SEO Analyzer Agent

An Apify Actor that runs a full SEO pass on any URL and returns structured data plus a Markdown report. Point it at a page and it will:

1. On-page / technical audit — crawl the page and flag issues across title, meta, headings, links, images, structured data, mobile & i18n. (No API key needed.)
2. SERP & keyword research — pull live Google results for the keyword and extract who ranks, People-Also-Ask, and related searches. (Uses apify/google-search-scraper.)
3. AI content optimization — generate an SEO content brief (title, meta, outline, entities, gaps, internal links, priority actions) with Claude. (Needs an Anthropic API key.)

Each job is independently toggleable, so it runs fine with just the audit, or all three together.

Grounded in primary sources — live corpus lookup

The AI steps are grounded in the 2024 Google Content Warehouse leak (14,022 attributes) and 5,456 Google search patents, bundled in src/leakdata/ from the Missing Intent SEO skill.

On every run, the Actor queries the real corpus for the page's situation: it detects intent (commercial/reviews, local, YMYL, informational, thin, media-heavy) and retrieves the actual matching ranking signals (with Google's own descriptions) and patents (with their actions). Claude then follows an evidence rule — cite the real leak attribute (Module.attributeName) or patent (US…), treat undocumented fields as existence-only, never invent weights — and applies Signal Projection (optimize the behavior a signal measures, not the metric). The content brief and keyword insights each return an evidence list with those citations, so every recommendation is checkable.

Two modes (via researchMode): Agentic (default) gives Claude the corpus as tools (search_signals, search_patents) and lets it decide what to research and iterate — like the skill's router — with a structured synthesis at the end; the searches it ran appear in the report under "What the agent researched." Standard retrieves the matching evidence deterministically in one pass (faster, fewer API calls). Both surface a receipts section; if agentic research ever fails it falls back to standard automatically.

Query engine: src/leakdb.py (reuses the skill's whole-word/camelCase matching); per-page retrieval: src/evidence.py; agentic loop: src/agent.py.

Input

Field	Type	Notes
targetUrl	string (required)	The page to analyze.
seedKeyword	string	Primary keyword/topic. Inferred from the page title/H1 if omitted.
runOnpageAudit	boolean	Default true. Works with no API key.
runKeywordResearch	boolean	Default true. Consumes Apify credits (calls the Google SERP scraper).
runContentOptimization	boolean	Default true. Requires anthropicApiKey.
anthropicApiKey	string (secret)	Your sk-ant-... key. Stored encrypted; never logged.
model	select	claude-opus-5 (default) · claude-sonnet-5 · claude-haiku-4-5.
researchMode	select	agentic (default) — Claude queries the leak/patents with tools and iterates · standard — fast single-pass retrieval.
country	select	SERP localization (default us).
resultsPerKeyword	integer	Organic results to pull (default 10).

Output

• Dataset: one record with onpage, serp, keyword_insights, and content_brief sections.
• Key-value store: a rendered Markdown report under the key REPORT.

Run it locally

pip install apify-cli            # or: npm i -g apify-cli
cd seo-agent
apify run --input '{"targetUrl":"https://example.com/","runKeywordResearch":false,"runContentOptimization":false}'

That command runs just the audit (no keys/credits). Add anthropicApiKey and flip the flags on to enable the AI and SERP steps.

Deploy to Apify

apify login
apify push

Notes

• The AI steps default to Claude Opus 5 via the official anthropic SDK, using structured outputs so the brief is always valid JSON. Switch to Sonnet 5 or Haiku 4.5 for lower cost.
• The on-page audit is fully deterministic — no LLM, no keys — so it's cheap, fast, and reproducible.
• Keyword research relies on the apify/google-search-scraper Actor, which consumes platform usage.

Roadmap ideas

• Multi-page crawl (audit a whole site, not one URL)
• Competitor backlink & content-gap comparison
• Internal-linking graph across the crawled pages
• Rank tracking over time (scheduled runs)