Skip to content

Evidence

Every finding in a report ends in one evidence state. The state tells the reader how much the audit can prove.

The four states

State Use when Example
VERIFIED Directly observed or safely measured. The business appears at rank 4 in the Local Pack for 21 of 49 grid points.
STRONG SIGNAL The observation is real, but the business consequence cannot be proven from public data. The business is mentioned in AI answers but rarely cited; the effect on leads is not observable.
OWNER DATA REQUIRED The question needs Search Console, analytics, CRM, call, booking, or financial data. Did non-branded clicks become qualified leads?
NOT APPLICABLE The check does not reasonably apply to this business, configuration, geography, or audit mode. Local Finder is not configured.

NOT APPLICABLE is not a synonym for "unknown" or "provider failed." A missing observation is a validity-gate problem, not an N/A check.

Provenance

Every normalized source records enough metadata to interpret and reproduce it:

{
  "id": "google-maps",
  "provider": "dataforseo",
  "type": "serp",
  "method": "standard",
  "endpoint_family": "serp",
  "surface": "google_maps",
  "collected_at": "2026-01-01T00:00:00Z",
  "collector_version": "1.0.0",
  "location_method": "coordinate",
  "language": "en",
  "device": "mobile",
  "field_completeness": 80
}

Every observation points to a source ID. The scorer rejects an evidence reference that is not in sources.

For AI observations, provenance also includes:

  • platform,
  • product surface,
  • source_type (consumer_surface or model_api),
  • method,
  • provider,
  • prompt ID and text,
  • market and language,
  • repeat number,
  • model/version where available,
  • citation behavior.

Target resolution

Before scoring visibility, the collector resolves the target business. It uses, in order of strength:

  1. place ID,
  2. domain,
  3. phone,
  4. name alias,
  5. manual verification.

The audit records match_method and match_confidence for every target observation. A confidence below 0.80 does not count as a resolved target. If the business cannot be resolved reliably, the run fails a validity gate. It never converts an identity-resolution failure into "rank not found."

Confidence

confidence is not a synonym for "how strongly the analyst feels." It is a property of the evidence:

  • a deterministic check must be verified;
  • a judgment check may be verified or strong_signal;
  • a check with no evidence cannot be scored.

The verifier compares the confidence label to the check type and the rendered report. A mismatch fails verification.

Source methods and surfaces

A surface is not a method. The audit records both.

Surface What it means
google_local_pack The local-pack block in ordinary Google results
google_maps Google Maps results at a coordinate
google_local_finder Local Finder results where configured
google_organic Ordinary organic results
google_ai_mode Google AI Mode consumer/search answer
chatgpt ChatGPT observation through a model/API provider

Local Pack, Maps, and Local Finder can order the same businesses differently. The framework records them separately and never merges their ranks into one number.

AI evidence

Mention, recommendation, and citation are separate fields:

{
  "target_mentioned": true,
  "target_recommended": false,
  "target_cited": false,
  "target_entity_accuracy": "accurate"
}

Do not collapse them into "AI visibility": "yes".

Every repeat is recorded separately. The majority rule is documented in Scoring. A model/API response is never labeled as a consumer/search surface observation.

Missing observations

A missing observation is reported as not_collected, never as rank 0. The metrics engine excludes it from rank calculations, and the validity gate checks whether core coverage is sufficient.

If a provider fails for a small subset of observations but enough valid evidence remains, the audit can still be scored and the limitation is reported. If core coverage falls below the threshold, the run is INSUFFICIENT EVIDENCE.

What never enters evidence

  • DataForSEO login, password, or Basic Authorization values.
  • Raw provider responses from a real audit.
  • Cookies, browser profiles, HAR files, or session data.
  • Customer names, emails, phone numbers, street addresses, or review text.
  • Real client geo-grid coordinates in a tracked example.
  • Search Console, analytics, CRM, or revenue exports in a public example.
  • A model's free-form prose as a substitute for a check score.

Owner data

Owner-data questions are evidence of a different kind: the business knows the answer, and the public web does not. They are reported with publicScoreImpact: none.

The report should say:

Owner data required. Search Console can show whether these impressions produced clicks.

It should not say:

The business is probably losing leads because the owner did not provide data.

What the verifier checks

The verifier independently checks:

  • source IDs and provenance,
  • evidence and confidence on every applicable check,
  • target-match confidence,
  • validity gates,
  • AI source-type separation,
  • deterministic check derivation,
  • the query-universe hash,
  • report values and labels.

If a finding cannot be traced to a source and a check, it does not belong in the report.

Scorecard family boundary

This page covers discovery. The GBP Scorecard diagnoses profile trust and action; the Website Scorecard diagnoses website conversion. Read The Scorecard Family for the ownership matrix and the no-duplicate-scoring rule.