We implement AI agent systems for other businesses, so our content operation has to prove the model. The semi-automated pipeline we built has published 94 bilingual post pairs in English and German — no writer, editor, or content manager on staff.
The situation
We implement agent systems for other businesses, so our content operation has to demonstrate the model rather than just describe it.
The goal was never just Google rankings. The target is AI assistants: when a founder asks Claude or ChatGPT which companies implement agent systems for small businesses, we want to be in the answer. That channel — generative engine optimisation, or GEO — rewards content built for extraction: clear entities, explicit claims, sentences that hold their meaning in isolation.
A well-researched bilingual post pair takes 7–8 hours by hand. At that pace, three or four pairs a month is the ceiling — not enough to build a presence.
The research layer
A content research API queries nine sources per topic seed: Google autocomplete, People Also Ask, Google Trends, SERP results, Reddit, Hacker News, Wikipedia, the Google Knowledge Graph, and an LLM probe. The output is a ranked cluster plan: execution order, near-duplicate warnings, and seeds to cut because the signal isn't there.
Two of the nine sources are specifically for GEO. The Knowledge Graph signal shows whether a topic has an established entity. The LLM probe tests whether AI assistants actually answer the question — and whether a new post has a realistic chance of being cited over what already ranks.
The strategy is generated by the agent, not for the agent. We approve direction — not the analysis.
The production layer
Each post runs through a five-phase skill built in Claude Code.
The brief has a rejection condition. Six questions: audience, takeaway, argument, strongest claim, and two highlight candidates. The test: would a reader who just saw the title already know this? If yes, the brief gets reworked. Weak posts die here — not after four hours of writing.
Writing runs against a section map. Every section gets a heading and a one-sentence job before it's written. The intro is written last, once the article it has to frame exists.
The quality gate is eight separate passes. Clarity, Voice, So what, Prove it, Specificity, Heightened emotion, Zero risk, and AI extractability. The eighth reads every sentence in isolation: AI assistants cite sentences, not pages, so any sentence that leans on "this means…" or "it handles…" gets rewritten until it stands alone.
The German version runs in the same session. Formal Sie throughout, tool names in English, localised frontmatter — with an automated scan for informal pronouns before the file is saved.
The brief gets rejected before prose begins — not after four hours of writing.
What changed
- 94 bilingual post pairs published. At Orbit Media's surveyed average of 3 hours 48 minutes per post, that's north of 650 hours — a sixteen-week full-time writing sprint, running in the background.
- Every post passes the same eight-sweep gate. Quality no longer depends on how much time the week had.
- Topic selection is evidence-based. Every topic has a signal tier before a brief is written.
- Content work stopped competing with client work.
Where we stay in the loop
Two approval gates: the cluster strategy before posts are commissioned, and the brief before writing starts. Publishing is a human step too — every post is reviewed before it goes live. Everything between the gates runs without interruption.
The agent doesn't decide what we should be known for. Positioning and vertical choices are made before the pipeline starts.
Where we are now
94 post pairs are live — 188 MDX files in English and German. Wave 4 shipped; Wave 5 — 37 demand-validated posts — is queued for the pipeline.
The pipeline we run for ourselves is the one we build for clients.