Key facts
| Core tools | Keyword data, sitemap and crawl reader, content inventory, analytics reader |
| Primary outputs | Intent clusters, content briefs, internal link suggestions, technical issue lists |
| Publishing rule | Drafts only; a human reviews and publishes |
| Crawl hygiene | Respect robots rules, rate-limit requests and cache crawl output |
| Grounding | Briefs cite existing pages and real queries, not model memory |
| Models | 30+ models on one OpenAI-compatible API for clustering and drafting |
| Embeddings | Live embeddings and RAG for intent clustering and dedupe |
| Roadmap | Batch endpoints are coming soon for large crawl jobs |
TL;DR
- Position the agent as analyst and drafter; humans publish.
- Cluster queries by intent with embeddings instead of exact-match lists.
- Map every brief to a gap you can prove from your own content inventory.
- Audit internal links and technical issues on a schedule, not ad hoc.
- Track time saved and review quality; ignore volume-based vanity metrics.
How it works, step by step
- Export keyword data, sitemap, crawl results and analytics into a single working set.
- Cluster queries by intent with embeddings and label each cluster with its job to be done.
- Compare clusters against your content inventory to find real gaps and cannibalisation.
- Generate briefs that specify intent, outline, entities, internal links and success criteria.
- Run a scheduled technical pass: broken links, redirect chains, orphan pages and duplicate titles.
- Queue briefs and fixes for human review with priority based on business value.
- Measure outcomes per cluster and update the keyword set on a fixed cadence.
Try it yourself
Open the agent workflow designer →
Four jobs an SEO agent does well
Divide SEO work into four repeatable jobs. First, research: collect queries and cluster them by intent. Second, gap analysis: compare those clusters to what your site already covers and find the missing or weak pages. Third, briefing: produce a structured brief per target page with the required entities, questions and links. Fourth, technical hygiene: crawl your own site on a schedule and report broken links, redirect chains, orphan pages, duplicate titles and missing metadata.
All four are analysis-heavy and low-risk, which makes them good agent work. None of them requires touching production without review.
Clustering and gap analysis with embeddings
Keyword lists hide structure that embeddings reveal. Embed every query, cluster by similarity, then label each cluster with the intent it serves: informational, comparison, transactional or navigational. The same embeddings let you match clusters to your existing pages, so you can tell the difference between a true content gap and a page that already covers the topic badly.
- Intent labels drive the brief's format and the call to action.
- Cannibalisation checks catch two pages competing for one cluster.
- Entity coverage lists the people, products and concepts the page must mention.
- Link suggestions come from the embedding index, not from model recall.
Guardrails, limits and what to measure
Search engines reward quality and punish automation that produces thin pages at scale. Keep a human review gate on every publish, cap how many briefs the agent produces per cycle, and require each brief to link to at least one existing page and cite real query data. Respect robots rules on any crawl, rate-limit requests and cache what you fetch.
Plugsky provides the model layer: 30+ models on one OpenAI-compatible key with live function calling, streaming, JSON mode, embeddings and RAG. It does not crawl sites or connect to analytics for you, so keep those tools in your own pipeline. Use a fast model for clustering and a stronger model for briefs, and measure time saved, review pass rate and ranking movement per cluster. Plans and the free plugsky-micro and plugsky-lite models are on the live pricing page; batch endpoints are coming soon.
Honest comparison
| Task | Manual process | Prompt-only tool | SEO agent with review |
|---|---|---|---|
| Keyword clustering | Spreadsheet judgement | Noisy one-shot lists | Embedding clusters with intent labels |
| Gap analysis | Manual comparison | Ignores your inventory | Cluster mapped to existing pages |
| Brief quality | Varies by writer | Generic outline | Entities, links and success criteria |
| Technical audit | Periodic and partial | Not covered | Scheduled crawl with prioritised issues |
| Risk | Low | High if published unreviewed | Review gate on every publish |
Frequently asked questions
Can the agent publish content directly?
It should not. Drafts and briefs go to a human editor, because publishing thin or inaccurate pages at scale damages rankings and brand. Keep the publish action outside the agent's tool set.
How do I cluster keywords without a data-science team?
Use an embedding model to vectorise queries, then group by similarity threshold and label each group manually once. Re-run the same pipeline as the query set changes.
How often should the technical audit run?
Weekly for active sites and after every deploy. Cache crawl results and only re-check changed sections between full runs to control request volume.
Which model writes the best briefs?
A mid-to-frontier model with strong instruction following. Clustering, classification and metadata generation can run on small fast models for much lower cost.
How do I measure whether it works?
Track review pass rate, time saved per brief, ranking movement per cluster and organic conversions. If briefs are rewritten heavily, refine the brief template first.
Is this against search engine guidelines?
Automation itself is not the problem; low-quality output is. Human review, original value and accurate claims keep you inside the guidelines. Respect robots rules when crawling.
Does Plugsky crawl or connect to analytics?
No. Plugsky provides models and the API layer, including embeddings and function calling. Crawlers, keyword data and analytics connections are integrations you build or buy.