ZEFT Ads × SEO Audit

We have built an audit system in which an AI agent works through multiple specialist perspectives in sequence — advertising accounts, measurement, search, page speed. Its first target was ZEFT, a social gifting service we build and operate ourselves: full account access, and complete freedom to publish what the system finds. On the surface everything looked healthy: ads were delivering, orders were coming in, and the site rendered fine on every device.

The audit still surfaced measurement configuration issues that surface-level numbers never show. Purchase tracking had been disabled at some point, and past results were recorded at a flat value regardless of actual order size — so automated bidding was operating without knowing what a sale was actually worth. We rebuilt the loop so each completed payment reports its real order value back to the ad platform.

Search had the same shape of problem: left over from a past system migration, many pages looked normal to human eyes but were unreadable to search engines, and the files that guide search crawlers were broken. After the fix, indexable pages went from 169 to 737, and 163 published articles appeared in search results for the first time.

One finding tied both halves together: slow loading on mobile (8.5 seconds to show the main content, versus 1.7 on desktop). That single lever suppresses search ranking and ad efficiency at the same time — fixing one thing improves both. This is why we audit ads and search as one discipline, not two.

None of this is exotic. Every migration and redesign lets configuration drift, and everything keeps looking normal on the surface. These issues only show up when audited against real data — and because the audit is a system rather than a one-off engagement, the same discipline can be run again, on any account.

Technical details

The audit system: an AI agent applies a sequence of specialist skills — account structure, measurement, technical SEO, page performance, and Japan-specific ad compliance — using check items we maintain in-house for the Japanese market.

Measurement: purchase conversion actions had been disabled, historical conversions were recorded at a flat value, and the only event still firing was a low-value micro-event. We implemented a server-side loop that returns each order's actual value from the payment webhook to the ad platform, restoring the basis for value-based bidding.

Search: a platform migration had left the catalog client-side rendered, so crawlers could not read the content. We rebuilt those routes as server-side rendering, moved robots.txt and sitemap generation onto the database, and emit structured data (JSON-LD) in the server response.