1. The two WBP AI principles
Principle 1 — Explicit consent required. No WBP feature calls an AI model without you opting in explicitly. Two integration modes: BYOK (bring your own API key, calls go directly from your server to the provider) and Managed AI Service (calls route through WBP's server on a paid plan). Both require a deliberate action. There is no silent AI in a default install.
Principle 2 — AI output is not trusted by default. Every AI-generated field runs through a four-stage validation pipeline before it's eligible to publish: Quality scan → Accuracy check against your WBP Knowledge Base → E-E-A-T evaluation → optional human review. AI convenience never lowers the quality bar.
2. The four models we publish for
ChatGPT (OpenAI)
Most widely used across WBP. GPT-4o-mini default, GPT-4o for depth, o1/o3-mini for reasoning. Powers Content Generation, Auto-Tagging, Planner, Quality scoring.
Claude (Anthropic)
Constitutional AI training. Lower hallucination, strong refusal behaviour. Used for technical / YMYL content and quality assessment.
Gemini (Google)
Google's AI for Google's search — a transparency-first tension. Same quality gates as any other model, no ranking-advantage claims.
Disclosure best practices
Four disclosure formats built in: LLM Summary Box widget, automatic disclosure text, AI Use Policy page, schema aiGenerated markup.
3. AI output validation pipeline
- 1
Quality scan
AI output is scored on info density, comprehensiveness, unique value contribution, and SEO-first-content detection. Failing the scan kills the row.
- 2
Accuracy check
Claims are cross-checked against your WBP Knowledge Base (your first-hand sources, docs, product specs). Contradictions are surfaced before publish.
- 3
E-E-A-T evaluation
EEAT Engine grades the output: is there evidence of experience, expertise, authoritativeness, trust. Missing signals block the row.
- 4
Optional human review
For YMYL, high-risk topics, or when the pipeline flags uncertainty, the row queues for human review before publishing.