AI Transparency
We use AI to measure AI visibility. That creates an obligation to be transparent about what our systems do, what data they use, and where human judgment governs.
AI Models We Use
Different models for different tasks, selected by accuracy, cost, and reliability. Routed through LiteLLM for unified logging and cost tracking.
OpenAI GPT-4
LLM probe target in RevScore IQ assessments. Also powers Content Engine drafts with strong instruction-following for BLUF-formatted output. All drafts pass human review.
Google Gemini
Probe target and primary model for rule actionability scoring. Gemini Flash delivers 90%+ accuracy for content optimization analysis at a fraction of larger model costs.
Anthropic Claude
Probe target and complex analysis engine. Excels at content quality evaluation, entity verification analysis, and detailed narrative sections of RevScore IQ reports.
Perplexity
Probe target exclusively. Real-time web search and citation model reveals content accessibility issues that other engines miss. Critical for measuring retrieval-based visibility.
What AI Decides vs. What Humans Decide
AI Handles
Running LLM probes, calculating K-Scores, computing Trinity composites, crawling for technical signals, generating initial content drafts, identifying schema gaps, and aggregating entity data.
Humans Decide
Strategic recommendations, action prioritization, final content approval (RevHumanize gate), client communication, assessment tier recommendations, probe query design, and edge case resolution.
The critical principle: AI never makes business recommendations autonomously. Every recommendation is reviewed by a human analyst who validates against domain expertise. AI provides data and patterns. Humans ensure contextual appropriateness.
Four Provenance Tiers
Every data point in a RevScore IQ report carries one of these labels. Enforced at the code level across every platform module.
LIVE
Real API call made during this assessment run, with call ID logged. Standard for all active components: LLM probes, DataForSEO queries, crawl requests.
CACHED
Data from a prior run with collection timestamp displayed. Used when the current call failed or was rate-limited. You always see how old the data is.
ESTIMATE
Derived or inferred value with the calculation formula disclosed. Used when direct measurement is unavailable but a reasonable estimate exists from other data points.
UNAVAILABLE
API call failed. No data collected. No substitution made. The score component shows NULL rather than a fabricated number. We never silently fill gaps.
Anti-Hallucination Commitments
Codified in Master System Prompt v3.1 Part 12. Engineered constraints, not guidelines.
No Fabricated Scores (12.2)
If an API call fails, that component shows UNAVAILABLE. We never derive scores from other scores without independent data. A score without data is worse than no score.
No Silent Substitution (12.3)
When a data source fails, we surface the failure. DataForSEO error? Dimension shows the failure. Probe timeout? Result shows the timeout. No comfortable fictions.
No Unsubstantiated Claims (12.5)
Reports never use phrases like "studies show" or "experts agree" without citing the specific source. Forbidden phrases are auto-flagged by our integrity scanner.
Halt on Verification Failure (12.7)
If our QA system (RevProof) is unreachable, the assessment halts and an integrity alert fires. We never auto-pass unverified reports. Occasional delays beat false confidence.
The RevHumanize Quality Gate
Three-Stage Review
Every AI-generated draft passes: (1) automated quality scoring for BLUF compliance and fact density, (2) AI-powered factual accuracy review, and (3) human editorial review with absolute veto authority. Content failing any stage is returned for revision.
Disclosure
We never publish AI-generated content without disclosing AI involvement. Clients know the Content Engine produces AI drafts. Published content includes appropriate AI disclosure.
Data Handling
We only analyze publicly accessible information. No backend access, CMS, or analytics required. Assessment data retained 90 days minimum for trend tracking. Never sold, shared, or used for model training. Client-authorized integrations (GSC, GBP) use minimum necessary scopes with revocable tokens.
Transparency Updates
Reviewed quarterly. Material changes published within 7 days of implementation.
Q1 2026
Added Perplexity as fourth probe target. Added call_id logging for all LIVE data. Published anti-hallucination commitment (MSP v3.1 Part 12).
Q4 2025
Migrated content quality scoring to Gemini Flash. Updated RevHumanize to three-stage process.
Q3 2025
Initial publication. Documented GPT-4, Gemini, and Claude usage. Established data provenance tier framework.
See Transparent Scoring in Action
Every number in your live assessment carries its provenance tier label. Request a demo and see exactly where the data comes from.