Back to case studies
Case Study
GEO Content for Oxford Online Pharmacy
Healthcare
Claude
n8n
Airtable
The brief
Oxford Online Pharmacy needed its articles to show up when people ask AI assistants about weight loss, hair loss and erectile dysfunction. In medicine every claim has to match an authoritative source and pass regulatory review, and nothing went live without the pharmacy’s sign-off.
Victoria Rusnac at Applied AI Co designed the pipeline. Before automating it, she brought me in to pressure-test the prompts and the GEO approach.
How the pipeline works
Every article starts from a question people actually ask AI assistants. One agent researches it, another turns the research into a brief, and a third writes the draft. Keeping research and briefing apart is what stopped the hallucinations.
QA agents then check the draft for writing quality, medical accuracy, compliance, links, and whether each target question gets an answer an assistant could quote. They fix what they find. A copywriter edits every article, the pharmacy approves it, and every correction lands in a log the writer reads before the next one.
It started as a chain of Claude prompts, one per stage, run by hand so every output could be checked before anything was automated. Once the stages held up, the chain moved into n8n, which runs the agents in sequence, and Airtable became the control panel where the team tracks every brief, draft and QA report.
The result
A copywriter now edits an AI first draft in 1 h 15. Writing the same article from scratch took four hours.
What I did
I reviewed every prompt, every stage and the way they hand work to each other, with one question in mind: would an AI assistant cite what this produces? These are the changes that came out of it.
A citation check before anything ships. Following the brief isn’t the same as being citable. A final check now runs every article against its target questions: is there a direct answer, could an assistant lift it as it stands, does it use the question’s own words? If not, it goes back.
A pipeline that learns. I suggested tracking each article’s AI citations at 30 and 90 days and feeding the results into the next plan. Victoria took it further: every edit the copywriter and the pharmacy make is now logged, and the writer reads that log before each new article.
One voice, start to finish. Two prompts were pulling every draft in opposite directions, one towards a clinical register and one towards a conversational one. The step that fought the brief came out.
Cleaner context for every prompt. Each step was writing its notes into the article itself, so every later prompt had to read past them. I recommended moving those notes into a separate audit trail, which means less noise in every prompt and an easier move to n8n.
