AI IN MI I BY PETER DODS What the Audit Uncovered
R ecently, a garbled trademark symbol led me to a five-year-old problem I didn’t know I had. A few columns ago, I shared how I’d built an agentic email system in n8n that read cus- tomer inquiries and drafted responses, and it turned me into a full-time supervisor for a system meant to save me time. My conclusion was that agents work well under active supervision, the way agentic coding does, and fall apart the moment you hand them real autonomy. I said I’d remain the human in the loop, and I meant it. Last week, that opinion softened. Here’s why. THE EXPERIMENT
what you’d expect: thin product de- scriptions, an inconsistently updated blog, an agents.md file still sitting as empty boilerplate. My catalog also had no structured specifica- tion data, and digging into why led somewhere I didn’t expect. Shopify didn’t have native meta- fields when we built our site, so the site’s structure has been quietly out of date for four years. This was structural, sitting underneath everything else on the list. THE PRIORITIZATION CALL That left two paths. I could chase the AEO work: agents.md, struc- tured specifications, review mark- up, everything that makes a store legible to an AI shopping agent. Or I could fix product categorization and upload consistency, the un- glamorous plumbing nobody sees. The traffic numbers made the decision for me. AI assistants were sending a fraction of a percent of my sessions, and there’s no point sending more shoppers, human or artificial, to a site whose product data is inconsistent. Categorization and upload consistency come first because that’s what everything else is built on. WHAT THIS REALLY CHANGED I didn’t ask Claude to find my cat- egorization problem. I asked it to run an AEO audit, and it found a much larger issue. That’s the shift agentic AI represents for small busi- ness owners. It can look across an entire business at once. I’d spent years fixing mojibake one symbol at a time and never once asked why it kept happening. What else needs updating that I’ve never thought to ask about? MI Peter Dods is the owner of Honolulu- based Easy Music Center.
I connected Claude Cowork to my Shopify store, which took about as long as clicking a connector and logging in and asked it to run an AEO audit. I expected a surface-level report. It came back as a prioritized list of real problems, and we started fixing them the same day. WHAT THE AUDIT DELIVERED The format won me over first. Every finding was broken down by who it affects, its impact and the effort to fix it. Instead of a wall of text with no obvious starting point, I could see at a glance which changes were high impact and low effort. THE MOJIBAKE Buried in the findings was a name for something that had
bothered me for years. Mojibake is what happens when a special character like ™ or © gets mangled into a garbled string that makes a product page look hacked. It usually starts when someone opens a product export in Excel and resaves it, quietly corrupting those characters before the file is ever re-uploaded. I’d catch one occasionally, fix it by hand and hope I wouldn’t come across another. Claude told me I had 243 of them across roughly 8,000 SKUs — a countable problem I’d been treating as unsolvable. I had it fix one instance so I could check the work personally. Claude then asked about a bulk
“I didn’t ask Claude to find my categorization problem. I asked it to run an AEO audit, and it found a much larger issue.”
upload. I uploaded the reviewed batch of 10. It worked perfectly. From there I kept accelerating, moving from line-by-line checks to casual glances. The pattern held, and we cleared all 243 in short order. What made this work was the pace. Trust was handed over one small batch at a time, scope growing only after each batch held. The n8n agent never got that. THE BIGGER DISCOVERY Chasing the mojibake pulled me into the rest of the findings. Most were
26 I MUSIC INC. I NOVEMBER 2026
Powered by FlippingBook