PDP Compliance Checker
Pilot · Function: Marketing · Content · Legal · Where it runs: Vertex AI Agent Engine (ADK)
Agent that scores Product Detail Page (PDP) content against Scotts's Legal + Regulatory compliance framework — 16 rules across 7 categories — using the product label as the source of truth.
Problem
PDP content (Amazon listings, web copy, packaging blurbs) has to comply with Scotts's Legal + Regulatory framework — claims must be substantiated by the product label, disclaimers must be present, quantitative and environmental claims have specific rules, OMRI vs. "Certified Organic" can't be used interchangeably, and so on. Compliance review used to be manual, slow, and inconsistent across batches; Legal and Content teams had to read every line by hand.
What it does
- Loads two inputs: the extracted product label (regulatory source of truth) and the product copy (what's about to ship on the PDP).
- Evaluates every content field against all 16 compliance rules spanning 7 categories:
| Category | Rules | Weight |
|---|---|---|
| Truthful and Accurate Claims | 5 | 30% |
| Complete Claims with Disclaimers | 2 | 20% |
| Quantitative Claims | 1 | 10% |
| Environmental Claims | 1 | 10% |
| OMRI Claims | 1 | 5% |
| Action Verb Accuracy | 1 | 10% |
| Verbatim Meaning Preservation | 5 | 15% |
- Produces a structured score per content field with rule-by-rule violations, weighted to an overall compliance score.
- Two variants: onetime (evaluates entire content from scratch — used for batch reviews and pre-publish gates) and field-level (re-scores a single edited field — used during interactive content editing).
- Pydantic-validated output so downstream tools (Content review dashboards, batch scoring pipelines) can consume the score programmatically.
Co-built with
- Content team — supplied the PDP corpus, defined the editing workflow, and validated outputs against real Amazon listings
- Legal — authored the 16-rule framework, weighting, and the per-rule "what to check" criteria; signs off on rule changes
- AI CoE — built and operates the agent on Vertex AI Agent Engine
Who owns it
- Business owner: Content team + Legal (joint)
- CoE AI Solution Lead: [TBD]
- Tech stack: Google ADK (Agent Development Kit) Python agent, Pydantic schemas, Vertex AI Agent Engine runtime
Where it runs
- Source:
ContentGen/compliance_checker_onetime_agent— ADK Python agent (root_agentexport pattern) - Compliance framework doc: lives in the ContentGen repo (
COMPLIANCE_FRAMEWORK.md) — Legal-authored, versioned alongside the agent - Deployment: Vertex AI Agent Engine
Data it touches
| Data source | Classification | Access path |
|---|---|---|
| Product labels (extracted text) | Public + Regulated | Read-only, supplied as input to the agent |
| Product copy (draft PDP content) | Internal | Read-only, supplied as input |
| Compliance scores | Internal | Returned as structured output for downstream review |
Governance status
- AI Governance Committee reviewed: Yes — pilot use for Amazon listing scoring; Legal owns the rule framework, so changes route through Legal sign-off
- Human-in-the-loop required for: All compliance scores are advisory — humans review violations before content ships. Agent outputs scores + flagged rule violations; the publish/no-publish decision is human
- Vendor / model risk noted: Standard Vertex AI / Gemini risk profile
Patterns reused
- Label-as-source-of-truth pattern shared with MyLawn Brain — both initiatives treat the product label as the regulatory authority
- ADK Python agent + Pydantic output pattern reusable for any rule-scoring agent (content review, ad-claim checks, packaging review)
What's next
- Wider rollout beyond Amazon listings — Scotts.com PDPs, packaging blurbs, ad creative
- Tighter integration with the content-generation workflow so newly drafted copy is scored at draft time, not at publish gate
- Legal review cadence on the rule corpus as marketing language evolves
Changelog
- 2026-05-21 — Initial entry on the hub.