Programmatic SEO without getting nuked in the next Helpful Content update. The March 2024 Google update took out 95–98% of traffic for zero-oversight pSEO sites; hybrid sites with review-gated templates recovered within hours. The differentiator is human-in-the-loop, not scale.
Source synthesis: post-HCU forensics, Lily Ray on AI Overview substrate (A single seeded fake claim can self-confirm in AI Overviews), the citation/mention gap (Citation rate and mention rate are different metrics; comparative content closes the gap).
Non-negotiable rule set
1. Mandatory human edit pass per generated page. No auto-publish.
2. ≥500 unique words per page that are not template boilerplate.
3. ≥30% differentiation between any two pages from the same template (measured on paragraph-level embedding distance).
4. Monthly pruning. Any pSEO page below N impressions after 90 days is noindexed or deleted.
5. Weekly signal monitoring. Search Console impressions + positions trend per template.
Template structure
Each template lives in a directory with:
template.md— skeleton with{variable}placeholders.variables.csv— the combinatorial variable set.differentiation-rubric.md— what makes pages non-duplicate.review-checklist.md— manual pass criteria.
Build loop
template builder agent
↓ assemble variables + template + examples
↓ generate candidate page
↓ differentiation audit (embedding distance)
↓ human-value audit (does this answer a real question?)
template-publish gate (human approves the TEMPLATE, not each page)
↓ template unlocked for batch generation
seo writer → page instances with mandatory copy-reviewer pre-check
↓ sample review (1 in 10)
↓ final brand-voice check at template level
publish via LP builder or content-publish flow
Safety gates
- Template-level gate. A reviewer approves the template after auditing 5 generated sample pages against the differentiation rubric and human-value test. Only after approval does the template unlock for volume generation.
- Per-page sample gate. 1 in 10 generated pages requires a human read before publish. Randomized sample, not deterministic — prevents gaming.
- Post-publish watch. Any template whose CTR or position regresses past a threshold gets flagged; investigate or kill.
Common failure modes
- Template drift into spam. Same format, no real signal. Mitigation: human-value test at the template gate.
- Cannibalization. Two templates target overlapping SERPs. Mitigation: search-intent audit across all live templates before adding a new one.
- Scaling without monitoring. 10,000 pages, no oversight. Mitigation: weekly prune job + Search Console trends per template.
- Boilerplate over 50%. Differentiation drops, embedding distance fails the threshold.
- Auto-publish. Even one bad batch tanks domain trust under HCU.
- Variable sets that produce empty pages. "Best [tool] for [niche]" with zero data when the niche has no real evidence — page is filler.
Quality gates
- Human approves every template.
- Every page over 500 unique non-boilerplate words.
- Differentiation ≥30% on embedding distance.
- 1-in-10 random page review before publish.
- Monthly prune of underperforming pages.
- Weekly Search Console trend review per template.
What this connects to
- Keyword research and search-intent map upstream (avoid cannibalization).
- Copy reviewer for voice and claim validation.
- AEO relevance engineering for passages and schema on the long-tail.