SEO helpful-content proxy

Business value

• Closes the feedback loop on helpful-content updates. You don't have to wait for Google to demote a page to know it's weak — the signals are visible in the draft.
• Prioritizes editor time toward the pages most likely to cost you in the next core update, not the pages with the highest keyword volume.
• Gives the rewrite agent something to chase: specific evidence spans instead of a vague 'add E-E-A-T' note.

Agent job to be done

Be a helpful-content reviewer. For the target query, is this page genuinely useful — concrete examples, original information, evidence, named sources? Or does it summarize what every other page on page one already says?

format: articleintended_use: publishdomain: SEO helpful content / pre-indexation QA

When to call VeracityAPI

Before initial publish, and again before pushing a refresh of an already-indexed page. The refresh case matters: most helpful-content damage happens when a generator-driven 'refresh' replaces firsthand sections with paraphrase.

What text to submit

Primary content, title, meta description, H1/H2s, intro, key comparison sections, conclusion. Exclude global nav, sidebar, footer, author bio, and related-posts blocks. Include FAQ schema text — that's where slop hides.

Decision policy

• allow: low risk AND content_trust_score ≥ 0.65 AND at least one named example per H2.
• revise: medium risk OR evidence flags thin comparison tables, paraphrased summaries, or unsupported best/safest/cheapest claims.
• human_review: high risk on YMYL pages (health, money, safety, legal), OR pages targeting commercial-intent queries above $5 CPC. The cost of getting these wrong is too high for autopublish.
• Section-level rule: if any H2 section scores high risk, treat the page as high risk even if the average is medium.

Automation recipe

• Content agent generates or refreshes the page.
• Extractor splits by H2 and submits each section separately to /v1/analyze.
• Worker maps evidence spans back to heading + character offsets so editors can jump to the bad paragraph.
• Revision agent adds firsthand details (named examples, data tables, screenshots, named sources) before rescoring.
• Page only enters the publish queue when every section is allow.

Evidence spans agents should inspect

• 'thin_comparison' — comparison tables where rows just list features without differentiating value
• 'paraphrase_summary' — paragraphs that read like a competitor's H2 reworded
• 'unsupported_superlative' — best/safest/cheapest without a named comparison or measurement
• 'missing_firsthand' — entire sections without a named example, screenshot, data point, or original observation

Policy pseudocode

if (result.recommended_action === "allow") continueWorkflow();
if (result.recommended_action === "revise") rewriteWith(result.evidence, result.recommended_fixes);
if (result.recommended_action === "human_review") queueForHumanReview(result);
if (result.recommended_action === "reject") discardOrRebuild();

KPIs to track

• pre-publish block rate by section
• % of pages improved (specificity_risk drop) on second pass
• indexed-page ranking delta for gated vs. ungated cohorts (set up an A/B at the URL-path level)
• organic traffic retained 90 days after a core update
• editor queue size — needs to stay sustainable

What can go wrong

• This is not a direct Google classifier. The signal is helpfulness — which correlates with helpful-content updates but doesn't predict them.
• Helpful content includes UX, internal linking, reputation, and user-satisfaction signals VeracityAPI can't see.
• Use evidence spans to fix pages. Chasing the score directly will push you into over-specific, brittle copy.

Cost and latency notes

Analyze only is $0.005 per 1,000 characters; Analyze + revise with auto_revise=true is $0.010 per 1,000 characters. Both round up to the nearest 1,000 characters. Short captions/emails usually cost $0.005; longer pages or chapters scale linearly by length. Current v0.1 latency is LLM-bound, so batch/concurrent orchestration is recommended for high-volume pipelines.