How To Build a Content Team of SEO AI Agents (n8n, OpenAI, Aidbase)

A fully autonomous SEO content pipeline can be built by chaining AI agents for keyword discovery, topic planning, deep research with citations, retrieval of private internal knowledge, long-form drafting, thumbnail generation, and automated publishing—using n8n plus OpenAI, 8base, and Replicate. The practical payoff is speed: daily blog posts for products like Feed Hive, Link Drip, and Abbase can be produced with little to no human intervention after setup. The bigger question is whether this kind of mass-produced AI publishing triggers Google’s spam filters or deranks sites indefinitely; the workflow’s creator argues the risk can be reduced by focusing on usefulness, adding verifiable references, and injecting unique internal knowledge that competitors can’t copy.

The approach starts with topic selection and keyword planning. Instead of relying on brittle “fully autonomous” long-tail keyword research, the workflow uses SERP data via SERP API (Google results returned as JSON) and AI search via Perplexity’s sonar models or OpenAI’s search preview model. Two operating modes are considered: manual keyword research followed by tightly instructed writing, or a more autonomous path where OpenAI search generates content ideas from a product/brand description. The creator favors the latter for autonomy, accepting less control and a smaller chance of ranking well per post in exchange for hands-off publishing.

To prevent the system from repeating itself, the workflow adds a deduplication step. Earlier attempts using vector databases to compare “idea similarity” didn’t work well in practice, so the system instead feeds the research agent a simple archive of summaries from prior posts pulled from the CMS. In the example setup, Strapi provides the blog archive via an HTTP request node, and the agent receives the archive as a JSON string before generating new topic plans.

Drafting initially produced generic, shallow articles—an outcome that would likely resemble spam—until the workflow inserted two additional research layers. One agent performs external research to gather statistics, facts, and citations for the specific post title and outline. Another agent retrieves unique internal material from a private knowledge base using RAG (retrieval-augmented generation). For the internal layer, 8base is used to train a chatbot on selected sources (for example, Feed Hive’s website and help desk, plus internal YouTube resources) and on custom FAQ entries. The workflow then calls the 8base chatbot through an API endpoint inside n8n, so the final draft can include business-specific insights that aren’t available elsewhere.

Once the post is researched and written, the pipeline generates brand-consistent thumbnails. The creator abandons a purely AI-generated “text-on-image” approach and instead uses Replicate’s Flux model (Black Forest Labs’ Flux) to generate the base image, then uses a custom NodeJS API built with the canvas library to compose the thumbnail using a theme, highlighted words, and overlay text. Finally, the system publishes to Strapi (and optionally triggers Feed Hive social posting), with n8n scheduling the whole chain to run daily, weekly, or on any cadence.

The overall message is caution: the setup is experimental and potentially risky for businesses where SEO is the primary revenue channel. For lower-performing or “dead” blogs, the creator frames the automation as a calculated gamble—worth trying if the content can be made genuinely helpful, well-sourced, and uniquely informed by internal knowledge.