Building a Content Pipeline That Runs On Autopilot

Marketing teams often rely on a CMS like WordPress paired with a personalization tool like Mutiny. These tools are designed to swap out a headline or a call-to-action on an existing page. They cannot generate the entire body of an article from scratch, tailored to a specific user segment. If you want to create 50 landing pages personalized for 50 different job titles, you are still stuck writing 50 full articles by hand.

For example, consider a B2B SaaS company trying to target both marketing agencies and software engineering teams. A marketer using HubSpot can create two separate landing pages. But when a new long-tail question emerges like "how to manage agile sprints for a remote agency," creating a bespoke, high-quality article takes days. The workflow is manual, slow, and cannot react to the thousands of specific questions potential customers ask every day.

Some teams try a headless CMS like Contentful, thinking it will solve the problem. While it decouples content from presentation, it still requires a human to write every content entry. The bottleneck is not the technology for displaying content; it is the human-powered process of creating it. You cannot hire enough writers to create a truly personalized content experience for every niche question in your market.

The structural failure is that these platforms are built on a document-centric model. They manage discrete pages created by people. An automated pipeline operates on a data-centric model. The system treats a question as a trigger, customer segment data as context, and generates the final page as an output, a process that scales infinitely without adding headcount.

We built our own AEO pipeline based on a data-centric model, and the same principles apply to building a system for content personalization. The first step is discovery. We map your target audience segments and identify high-value question sources, whether from industry forums, Google PAA, or your own customer support logs from a tool like Intercom. This audit defines the data inputs for the entire system.

We built the core pipeline in Python, orchestrated by GitHub Actions for daily execution. The process begins by pulling new questions and embedding them to check for semantic duplicates against a pgvector index in Supabase. For personalization, the system injects segment-specific data (e.g., industry terminology, relevant pain points) into a prompt for the Claude API. This ensures the generated content speaks directly to the target persona. A 100-page run consumes roughly 3 million tokens and completes in under an hour.

The delivered system is a fully automated engine. Each generated page passes an 8-point QA check that uses the Gemini API to score answer relevance and the Brave Search API to verify web uniqueness. Approved pages are deployed instantly to Vercel with correct Schema.org structured data and submitted to search engines via IndexNow. You get a dashboard that shows content production volume and a 9-engine Share of Voice monitor that tracks how your content ranks in AI search results.