Nine AI agents, one SEO department.
How a multi-agent pipeline plans, writes, reviews, illustrates, tags, publishes, and measures SEO content for a catalog no human team could cover page by page — with editors approving batches, not babysitting drafts.
- 5,000
- Products
- 100
- Services
- 30
- Locations
- 50
- Blog categories
The problem at this scale
5,000 products sold across 30 locations with 100 services means the theoretical page inventory runs into six figures. Every page needs current copy, correct schema, internal links, and imagery — and search competitors refresh weekly. A human content team can either cover a fraction well or everything badly. This client's organic traffic had plateaued for two years because production, not strategy, was the bottleneck.
The agent team
How the nine agents collaborate.
Each agent has one job, defined inputs, and defined outputs. They communicate through a shared task queue and vector memory — never by passing free-form text and hoping.
Planner Agent
The editorial brain. Reads the site inventory, keyword universe, and business priorities, then produces a prioritized content plan: which pages to create or refresh, in what order, and why.
In: Site inventory (5,000 products × 30 locations), keyword research memory, seasonality data, business priorities
Out: Ranked task graph — e.g. "refresh 240 product pages losing rank; create 30 location × service pages; 12 blog briefs"
Research Agent
Feeds every downstream agent with evidence: SERP analysis, competitor coverage, People-Also-Ask questions, and internal search data for each planned page.
In: Task briefs from the Planner, live SERP data, competitor sitemaps
Out: Research dossiers with target keywords, entities to cover, questions to answer, and competitor gaps
SEO Agent
Turns research into on-page specifications: title, meta, H-structure, target/LSI keywords, internal-link targets, and canonical rules — consistent across the entire catalog.
In: Research dossiers, site-wide SEO rules, current page audits
Out: Per-page SEO specs the Content Agent must satisfy; validation checklists for the Review Agent
Content Agent
Writes at catalog scale in brand voice — product descriptions, service pages, location pages, and blog posts — always from the SEO spec plus structured product data, never from imagination.
In: SEO specs, product/service data from the PIM & CMS via MCP, brand voice guide in memory
Out: Draft pages with structured sections, FAQ blocks, and internal-link anchors
Review Agent
The quality gate. Checks every draft for factual consistency with product data, spec compliance, duplicate content across the 5,000-page catalog, and brand tone. Rejects and returns with reasons.
In: Drafts, source data, SEO specs, similarity index of existing pages
Out: Approved drafts, or rejections with structured feedback the Content Agent can act on
Image Agent
Produces or selects imagery per page: renders product-variant images, generates blog illustrations, compresses everything, and writes descriptive ALT text tied to the target keywords.
In: Approved drafts, product image library, brand style rules
Out: Optimized images with ALT text, uploaded to the CDN, referenced in the draft
Schema Agent
Generates JSON-LD for every page type — Product, Service, LocalBusiness (×30 locations), Article, FAQ, Breadcrumb — and validates it before it ships.
In: Approved page + structured source data
Out: Validated schema markup embedded per page
Publishing Agent
Pushes approved pages to the CMS through its MCP server, sets canonicals and redirects, updates the sitemap, and requests indexing — on a throttled schedule that looks natural to search engines.
In: Fully approved pages with images and schema
Out: Live URLs, updated sitemap, indexing requests, publish log
Analytics Agent
Closes the loop. Watches rankings, impressions, clicks, and conversions per page, attributes wins and losses, and writes findings back to memory — so next week's plan is smarter than this week's.
In: Search Console, GA4, rank tracking via MCP
Out: Weekly performance report, re-optimization tasks fed back to the Planner
Workflow
The SEO automation pipeline.
The plumbing
Memory, MCP, approvals — what makes it enterprise-grade.
Memory
A shared vector database (Qdrant) holds the keyword universe, brand voice, past performance, and every published page's fingerprint. Agents read before acting and write after — the system compounds knowledge instead of repeating work.
MCP
Every external touch — CMS, PIM, Search Console, GA4, image CDN — goes through MCP servers with scoped permissions. The Content Agent can read product data but can't publish; only the Publishing Agent can, and only after approval.
Human approval
Editors review batches in the dashboard, not individual documents in email threads. New page types require approval; proven templates graduate to spot-check sampling. Nothing indexes without a human having defined its guardrails.
Notifications
Slack digests announce completed batches, rejected drafts needing decisions, and ranking anomalies. Approvals happen inline — one click from the notification.
Dashboard
One view: pipeline status per agent, pages awaiting approval, publish velocity, and per-page performance since launch — with drill-down to every agent action for audit.
Analytics loop
The Analytics Agent doesn't just report — it files re-optimization tasks. Pages that underperform re-enter the pipeline with diagnosis attached. That feedback loop is what makes the system improve monthly.
FAQ
Common questions.
How do the SEO agents avoid publishing low-quality AI content?
Three controls: the Content Agent writes only from structured source data and an SEO spec (never free generation); a separate Review Agent checks factual consistency and duplication across the catalog; and humans approve every new page type before it can publish, with spot-check sampling after trust is earned.
How long does an enterprise SEO automation system take to build?
A pipeline like this one deploys in 8-12 weeks: weeks 1-2 for MCP integrations (CMS, PIM, Search Console), weeks 3-6 for the agent pipeline on a subset of pages, then scale-up with human review tightening or loosening per page type.
Does this replace an SEO team?
It replaces the production grind — writing, tagging, schema, publishing at catalog scale. Strategy, brand judgment, and approval stay human. One editor effectively directs the output that previously required a content team of eight to ten.
Have a catalog search engines are ignoring?
We'll map an agent pipeline to your inventory in one call — or email hello@aiplusagentic.ai.
Book an Architecture Call