The number nobody questioned
A patient site's paid search program reported 4,095 clicks and 3,121 conversions over a single period — a 76.21% conversion rate. The monthly reports went unquestioned because they looked like success. Data that flatters does not get audited. Data that disappoints does. That asymmetry let the measurement failure persist for quarters.
What it actually was
No fraud, no single bug. Ordinary configuration decisions compounded:
- Engagement events from the analytics layer were promoted to conversion status
- Scroll milestones, video starts, and treatment-center locator interactions were all counted as conversions
- Some user actions registered through multiple paths, letting a single visit count more than once
- A platform migration created tracking gaps that distorted historical comparisons
- Automated bidding optimized toward the inflated signal, systematically wasting budget
The fix was subtraction
The rebuild narrowed conversions to native actions tied to real business outcomes: treatment-center searches and doctor discussion-guide downloads. Using a Proxy ROAS framework, the account moved from $1.10 to $2.87 within 90 days.
Four checks for your own account
- Inventory the conversion sources. List every action counted as a conversion and where it originates.
- Check the counting method. Determine whether actions count once per click or per occurrence.
- Sanity-check the denominator. Conversion rates above 20% on cold traffic warrant investigation.
- Audit the value weighting. Make sure different actions carry proportional algorithmic weight.