Search Engine Land contributor Tania Brown spent months building and rebuilding an AI content pipeline in Claude Code. It now supports article updates and production for a company blog and external publications, usually getting pieces about 95% of the way to publication.
The lesson underneath it: the hardest part is not getting AI to produce an article. It is figuring out what the finished article needs to look like, then building the workflow and inputs that reliably get you there.
This is not a quick build, and with Google aggressively noindexing commodity content it may not pay off for every brand. Brown mitigates the risk with research, multiple human quality gates and AI fact-checkers.
If you already have many of the component documents, you can build in stages and iterate — and reuse agents across other workflows later.
A working pipeline needs to produce useful, original content in your brand voice. Pieces should help your ideal customer profile, describe your business accurately, and sound human. Ranking and citation potential is the bonus.
Once you know the target, list what it takes to get there. Values that never change get hard-coded as constants; topic, angle and keyword arrive as per-run inputs.
Mirror your existing content process, then add gates for the two problems AI introduces: robotic language and hallucinated facts. An orchestrator document describing the full workflow and each agent's responsibility keeps steps from being skipped.
A keyword and an angle start the research. If you have multiple ICPs or a product that must be mentioned, specify it up front. Start with one content type and get it working before adding branching for others.
An agent researches the topic, what your brand has already published on it, and current SERPs to identify gaps. Output a dossier for downstream agents. Supply it with a list of reputable sources to check first, sources to avoid, timeliness and sample thresholds, your sitemap, and links to your original data.
Generate the outline, then review it. You see the shape of the finished piece before spending tokens on it, and you decide to scrap, revise or continue.
Feed in the dossier, outline, voice guide, ICP information, case studies and first-party research. Give the model an example of high-quality content written from an outline and have it analyze narrative flow, logical progression and word choice.
Run editor passes before handing anything to a human. Brown's key finding: one editor handling structure, coverage and style together produced worse output than splitting it into one agent for structure and coverage and another for phrasing and AI tells. Asking an editor to also fact-check "resulted in two poorly done jobs."
Run each in a fresh context window.
The most valuable part of this build is not the tool choice — it is the observation that overlapping responsibilities degrade every one of them. That is the same structure behind why people keep rewriting AI drafts, covered in Why You Keep Rewriting What AI Drafted — the Roles Overlap, Not the Sentences.
The second is the verification gate. Automated pipelines are good at declaring themselves done; proving it is a separate step you have to build — see The Missing Step in Vibe Coding Is Verification — How to Make 'Done' Provable.
And the condition Brown states plainly: an LLM can write the agent docs quickly, but you still have to supply the context that makes the content worth reading. Without ICP information, examples of quality work and first-party research, no workflow produces non-commodity content. Nothing should publish without passing through a human either.
Not generating text. It is defining what a finished, publishable piece must look like, then designing the inputs and order of operations that reliably reach that standard.
One editor handling structure, coverage and style produces worse output than two specialized agents. Fact-checking bundled into editing degrades both jobs, so it needs its own adversarial agent.
One at the outline stage, so you can scrap, revise or continue before spending tokens on a full draft — and a mandatory one before publication.
Adversarially. It should assume every claim in the draft is wrong and actively try to disprove it, which is what catches hallucinated statistics.
To apply what you just read to your own site, start with a free audit of where things are now.
A strategist replies within 24 hours on business days.
New research on AI search optimization terminology: 81% of marketers still say SEO, budgets already sit at 24%, and terminology fluency ranks last as a credibility signal at just 9%.
An E-E-A-T checker was never automatable because there is no API for it. Here is how an AI coding agent read Google's quality guidance, crawled a page sample, and produced a scored audit — plus where it stops.
A GEO audit is not checking whether your brand appears in an AI answer. Here is a nine-point GEO checklist split into four on-site items, four off-site, and one measurement layer.