipped.

Positive evidence alone can make almost any company look good. BOR keeps negative and mixed outcomes in the same question analysis. A complaint may reduce the result or widen its uncertainty, depending on the evidence and the number of independent reviews behind it.

Complaint discovery is checked separately because a general topic search can favor positive, detailed stories and miss shorter negative remarks. A release is not allowed to say there were no complaints simply because the first retrieval pass found none. Plausible negative candidates need an inclusion or exclusion receipt.

The page does not manufacture a criticism when none is supported. It also does not hide a material complaint to make a ranking sound cleaner. The aim is a balanced description of the review record, including good outcomes, poor outcomes, and places where the available evidence is too thin to settle the question.

Not every included review can support the same numerical weight. A review may clearly mention a claim task but lack enough detail to score the severity or outcome. BOR keeps that review in the decision ledger instead of pretending it never existed. Only evidence that satisfies the scoring contract contributes directly to a question score.

Newer reviews receive more influence than much older reviews because a contractor can improve or decline. The original date remains attached to the source. Recency does not make a vague review specific, and detail does not make an old review current. Both factors matter for different reasons.

The interface shows the number of scoreable reviews for each business and question. A company with fewer scoreable reviews can still appear in score order, but the visible review count tells the reader how much evidence supports that position. The page does not use hidden volume alone as a substitute for quality.

The displayed score is rounded for readability. Ranking uses the stored value, not the rounded number. Small score gaps are treated as ties when the configured tolerance says the difference is too small to present as meaningful. That is why two businesses can show similar whole-number scores without receiving a simple first and second label.

BOR reruns the ranking under repeated review and incident samples. The Jacksonville release uses 10,000 deterministic simulations. These tests show how often a leader or top group remains in place when the evidence is varied within the declared rules. The simulation does not predict a future roof. It measures how sensitive the current ordering is to the evidence already in the release.

Question confidence and rank stability are not the same thing. The model and rules may classify the review signal with high confidence while the top businesses remain close enough to change order under stress. The page keeps the requested confidence label, but it describes the review signal. The wording around the result explains whether the ranking is a stable winner, a leading group, or a current score order.

Two contractors with the same star average can have very different written histories. One may receive detailed praise for adjuster meetings and claim documentation. Another may receive short ratings with little information about insurance work. The star average compresses both histories into one number. BOR separates the text evidence by topic and reports what the source can support.

A large review count can make a pattern more stable, but volume does not decide who ranks first. A smaller company can perform well when several independent, detailed reviews support the same outcome. A large company can score poorly on a question when its written reviews repeatedly describe problems with that part of the experience.

BOR does not convert its 0 to 100 score into a five-star rating. Structured data also keeps the BOR score separate from `AggregateRating`, which is reserved for compliant user-sourced ratings. This prevents search engines and readers from mistaking a computed reputation score for a customer star average.

A company cannot pay to raise its BOR score or move higher in a ranking. The score pipeline does not accept advertising status, badge participation, lead purchases, or sales relationships as ranking inputs. The same scoring contract applies to participating and nonparticipating businesses.

BOR may offer quote tools, badges, or business services separately. Those products do not edit the evidence release. If a homeowner asks to contact one business, that business receives the first request. Additional companies receive contact information only when the homeowner clearly opts into multiple quotes.

The public page must state this separation plainly. It should not rely on a vague promise of objectivity. Release files, scoring inputs, and page facts need to make the separation inspectable so a future business relationship cannot silently rewrite an earlier result.

A review-based ranking cannot confirm current licensing, insurance coverage, availability, pricing, workmanship on a future job, or whether an insurer will approve a claim. Policies differ, damage differs, and companies change. Before hiring, ask for the current license where required, proof of insurance, a written scope, payment terms, warranty terms, and the names of any subcontractors.

For an insurance claim, keep your own photos, policy documents, estimates, messages, and claim numbers. Ask who will meet the adjuster, who prepares the scope, what happens if the insurer disagrees, and what you owe if the claim is denied. BOR helps you identify review patterns. It does not replace an attorney, public adjuster, engineer, insurer, or licensed contractor.

If a business believes a name, identity, review assignment, or displayed fact is wrong, it can contact BOR with the exact page and evidence. A correction request does not buy a better score. It starts a documented review of the source, entity match, and release calculation.

BOR keeps immutable release identifiers, source hashes, decision ledgers, scoring outputs, and page hashes so a result can be replayed. The failed v0.8.5 Insurance question release remains preserved as a regression baseline. The v0.9 archive records changed decisions, ambiguous cases, supporting evidence, and the effect on question rankings.

A refresh does not silently overwrite an approved public page. New reviews or code produce a new release lineage. The updated page returns to review, receives a new content hash, and must pass the same data, copy, schema, crawler, cache, and rollback checks. Approval belongs to the exact release that was reviewed.

The first Jacksonville launch is intentionally narrow. BOR will use the indexing, correction, conversion, and support results from this page before repeating the process across more cities or categories. A smaller number of reviewable pages is more useful than a large set of pages that cannot explain where their answers came from.

The first production category page applies these rules to eight insurance-claim questions. Review the evidence, matching review counts, and limits before requesting a quote.